Off the Yellow Brick Road

Get clear which road you’re on. Then stop watching the other one. Focus, focus, focus is the new location, location, location.

WHERE WE LEFT IT

Three arguments here over three weeks. CRE is two industries under one name, and AI hits each differently. The middle layer of PropTech is hollowing out as buyer-builders take it in-house. The consortium deals are a game mid-tier operators should refuse to imitate.

For years, CRE technology came from two places: a startup sold you software, or you built your own. The incumbents sat beneath both as the systems of record. AI has added two more players. The labs are taking the horizontal work directly. The consortia are taking the repeatable middle. Five kinds of player now compete across the same ground, and the only useful question is where each should stand.

A piece from Andreessen Horowitz, published the same week as the last of those three, gives us the map to answer it.

WHAT THE VC SIDE JUST SAID

Joe Schmidt, partner at a16z, published Avoiding Death on the Yellow Brick Road on 27 May. It is, in Roger Martin’s terms, a where-to-play map. The labs own the Yellow Brick Road: horizontal, low-step coworker work that improves with raw model capability. Everything else lives in the Rest of Oz: vertical, multi-step, regulated, judged on customer P&L rather than benchmarks. Both can win. Most application-layer startups will die because they walked onto the road.

The Rest of Oz survives the labs on four defences, Schmidt argues: data flywheels, model routing across vendors, cost optimisation across model tiers, and governance that absorbs regulatory complexity for the buyer. Three tests tell you whether you are on it. The tools-and-steps test: how many steps does the work take, and how complex are the tools beneath it? The system test: are you the system the customer runs work through, or a tool sitting on one they already have? The P&L test: are you judged on customer outcomes, or on benchmark scores?

Apply this to CRE and the diagnosis from the last three pieces survives intact. The middle layer of cognitive workflow software dies. Generic AI-for-X PropTech fails all three tests. As last week’s piece argued, the buyer becomes the builder where the work carries its proprietary edge. Schmidt adds the other half from the venture side: the startup wins where it owns the system of work, not where it wraps a tool around someone else’s system. Two people landing on one diagnosis from opposite ends of the industry is the closest thing to validation an argument like this gets. The three tests are also sharper than anything in the trilogy; use them on whatever you are buying or selling.

THREE TERRAINS, FIVE PLAYERS

Translate the map to CRE and there are three terrains, not two. The Yellow Brick Road carries generic coworker work. Between it and the deep country sits Mid-Oz: vertical-ish work, repeatable and configurable enough to standardise across firms. The deep Rest of Oz is the domain-specific, multi-step, regulated, judgement-heavy work that cannot be built from generic capability alone.

Now put the five players on that ground.

The labs own the Yellow Brick Road and reach into the edges of Mid-Oz. The consortia — Anthropic’s enterprise AI venture with Blackstone, Hellman & Friedman and Goldman Sachs, and OpenAI’s Deployment Company — are built to industrialise Mid-Oz, standardising repeatable workflows across the companies they deploy into. They have the capital to do it. They are not the only ones who can. The deep Rest of Oz belongs to two kinds of player: operators who build their own systems, and startups that own a single painful, regulated, commercially material process end to end and get better at it with every deployment. The incumbents sit beneath all of it, across the systems of record.

That is the where-to-play map. The two new entrants, labs and consortia, arrived with clear remits. The three players who were already here — incumbents, startups and operators — now have to relearn where their advantage lives. Take them in turn.

THE INCUMBENTS HAVE LOST THEIR FREE PASS

The first is the incumbents. Yardi, MRI, Argus, CoStar, MSCI/RCA hold the systems of record and the data primitive layer. Schmidt predicts they survive: whoever owns the system of work survives the labs. The position is theirs to lose, and it is defensible. They are also, on any fair read of their fifteen-year record, the firms in CRE least obviously equipped to execute on AI at speed.

That did not matter before. They were unassailable. Switching costs measured in years, certifications in decades, relationships in golf-course-time. They did not need to be good at innovation because nothing they did was contestable. Argus has long been the sector’s byword for standing still. None of these firms has been a natural home for ambitious AI talent. None of it cost them their position.

It does now. The position has not become indefensible; the cost of defending it has gone up. While still unlikely, it is no longer impossible that an embedded operating system gets supplanted, worked around, or abstracted away from being critical. The three challenger paths I named last week — operator-spinout architectures, specialist services-as-software vendors, and lateral entries on incumbent failure points — can each take meaningful share without dislodging the incumbent outright. None of those routes had credible AI-driven economics eighteen months ago. All of them do now.

So the incumbents win this round by default, and have to play the next one properly. The free pass on execution and innovation is the thing that has changed. Should win is carrying real weight in that phrase. Whether they will is no longer the irrelevant question it has been for fifteen years.

THE EASY STARTUP IS DEAD

The second is the startup, the player Schmidt is really writing about and the one whose ground has moved most. His argument is widely misread as “the labs kill startups”. It is narrower. The labs kill the lazy startup. A model wrapper, a reporting overlay, a generic AI-for-real-estate copilot on someone else’s system: the road absorbs all of them.

The startup that survives owns a system of work. It takes one painful, multi-step, commercially material CRE process and becomes the place that process happens, capturing the workflow data, the governance and the record of what was done. It picks ground where the incumbent is too slow to follow, the operator cannot justify building alone, and the labs cannot learn the domain from outside. And it gets better at that one thing with every deployment, which is the one moat Schmidt thinks survives contact with the labs.

That is a narrower opening than PropTech has enjoyed for a decade. It is also a more defensible one. The winning CRE startup will not sell AI into the industry. It will take a slice of the industry’s work and own it end to end.

THE OPERATOR DEFAULT IS DATED

The third is the operators. The default setting for almost every CRE operator today is buy. That setting is rational. It is also dated.

When sophisticated operators made their AI decisions through 2024 and 2025, the build path was not credibly open to a mid-tier firm. Agentic infrastructure was immature. Models could not yet do agentic work reliably enough for a mid-tier firm to bet on them. Skills frameworks barely existed. Forward-deployed engineering for mid-market operators barely existed. Buying from an incumbent or a specialist vendor with the infrastructure already in place was the right call.

It is a different call today, and will be different again in 2027. AI capability is improving fast enough that any tooling commitment is now a bet on which curve is steeper: the buy partner’s incremental AI delivery, or the operator’s own build option as the underlying tools become generally available.

Sam Altman has made this point to founders for years. Build assuming the models keep getting better, not assuming they stay still. He reckons 95% of founders should bet on continued improvement; most bet the other way and end up as the “OpenAI killed my startup” meme. The advice travels to a sophisticated CRE operator unchanged. What is impossible to build in-house this year may be routine next year. The work an operator treats as its distinctive edge may be buildable in-house the year after.

None of this says build everything, or that buying today is wrong. It says the buy decision now needs making with one eye on whether what you are buying is still the right thing across the life of the commitment.

For now, an operator’s tacit knowledge is proprietary. The consortia are the most visible mechanism that changes it. Forward-deployed engineers embedded across enough client firms observe how each one operates, then aggregate what they see into a standardised method the whole industry can buy. Every deployment makes the next operator’s edge a little less distinctive. The window between proprietary and commoditised is open now, and the cost of building tools to capture your knowledge in your own systems before it shuts is falling fast.

DIRECTION OF TRAVEL

Put the map and the three moving positions together, and the takeaway is easy to state and uncomfortable to act on.

The field has five players now, not the two it had when the only question was build or buy. The labs own the road. The consortia are industrialising the middle. The incumbents hold the systems of record but have lost the free pass that let them coast. Operators and startups contest the deep country, where the real systems of work get built.

So the question is not whether AI changes your corner of CRE. It is which of the five you are, and where on this ground you can win. Most of the industry will treat that as too early to act on. By the time they do not, the positions will have been taken.

The Wizard is buying road. The land is shifting faster than the road is being laid.

The best new PropTech is a CRE company

Fifteen years of 'buy don't build' was the right answer. AI has flipped it.

Fifteen years of ‘buy don’t build’ was the right answer. AI has flipped it. The most consequential new property technology of the next five years will be built inside the firms that need it, not sold to them.

I have written about PropTech, and argued with PropTech founders, for three decades. The assumption underneath the whole industry has been that CRE firms need technology, but cannot build it. That assumption is structurally breaking, and not for the reason most commentators are reaching for. The cause is more interesting than build cost alone: the ‘firm’ itself is becoming something different, and the people inside it are becoming something different. What follows is what that means for which PropTechs survive, and for what your firm should be doing inside its own walls.

Executive summary

The PropTech industry has spent fifteen+ years on one assumption: CRE firms needed software they could not build themselves. That assumption is now breaking. Build cost has collapsed for domain experts who hold the relevant knowledge, and the working life of the typical CRE professional is reshaping around agent orchestration in a way that makes building feel adjacent to using. Together those two shifts have inverted the ‘buy don’t build’ default. The middle layer of PropTech, the workflow tools and reporting overlays and generic AI-for-X products, is now contestable from inside the buyer. Data networks, embedded operating systems and deeply specialised regulatory tools remain on the buy side. PropTech founders need archetypes that survive this shift; CRE firms need to recognise that their next competitor for cognitive workflows is increasingly themselves.

WHAT HAS ACTUALLY CHANGED

The argument rests on two things happening at once. They reinforce each other, but they need to be seen separately to be understood properly.

The build cost has collapsed
A senior surveyor with access to a properly configured reasoning substrate, a small library of firm-specific skills, and an afternoon free can now produce working tooling that a Series A PropTech would have needed eighteen months and three engineers to ship. Run a ‘Layer 1’ deployment inside most firms that have done the data work, and the demonstration sits on a desktop within days or a few weeks.

That collapse changes the economics of in-house build in a way the industry has not yet absorbed. The old buy-versus-build calculation assumed building required engineers the firm did not have, infrastructure it did not run, and time it could not afford. All three lines have shifted. Two of them have shifted profoundly. The CRE firm that runs the numbers today reaches a different answer than it did in 2020.

That is the supply-side change. It is real and increasingly understood.

The firm is becoming something different
The more interesting change is what the firm itself is becoming, and how that affects who can build what.

Over the next three to five years, the working life of the typical CRE professional will reshape around agent orchestration. The substance of their job, not a layer on top of it. The discipline is curation. The day’s work is orchestration. Scope the task, configure the agent, validate the output, decide what to do with it. That is a different kind of professional, doing a different kind of work, embedded inside a different kind of firm.

Which matters for the buy-versus-build question. The muscle memory of curating an agent is adjacent to the muscle memory of building one. The capability ladder has fewer rungs than it used to. Curating output is the entry-level discipline. Codifying workflow is the next. Building internal tooling is the next after that. Each rung is a smaller step than it was, and the people climbing them are already on payroll.

What this means in practice is straightforward and largely unannounced. Knowledge workers are beginning to build their own software. Not full enterprise systems. Not the platforms procurement signs annual licences for. The specific cognitive tools they need to do their work better tomorrow than they did today. A property lawyer encoding her firm’s approach to lease abstraction as a skill. An asset manager configuring a covenant-watch coworker that flags exposure shifts across the portfolio. A development director building an agent that runs first-pass appraisals against the firm’s house assumptions. None of them call this ‘building software’. They are.

The same shift is visible in the startup ecosystem. The domain expert who used to need a technical co-founder, a contract dev shop, or eighteen months of runway can now be the hybrid technical founder. The surveyor who knows exactly why ARGUS produces unreliable reversionary valuations in a rising yield environment can build directly against that problem. The fund manager who has watched three IC processes fail at the same point can build the fix herself. Both phenomena have the same cause. Both have the same consequence for the buy-versus-build question.

This is the organisational shift that makes the inversion structural rather than situational. The professional class in CRE is not being replaced by AI. It is being restructured around it. That restructuring is the precondition for in-house build to feel natural rather than foreign.

One caveat worth holding in mind. This describes firms that have done the organisational work to support agent orchestration. Most have not yet. For firms still running on conventional structures, deploying AI on top produces automated dysfunction at machine speed, not compounding capability. The thesis applies most strongly to firms that have crossed the organisational threshold, less to those that have not, and not at all to firms that refuse to. The middle layer of PropTech does not die uniformly. It dies at the frontier first, with the customers vendors most need to retain.

For commercial real estate, this restructuring maps directly onto Quadrant B (see previous newsletters) of the CRE Automation Matrix: ‘verifiable cognition’, where firm-specific tacit knowledge gets encoded into firm-specific agents that handle the high-value cognitive work the firm actually does. Quadrant B is where the durable differentiation lives, because the knowledge being encoded is yours and not your competitor’s.

It is also where the buy-versus-build calculus tilts most sharply toward build. The value of Quadrant B work is precisely what makes it un-standardisable. Any vehicle that promises Quadrant B at scale will, by its economics, push toward generalised patterns that work across many firms. The patterns that work specifically for you have to be built by people who know what ‘specifically for you’ actually means. Mostly that is you.

One further dependency. Quadrant B sits on top of the lower layers of the stack. The data foundation has to be reasonable. The reasoning substrate has to be deployed. Grounded retrieval has to be working over firm data. None of that is exotic in 2026, and none of it is free either. The firm that wants to do meaningful Quadrant B work in-house needs to be comfortable with most of the layer cake, not just the top tier. The full architectural argument is in *CRE AI Is a Layer Cake*. The short version: sequencing matters, and the firms that try to start at Quadrant B without the lower layers in place produce the productivity theatre they were trying to avoid.

WHO MAINTAINS WHAT

The obvious objection to the argument so far. Prototypes are cheap. Enterprise capabilities are not. Maintenance, governance, audit, security, integration. None of these has been collapsed by AI to anywhere near the degree first build has. Anyone who pretends otherwise has not actually shipped software inside a regulated firm.

This is half right. The objection assumes the in-house build resembles the old in-house software project: bespoke infrastructure, custom code, internal devops, full-stack ownership. That is not what serious firms are building.

What they are building sits in the upper layers of the stack, on top of platforms the model providers maintain. Configured agents. Skills libraries. Projects. Grounded retrieval over firm data. The model, the orchestration, the security baseline, the inference reliability. All of those are maintained by the platform, not by the firm. The firm maintains the configuration, the prompts, the data connections, the access controls, and the management structure that curates the lot.

The shape of that management structure matters more than most operators realise. The firms doing this well will have built a distributed fluency model: domain experts as builders inside business functions rather than seconded to a central AI team, stewards governing a skills library treated as institutional memory, evaluation infrastructure detecting drift before it bites. The talent-dependency problem of ‘one analyst built it, nobody understands it after she leaves’ is real for firms that have not done this organisational work. It is largely engineered out for the firms that have.

That is a different maintenance problem to the one PropTech sales teams have spent fifteen years warning operators about. The cost is real. The shape is manageable. Any firm that can run a quality management system for its asset data can run one for its skills library and its agent configurations. Where deeper maintenance is genuinely required (bespoke integrations, regulated workflows with full audit chains, mission-critical systems), you bring in partners who manage those workloads professionally. The same way you bring in audit, legal or quantity surveying for the work you do not staff internally.

What you do not do is build random demo-quality apps and pretend they are enterprise systems. That route was always a bad idea and it still is. The category of build worth doing is the category that lives in your firm’s distinctive judgement and benefits from compounding inside the firm. Quadrant B work. Not everything. Not infrastructure. Not commodity workflows. The cognitive workflows where the value sits in your house view of the world.

AI has not made governance optional. It has made ownership newly plausible.

THE INVERSION

The default has flipped.

For fifteen+ years, the right answer to a CRE technology question was almost always ‘buy, not build’. The reasoning was solid. Building required engineers the firm did not have, infrastructure the firm did not run, and time the firm could not afford. The PropTech industry was the rational outsourced answer to a problem the industry was not equipped to solve internally. 

That answer was right. It is no longer right for a meaningful slice of what PropTech used to sell.

The slice that has flipped is the middle. Workflow tools that automate cognitive labour. Reporting and dashboard products. Analytics overlays that summarise what other systems already hold. Document-handling utilities, IM drafting accelerators, lease review tools, and most of what is currently described in pitch decks as ‘AI for [X workflow]’. These products were defensible when the alternative was a serious in-house dev project running into years. They are increasingly indefensible when the alternative is two analysts, a Claude subscription, and a fortnight.

The default has not flipped everywhere. Three categories remain firmly on the buy side. Data primitives and networks (CoStar, CompStak, the major brokerage networks) cannot be replicated in-house at any cost, because the moat is contributed data the firm does not own. Embedded operating systems (Yardi, MRI, RealPage, ARGUS) carry years of accumulated process logic, regulatory certifications and integration depth that no internal team can recreate inside a sensible budget. Niche cognition tools with deep regulatory specificity (UK SDLT optimisation, US 1031 exchanges, German Wohnen rules, BNG tracking) live in markets too narrow and too specialised for any single firm to justify in-house work, and the specialist who serves the whole market amortises the rule library across all of them.

Outside those three categories, the default has moved. The answer is now ‘build the bits that are uniquely yours; buy the bits that aren’t’, and the boundary has shifted to give ‘uniquely yours’ a much larger surface area than it had five years ago.

This sounds like a small adjustment, but in reality it is a structural one. The PropTech industry was built around an assumption that the buyer could not be the builder. For the next five years, the buyer increasingly is the builder. The competition for most middle-layer PropTech is no longer another PropTech start-up. It is the CRE firm’s own internal capability, supplied by people who already work there.

Both sides of the market have to absorb this.

WHAT SURVIVES

The inversion clarifies rather than kills PropTech.

What remains genuinely defensible falls into four archetypes. Three exist outside the inversion entirely. The data, infrastructure and specialist categories no in-house build can replace. The fourth is the shape new PropTech takes when the inversion lands: operator-built tools that get spun out.

1. Data primitives and networks
Multi-sided businesses where the moat is contributed or observed data that no single participant could assemble alone. CoStar, CompStak, MSCI/RCA, the major brokerage transaction networks. AI makes genuine network-contributed data more valuable, not less, because it becomes the structured ground truth every AI agent in the industry consumes. However, the category is narrower than the named brands suggest. Data that was merely hard to assemble, involving proprietary scraping, manual research teams, ops you used to need scale to fund, is now increasingly easy to reconstruct. Some businesses currently classed as data primitives sit on the weaker side of that line.

2. Embedded operating systems
Yardi, MRI, RealPage, ARGUS, Tramps, Qube. Workflow lock-in across years, regulatory certifications, complex switching projects, integration depth no internal team can recreate. AI strengthens the incumbents who layer it on top of their existing footprint, because the existing footprint is the asset.

A frontal AI-first challenger PMS is unlikely to win. However, an agent-native operating layer that starts at a specific incumbent failure point, captures workflow-specific data, and gradually becomes the system of intelligence above the system of record is a different proposition, and remains plausible. Particularly in various niches. Mid-market BTR. European multifamily compliance. Single-jurisdiction lender reporting nobody serves well. The lateral entry, not the frontal attack.

3. Niche cognition and compliance tools
The regulatory and structural complexity that generalist AI gets systematically wrong, and that no single firm can justify tracking in-house. UK SDLT optimisation for cross-border investors, German Wohnen rules, US 1031 exchanges and LIHTC modelling, MEES compliance, the UK service charge code, biodiversity net gain, embodied carbon under RICS WLCA, SFDR and EU Taxonomy reporting.

The moat is two-sided: deep edge-cases a generalist team would get wrong, and a moving regulatory target the specialist amortises across hundreds of customers. Ideally founded by a domain expert who personally encountered the problem and now codifies it. The CRE firm trying to track this in-house discovers what ‘compliance debt’ looks like.

This category sits firmly on the buy side. The specialists who serve the whole market amortise the rule library and the regulatory tracking. You would not.

4. Operator-built tools that get spun out
The route to market that probably defines the next five years. Tools built inside major CRE firms (Blackstone, Brookfield, Hines, Greystar, Landsec, and their equivalents) for their own use, validated against billions of AUM, then spun out as commercial products. Operator credibility. Workflow validation against assets that actually exist. Pre-built distribution through the originating firm’s network.

Spinouts work best when distribution is non-competing: a different asset class, geography, or market segment from the originating firm. Vicinitee, which I founded and co-developed with British Land and which Equiem later acquired, ran on exactly that logic. Most of the credible new PropTech I expect to emerge between now and 2030 will come through this path, rather than through VC-backed SaaS founded outside the industry.

The pattern underneath the four
The four archetypes share a test. The PropTech that survives is the kind an incumbent cannot easily extend into, an in-house team cannot easily build, and a generalist AI cannot easily replicate. Most of these pass two. The best pass all three. Anything that fails all three is exposed, regardless of pitch deck or capital raised.

For a tech founder, the archetype tells you whether the idea has a future. For a CRE firm, it tells you what is still worth buying, and what to start building.

HOW INCUMBENTS GET OUTMANOEUVRED

The first two archetypes look unassailable. Data primitives like CoStar and embedded operating systems like Yardi have spent decades accumulating moats that no challenger can match head-on. The conventional wisdom says do not attack them, route around them.

The conventional wisdom is right. There are three credible ways through.

1. The over-the-top intelligence layer
Features accelerate a step in an existing workflow. Agents own the workflow and make the step structure obsolete. That distinction matters most where the workflow lives across multiple incumbent systems.

The CRE technology stack is a mess. Yardi, MRI, RealPage, ARGUS, CoStar, and dozens of point solutions. Each holds part of the picture. None can credibly offer the integrated reasoning layer above the lot, because each has an interest in privileging its own data. A vendor whose business model depends on its ecosystem cannot honestly treat its own data as one input among many.

That structural conflict is the opening. An intelligence layer that ingests data across legacy systems, resolves their inconsistencies, and provides unified reasoning is something the incumbents cannot match without cannibalising their own subscription revenue. The reader who has watched a quarterly portfolio review knows the operational gap intimately. The data lives in seven systems. The decision needs all seven.

There is a caveat worth flagging. MCP and standardised agent protocols are compressing the integration work this play depends on. The intelligence layer was a moat in 2024 because connecting systems was hard. It is increasingly available as infrastructure rather than as a defensible product. The play still works, but more as a capability CRE firms build for themselves than as a venture-scale category. The buyer-is-the-builder logic of this whole piece bites here too.

2. Attacking specific incumbent failure points
Frontal attacks on embedded operating systems are nearly dead. Lateral entries are not. The incumbents have structural blind spots, and challengers who concentrate on a single underserved segment can build credible positions before the incumbent notices or chooses to respond.

Mid-market BTR. European multifamily compliance. Single-jurisdiction lender reporting nobody bothers to do well. Niche operator categories where the incumbent product is genuinely poor, the customer base too small to interest the incumbent but too important to leave underserved. The play works because the incumbent’s economics push it toward broad horizontal capability, not toward deep vertical excellence. The gap is permanent. The opportunity is not.

3. Services-as-software
The most underappreciated of the three. Buyers do not want software. They want the work done. The traditional PropTech model sold a tool that helped a customer do work. The services-as-software model sells the completed work itself, with AI doing the heavy lifting inside the vendor.

The customer receives the maintained compliance evidence file, the lender reporting pack, the rent reconciliation, the planning submission draft. The customer sees outcomes, never software. The vendor takes the professional risk and prices the output rather than the seat. The model bypasses incumbent OS systems entirely. It competes for the result rather than the workflow.

The reason this is durable in CRE specifically: most of the high-value work is regulated, professional, and accountability-bearing. AI can do the cognitive labour, but somebody has to sign off. A services-as-software vendor that takes that liability captures margin no SaaS product could justify. The closest analogy outside CRE is the rise of in-house legal teams backed by AI-augmented external counsel: clients pay for the outcome and the accountability, not the tool that produced them.

What the three share
Each play routes around the incumbent’s strength by attacking from a direction the incumbent cannot credibly defend. The over-the-top layer attacks the closed ecosystem with openness. The failure-point play attacks horizontal breadth with vertical depth. Services-as-software attacks workflow ownership with outcome ownership. Each is a route through territory the incumbent built without considering this kind of competition could exist. AI is what makes all three economic. The incumbent has the data and the workflow. The challenger now has the cognition.

SO WHAT NOW
If you are running a PropTech, the question is whether your archetype is in the four. If it is, the question is whether you pass at least two of the three tests. If it is not, the question is how quickly you can move into one that is. The conventional answer is to raise more capital and grow into defensibility. The current answer is that capital does not buy you past any of the three tests. Adding scale to a category the inversion has hollowed out is a more expensive way to discover you have built the wrong thing.

If you are running a CRE firm, the question is which of your differentiating workflows you have allowed to live inside a vendor’s product. The default has moved. The vendor lock-in you have been managing was the right risk to take in 2018. It is the wrong risk to be carrying in 2027. The work of pulling Quadrant B back inside the firm is unglamorous, slow, and unavoidable. It also compounds. Buying rarely does, in the workflows where your judgement is the asset.

There is a larger question the industry has not started asking out loud. The PropTech category was built around venture capital and the assumption that scale follows product. If most of the surviving archetypes are not venture-scale, what happens to the category as a category? My current view is that PropTech contracts as a separate VC asset class over the next five years, and value capture migrates from venture investors back into the operator firms that now have the means to build for themselves. The category does not disappear. It gets absorbed.

The best new PropTech is a CRE company. Sometimes this is a literal claim about a tool spun out of a real estate firm with a £20bn portfolio behind it. More often it is the smaller claim that the most consequential property technology of the next five years was built inside the firm that needed it, by people who already worked there, and never reached the market at all.

For fifteen years, the smart firms bought what they could not build. For the next five, the smart firms will build what others would have sold them. The firms that see this first will spend the time quietly compounding the capability that the rest of the industry will eventually have to procure from someone.

Probably them.

Are you in the wrong half of Real Estate?

The 4-quadrant framework for surviving the hollowing out of the industry’s analytical core.

Last week we covered Pierson Ferdinand: 270 partners, no US associates, AI doing what juniors used to do. That was the firm-structure scale of a dynamic that has been building for two years. This week the capital-deployment scale arrived, and arrived with two frontier AI labs at once. The convergence forces a question on the industry: is real estate still one career? It plainly is not. The bifurcation has begun, and it is going to define the next ten years of who gets paid, who gets promoted, and which firms still exist in 2035.

EXECUTIVE SUMMARY

The first week of May 2026 produced three apparently separate signals: MetaProp Labs released a public catalogue of CRE-specific AI Skills; Anthropic announced an enterprise services firm with Blackstone, Hellman & Friedman and Goldman Sachs; and OpenAI reportedly raised over $4bn for a PE-backed deployment company carrying a reported 17.5% guaranteed annual return for its sponsors. Read together, they describe a single dynamic: AI is industrialising the codifiable, financial-instrument layer of CRE far faster than it is changing the physical-asset layer. This piece offers a two-axis framework: site-specific dependence and named professional accountability. The result is four quadrants with materially different futures. The strategic implication is that single-career framing is now actively misleading. Locating yourself accurately is the precondition for everything else.

THE WEEK THE BIFURCATION BECAME UNMISSABLE

Three things happened, and they look unrelated until you notice they aren’t.

MetaProp Labs published a public catalogue of CRE-specific AI Skills: machine-readable, downloadable, free to install across Claude, ChatGPT, Copilot and Gemini. Deals, asset management, leasing, accounting, legal, investor reporting. It is the public-tier expression of a much larger movement: codifiable industry expertise, packaged for portable consumption, available to anyone with a browser and an API key.

Anthropic announced a new enterprise services firm with Blackstone, Hellman & Friedman, Goldman Sachs and a consortium of further alt-asset managers. The firm will place AI engineers inside portfolio companies to bring Claude into core operations. Blackstone, for the avoidance of doubt, owns more commercial real estate than anyone else on earth.

The same week, OpenAI reportedly raised over $4bn from Brookfield, TPG, Bain and others to launch a competing deployment company, valued at $10bn, majority-owned by OpenAI, with a reported 17.5% guaranteed annual return promised to its private equity backers over five years. If accurate, that detail is the tell. It is a financial-engineering product as much as a technology rollout. The return is not earned on growth. It is earned on compression.

Three artefacts. Three scales: public-tier skills, consortium-tier deployments, capital-tier returns guarantees. One underlying dynamic the industry has not yet named clearly enough.

THE PUZZLE

The intuitive readings are familiar. The optimistic version: AI is industrialising in CRE, the pie is getting bigger, mass-market access to institutional-grade thinking is finally possible. The pessimistic version: the pyramid is breaking, most current jobs disappear, the industry is heading for the same compression as investment banking.

Both readings are incomplete. They share an assumption the events of this week ought to make untenable: that CRE is one industry on one trajectory.

Look at the three artefacts again. The MetaProp Labs catalogue is overwhelmingly aimed at the analytical layer of CRE: deals, asset management, investor reporting, accounting. Pierson Ferdinand is a law firm, a pure financial-instrument-layer professional services business. Blackstone’s deployment of Claude across BREIT, BPP and the European logistics platforms will hit the analytical and investor-reporting layers first, well before it touches a Manchester PRS scheme’s leasing operation or a Birmingham office’s facilities team. The same dynamic that is industrialising on one side of CRE is barely visible on the other.

Real estate has stopped being a single industry on a single trajectory. It is two industries that share a name. The week’s news is the moment that becomes impossible to ignore.

THREE REASONS TO READ THE NEWS CAREFULLY

Before going further, three caveats. None of them changes the structural argument. All of them change how to read the announcements themselves.

First: A press release is not an execution. Enterprise software is full of multi-billion-dollar consortium deployments that produced expensive PowerPoints and not much operational change. The pattern is consistent: capital and brand-name partners assemble around a real bottleneck, eighteen months of building, then collision with the operational reality of getting hundreds of mid-market companies to actually change how they work. Roughly 70% of large-enterprise digital transformation programmes miss their stated objectives, and the most common failure mode is precisely the one this consortium structure produces: capital-and-mandate-driven rollouts pushed down through portfolio companies whose operating teams didn’t ask for the system, don’t trust it, and quietly resist it. CRE operations are particularly exposed because the work is local, embodied, and depends on tacit knowledge that headquarters does not have.

Secondly: The model providers and their customers have inverted incentives. The frontier labs make money when their tools are used more: more tokens, more agents in production, more model calls per workflow. The customer makes money when business outcomes are produced with fewer tokens, smaller models where they suffice, and the minimum viable deployment that delivers the result. A deployment company majority-owned by the model provider, staffed by engineers measured partly on AI consumption, has a structural conflict of interest with the customer’s cost-efficiency goal. Forward-deployed engineers sound like service. They are also a sales channel, and a quiet conduit for everything those engineers see while they are inside your operation. There is more to say about this. It will be the subject of next week’s piece.


Thirdly: There is a financial-engineering smell to all this that will not sit comfortably inside an industry rediscovering itself as operational. Real estate has spent the last five years remembering that it is an operational business: the rise of operating partners over fund managers, the mainstreaming of BTR and life sciences and hospitality-led residential, the shift from spreadsheet returns to building-level performance. An AI deployment story sold by Wall Street, packaged with a guaranteed-return wrapper, marketed through portfolio mandates, is going to land badly with the people who actually run buildings. They are right to be cautious. The framework that follows holds regardless of whether any specific consortium succeeds, because the thing commoditising the codifiable layer of CRE is the model itself, available to anyone with an API key and a SKILL.md file. The consortium is downstream of that.

CODIFIABLE AND ACCOUNTABLE

Some knowledge travels in a SKILL.md file. Some doesn’t. The difference is what is being commoditised this week and what isn’t.

Codifiable knowledge is the kind you can write down as a procedure: comp set construction, T-12 normalisation, cap rate triangulation, debt sizing, lease abstraction, IC memo first drafts, valuation models, variance analysis, lender reporting templates. It travels well. It can be packaged, shared, and run by an AI agent with the right context. The MetaProp Labs catalogue is a public library of exactly this kind of knowledge. The new Anthropic-Blackstone and OpenAI-Brookfield deployment firms exist to industrialise the same kind of knowledge inside large portfolios at speed. Codifiable work is directly substitutable.

Non-codifiable knowledge is what Aristotle called phronesis: practical wisdom built case by case, mistake by mistake, consequence by consequence. The investor who senses a deal start to break in a way the spreadsheet does not yet show. The planning consultant who knows which conservation officer in which authority will sympathise with massing exceptions. The retrofit coordinator who knows which contractor on her list can actually deliver the airtightness numbers the model assumes.

Phronesis is not magic, and it is not immune to AI. It is formed through consequence. It comes from seeing real buildings, real tenants, real contractors, real planning committees, real lenders and real mistakes. AI can support this work, document it, search precedent around it, and remove much of the administrative drag. What AI cannot do is become the named person who has lived through enough cases to know when the written procedure is about to fail. Non-codifiable work is indirectly leveraged by AI: the administrative layers around it get stripped away, output rises, and the burden of named accountability sits on fewer humans, more squarely. That accountability becomes scarcer because output is increasing faster than trusted sign-off.

The asymmetry between these two kinds of knowledge is the engine of the bifurcation. Direct substitution on one side, indirect leverage on the other. Where a CRE role sits on that spectrum determines almost everything about its trajectory.

GOOD, BETTER, BEST

Skills, using Anthropic’s Agent Skills methodology, can be deployed at three tiers, and each tier compresses a different layer of work at a different speed.

Good is the public, free, generic skill. The MetaProp Labs catalogue. A SKILL.md file you download and install, designed to work on standard inputs with institutional-standard procedure. It gets a small landlord 80% of the way to institutional-grade thinking on a routine task. Genuinely valuable, historically unavailable. Compresses the floor of the analytical pyramid first and most violently.

Better is the private, firm-tuned skill. The same procedural backbone adapted to a specific firm’s data, conventions, templates, and house view. Built and maintained internally by someone who owns it as an asset. Compresses analyst headcount by 60-80% on the tasks it covers, and creates one new role: the person who builds, curates and improves the skill library. This tier is where mid-tier firms either build a moat or fail to.

Best is the engineered, agentic, system-level deployment. Skills composed into orchestrated workflows with memory, evaluation harnesses, monitoring, exception handling, sub-agents and feedback loops, integrated with the firm’s data substrate, observable enough to be trusted with consequential decisions. This is what the new Anthropic and OpenAI services firms have just been capitalised to build. It is, in practical terms, a moat for the largest sponsors, and a serious build problem for the operationally disciplined mid-tier firms that prefer to construct their own version internally rather than rent it from a consortium.

Three tiers. Different costs, different defensibility, different effects on different parts of the industry.

THE TWO AXES

The framework rests on two axes. Both are jurisdiction-neutral, although the regulatory texture varies: RICS in the UK, the appraisal regime in the US, Sachverständigenwesen in Germany, expert agréé in France. The structural axes hold regardless.

Axis 1: site-specific dependence. How much does the work depend on local, physical, asset-level, tenant-level or regulatory context that cannot be fully reduced to a document set? The answer ranges from “screen only” to “cannot be done off-site”.

Axis 2: named professional accountability. How much of the role’s value comes from a human or institution putting their name, licence, indemnity, balance sheet or regulatory standing behind the answer?

Two axes. Four quadrants. The bifurcation runs through them.

QUADRANT A: HIGH SITE, HIGH ACCOUNTABILITY

Building surveyors, planning consultants, building control, retrofit coordinators, fire engineers, conservation architects, MEP engineers with statutory sign-off, certified energy assessors, technical due diligence professionals. The protected professions of CRE, where the work is anchored to a specific building, a specific authority, and a specific named person who has to stand behind the answer.

This quadrant is the most resilient of the four, and arguably the strongest position in the entire industry. Demand is growing, driven by the largest infrastructure programme in human history: the decarbonisation of the built environment. The UK alone needs to retrofit 28 million homes and over a million commercial buildings to net zero by 2050. Add building safety remediation post-Grenfell, the conversion of redundant office stock to residential, the planning system’s chronic understaffing, the AI data-centre and logistics build-out, the housing crisis. None of this is being automated away.

The formation pipeline in this quadrant is also intact, almost by accident. A young surveyor learns by walking buildings with a senior surveyor. A planning consultant learns by sitting in committee meetings. The apprenticeship is embedded in physical work that AI does not displace, which means phronesis still develops through the channels that have always developed it. AI shows up here as augmentation: better measurement, faster reporting, fewer errors on routine paperwork. The role gets more leveraged. Viability stays.

Career advice for this quadrant: yes, with confidence. Build AI fluency to multiply your output, but do not mistake the AI for the job. The job is still the building, the local context, the regulation, and the named responsibility you sign your name to.

QUADRANT B: HIGH SITE, LOW ACCOUNTABILITY

Property managers, leasing agents, facilities managers, project managers below sign-off level, residential operators, hospitality-led workplace teams, tenant experience leads, asset operators in BTR, life sciences and logistics. The operational backbone of the industry.

This quadrant is resilient but compressing in its administrative layer. AI takes the paperwork fast: lease renewal correspondence, work-order triage, basic tenant queries, scheduling, rent collection chasing, comp gathering. What it leaves alone is the relational, operational, and physical work: walking the asset, talking to the tenant, supervising the contractor, negotiating with the local authority, handling the moment when something goes wrong at 2am. The role gets redefined toward the human-intensity end.

This is where the asymmetric outcome lives. Generic property management businesses compress. Operationally distinctive landlords expand. #HumanIsTheNewLuxury was always pointing here. Hospitality-grade residential, life sciences operations, experiential retail, members’ clubs, curated workplace: these buildings need more people per square foot than today’s mass-market institutional landlords employ, because the per-asset relational density is higher.

Career advice for this quadrant: yes, with the right operator. The serious operational landlords of the next decade win. Generic property management businesses lose. Pick carefully.

QUADRANT C: LOW SITE, HIGH ACCOUNTABILITY

Senior valuation partners signing for secured lending and fund reporting, IC partners with named responsibility, debt structuring at partner level, fund managers with regulatory accountability, named investment recommendation signers, partners signing audited reports. The senior judgement layer of the financial-instrument business.

This is the Pierson Ferdinand quadrant for CRE: partner-heavy, junior-light, AI-supported. Resilient at the top, compressing in the middle, breaking at the bottom. The senior roles survive because the legal and regulatory structures of the industry require named human accountability. As AI floods the world with analytical output, the scarce resource becomes the qualified human willing to take fiduciary responsibility for the answer. That role gets more valuable in the short run.

But the formation pipeline is breaking. The seniors of 2040 were supposed to be formed through the analytical apprenticeship that produced today’s seniors: ten years of comps, models, memos, IC support, supervised exposure to consequential decisions. AI is hollowing out exactly that apprenticeship faster than firms have built a replacement for it. Pierson Ferdinand has decided, openly, that someone else will train the lawyers it eventually hires. Most firms quietly cutting graduate intake and slowing trainee programmes are echoing the same choice without naming it.

Career advice: yes if you are already senior, very risky if you are trying to enter via the traditional pyramid. The traditional pyramid no longer reliably leads to seniority because the rungs have been removed. Entrants need to find one of the firms consciously over-investing in formation against the sector trend, or build seniority through an unconventional path: founder roles, specialist boutiques, family offices, in-house at a sophisticated owner-operator. The era of “join a big house, learn the trade, make partner in twelve years” is closing on this quadrant. It hasn’t fully closed. It is closing fast.

QUADRANT D: LOW SITE, LOW ACCOUNTABILITY

Analyst pools, research, junior IC support, model maintenance, comp pulls, BOV production, deck building, capital markets junior tiers, junior asset management analytics, lease abstraction, draft variance commentary, junior fund accounting, junior valuation modelling. The codifiable middle of the financial-instrument layer.

This is where the pyramid is breaking fastest and most visibly. The MetaProp Labs catalogue is principally aimed here. The new deployment firms are designed to industrialise here. Brynjolfsson, Chandar and Chen at Stanford’s Digital Economy Lab found a 16% relative employment decline for workers aged 22-25 in AI-exposed occupations through 2025, with wages in those roles actually rising. Fewer, more experienced workers doing more. This is not CRE-specific evidence, but it describes exactly the labour-market pattern one would expect if codifiable entry-level cognitive work is being automated first.

There is also a geography to this. Quadrant A is a local moat: the building stays, the planning officer stays, the regulatory regime stays, and the value sits in the named professional embedded in that local market. Quadrant D is a global commodity: a junior CRE analyst’s work is now competing in a global labour market that includes Mumbai, Manila, and any AI agent with an API key. The bifurcation is therefore also a geographic asymmetry. A is locally protected. D is globally exposed.

The career question follows directly. There is no stable long-run position in Quadrant D doing the work as currently structured. Three to five years is the maximum viable time at this altitude before the role’s value has fully migrated to the agents performing it. Drifting in Quadrant D means waking up at 35 with a CV full of work that an agent now does cheaply, instantly, and well enough.

But the quadrant offers three constructive paths out, and the framework should name them clearly.

The first is accountable judgement: build the AI fluency, the relationships, and the domain depth to break into Quadrant C before the formation pipeline closes. The traditional path, now compressed into a shorter window than it was for previous generations, but still real for those who move fast.

The second is relationship-led origination: reposition toward the parts of the financial-instrument layer where the work is fundamentally about sourcing, trust, and client relationships rather than analytical production. Capital raising, deal origination, investor relations, specialist brokerage. These roles sit nominally in low-site territory but are protected by the relational density AI cannot replicate.

The third is AI-enabled workflow ownership: move up the stack from producer to architect. Stop being the person who writes the lease abstract or the BOV; become the person who designs, curates, and continuously improves the firm’s internal skill library, eval sets, monitoring systems, and orchestrated workflows. This is the betterand best tiers of the good/better/best ladder applied inside your own firm. It depends on deep tacit knowledge of how this firm - its data, its templates, its edge cases, its house view - actually operates. That tacit knowledge is its own kind of non-codifiability. It is also, not coincidentally, exactly the work that RIRA’s verifiable cognition layer points toward, and one of the few genuinely scarce roles AI is creating rather than destroying. It deserves more weight than this framework can give it in passing. We will return to it next week.

The framework’s instruction to Quadrant D is therefore not “transition out”. It is “move up, sideways, or up the stack - but do not stay still”. Drifting is the only path with no future.

THE TAM ARITHMETIC

A counter-argument to all of the above is the one a thoughtful reader will raise here: even if the per-role labour content collapses, the market for CRE services is going to expand by 10x as AI brings white-glove services to mass-market clients. Total industry employment grows. The displacement is real but absorbed.

The argument is partially right and substantively wrong. A stylised example makes the point.

Imagine a UK retrofit advisory market with 2,000 specialists today, serving 10,000 assets a year at five-figure fees. Tomorrow, AI-enabled mass-market retrofit advisory could serve a million assets a year at three-figure fees. The market expands by 100x. But the labour intensity per asset has collapsed. If the mass-market service requires one human per 2,000 clients (because most of it is automated), that’s 500 humans serving the million assets. Total advisor headcount: a quarter of what it was. Total market revenue: meaningfully larger.

The pie expanded. The headcount that serves the pie shrank. This pattern shows up in robo-advisory in wealth management, in Squarespace versus the bespoke web design industry, in TurboTax versus high-street tax preparers. Total category spend rose. Total category employment fell.

Worse, the new TAM is rarely captured by the firms that currently serve the old TAM. The mass-market version of strategic asset management for small landlords will not be sold by JLL. It will be sold by a PropTech platform, possibly one that does not yet exist, with a payroll a fraction of the incumbents whose work it commoditises. TAM expansion is real. It does not automatically rescue the incumbent labour model. In most technology transitions, the new demand is captured by new operating models, new distribution channels, and much smaller teams.

The bigger pie is real. It is not the same as more jobs.

WHAT THIS MEANS FOR YOU

Locate yourself accurately on the two axes. Do not guess. Do not be optimistic.

If you are in Quadrant A (surveying, planning, retrofit, building safety, fire engineering, technical due diligence), you are in the strongest position in the industry. Build AI fluency to multiply your output, deepen your local market knowledge, and recognise that what you have is structurally valuable in a way most of your colleagues in other quadrants do not. Do not waste this position by treating it as ordinary.

If you are in Quadrant B (operational asset management, leasing, facilities, residential operations, hospitality-led workplace), pick your operator carefully. The serious operational landlords of the next decade will employ more people per square foot than today’s mass-market institutional ones do, because the per-asset relational density is higher. Generic property management businesses will compress. Operationally distinctive ones will grow.

If you are in Quadrant C (senior fiduciary roles in the financial-instrument layer), your individual role is durable for a decade or more. The problem is what comes next behind you, and whether your firm is doing anything serious about formation. Ask the question. If the answer is “we have outsourced training to AI”, you are working in a firm that is harvesting the last generation of seniors and not replacing it.

If you are in Quadrant D (analyst pools, junior research, model maintenance, deck building), you have a planning horizon of three to five years for the work as currently structured. Three real options: climb into Quadrant C, reposition into Quadrant A or B, or move up the stack into AI-enabled workflow ownership inside your current firm. The longer you drift without choosing one, the harder all three become.

WHAT THIS MEANS FOR FIRMS

The framework changes how a firm should think about itself, particularly if it operates across multiple quadrants. Most large CRE firms do.

The same firm cannot run a single formation model across all four quadrants. A firm with a Quadrant A surveying business and a Quadrant C investment business is, in capability development terms, two firms in one suit. The surveying business should keep training the way it has always trained: real buildings, supervised work, deliberate exposure to the kind of mistakes that build judgement. The investment business needs an explicit redesign of its formation pipeline because the apprenticeship that used to produce its seniors is gone. Treating them as one career path was always a polite fiction. It is now a strategic vulnerability.

Firms straddling Quadrants C and D have a particular problem and a particular opportunity. The problem: the C quadrant depends on a junior pipeline that the D quadrant is industrialising out of existence. The opportunity: cross-subsidising deliberate phronesis development from the firm’s other revenue lines may be one of the few defensible moats in a world where the codifiable analytical layer is commoditised. Pierson Ferdinand has chosen the opposite strategy. Quietly, most large firms are also choosing it without naming the choice. The firms that consciously do the opposite, that over-invest in formation, will produce the seniors of 2040, and will be paid premium prices to provide them to the firms that did not.

Build versus buy on AI deployment is also an open question, and the consortium narrative is trying to close it prematurely. My hypothesis: the operationally disciplined mid-tier players who build their own better-tier capability incrementally, around their own data, with their own engineers, without the financial-engineering overhead and without the model-provider conflict of interest, will compound advantages that the consortium-led buyers will struggle to match. That is a hypothesis. The next 24 months will test it.

The firms that take this path will need the people described in Quadrant D's third option to do the work. Both halves of the prescription are the same prescription, seen from opposite ends of the org chart.

WHAT THIS MEANS FOR THE INDUSTRY

The bifurcation needs to be named. The single-career, single-firm-strategy, single-industry framing currently dominant in CRE commentary is leading people into wrong quadrants by accident. Smart juniors are walking into Quadrant D because that’s where the entry-level jobs are visible, without understanding the floor is dropping. Mid-career professionals are betting on relational-sector expansion to absorb their displacement, without understanding the expansion will be captured by a different category of firm than the one they work in. Firms are running formation programmes designed for an undifferentiated career path that no longer exists.

This is also a policy and institutional question, not just a firm-level one. The IPF, INREV, ULI, BBP, RICS, RIBA, BCO have all set industry conventions before; they could set capability-development conventions now. A modern guild for AI-fluent retrofit professionals, or for phronesis-development in the financial-instrument layer, would be more useful to the next generation than another conference programme. There is more to say about institutional response and state policy, and it deserves its own piece. For now, it is enough to say the collective-action problem is real, the voluntary bodies are well-suited to address it, and most of them are not yet trying.

THE CHOICE

Two industries. One name.

The events of the past week made the split visible at three scales: public CRE skills at the catalogue level, consortium-tier deployment companies at the enterprise level, and a 17.5% return guarantee at the capital level. AI is not hitting the whole industry equally. It is breaking the codifiable apprenticeship on one side of CRE while augmenting the embodied, accountable professions on the other.

So the important career question is now this. Which part of real estate are you actually in?

If your work is codifiable, low-accountability and far from the asset, the floor is dropping. If your work is site-specific, relational, regulated or consequence-bearing, the tools are getting better and your leverage is rising.

The dangerous answer is the comfortable one: “I work in commercial real estate.” That is no longer specific enough.

The Pyramid Has Already Broken

Six months ago this was a hypothesis. Now it has 270-plus partners and no juniors.

In November 2025 I published ‘Human + Machine Organisational Architecture’. The argument was that capable AI would break the talent pyramid that knowledge-intensive firms have run on for a century, by destroying the economics of the junior tier without replacing the developmental function it served. I described a delayed catastrophe: feast in years 0 to 3, famine from year eight onward, when seniors retire and there is no internal bench to replace them. The piece was a hypothesis. As of this month, May 2026, it is now visible in the operating model of a fast-growing US law firm, and the macro evidence is catching up fast. So the time for treating this as a future risk has run out.

EXECUTIVE SUMMARY

Pierson Ferdinand is showing what the new model looks like: 270-plus partners, no US associates, and AI doing much of the work juniors once performed.

The firm-level logic is rational. The system-level logic is dangerous. If every firm cuts the junior tier and relies on AI plus laterals, the lateral market eventually runs out of laterals. The profession enjoys a short productivity feast, then discovers it has stopped manufacturing senior judgement.

The answer is not nostalgia for the old pyramid. It is a new Human + Machine architecture: AI for routine execution, humans deliberately trained in phronesis, and firms designed around the scarce capabilities the post-commodity economy will still reward.

But that cannot be solved by individual firms alone. Capability development is becoming a collective-action problem. Firms that train juniors bear the cost; firms that cut juniors can buy the benefits later. So the next stage of AI adoption requires a new social contract: professional bodies, training systems, reporting standards, and possibly tax or levy structures that reward the deliberate production of human capability rather than the short-term extraction of AI productivity.

The pyramid has already broken. The question now is whether we build a better architecture in its place, or strip-mine the human-capability base on which the next economy will depend.

THE THESIS, RECAPPED

Six months ago, in ‘Human + Machine Organisational Architecture’, I argued that the traditional pyramid in knowledge-intensive firms had two functions doing different work. The economic function: cheap junior labour subsidising expensive senior expertise. The developmental function: juniors learning by doing over eight to twelve years until they themselves became seniors. Capable AI obliterates the first function. A junior analyst at £50–70K is now dramatically more expensive than the AI agent that can do the same data extraction, drafting, and modelling for under £10K a year. The obvious optimisation, cut the juniors, destroys the developmental function. Result: feast for three years, famine ten years out.

The proposed alternative was a deliberate three-layer architecture: AI handles routine execution; humans focus on judgement at every level from day one; and the system itself - workflows, captured expertise, simulation libraries - becomes a unit of competitive advantage in its own right. The piece closed with an analogy. The traditional firm was a clockmaker’s workshop where apprentices polished gears for years before being trusted with assembly. The Human + Machine firm should be a flight simulator: AI handles the routine mechanics, the apprentices practise complex landings under close guidance from the start, and mastery is reached in half the time.

That was the argument. Six months later, it needs sharpening, and broadening. Here is why.

WHAT SIX MONTHS HAS DONE

In early November 2025, Claude Code did not exist in the form it does now. Opus 4.5 was on the horizon. Agentic AI - systems that take action, not merely produce text - was a research demo more than an operational reality. Today all of that is shipping in production. The category of ‘routine cognitive work’ has expanded faster than any reasonable forecast had it. First-draft documents, comps analysis, scenario modelling, code generation, contract review: AI is at parity with mid-level professionals across a widening band of bounded, reviewable cognitive tasks, and improving on a quarterly cadence.

That changes the timeline. My original framing had years 0–3 as productivity surge, 4–7 as hidden erosion, 8–12 as capability crisis. With this acceleration, the compression is significant. Firms making structural decisions today will see the consequences by 2032, not 2037. The pipeline-collapse mechanism is no longer a logical deduction. Recent work by Brynjolfsson, Chandar, and Chen (‘Canaries in the Coal Mine?’, Stanford Digital Economy Lab, October 2025) documents a roughly 16% relative employment decline for workers aged 22–25 in AI-exposed occupations. Crucially, wages in those roles rose. The pattern is unambiguous: firms are using fewer, more experienced workers. The junior tier is being thinned at exactly the rate the original argument predicted.

And then there is Pierson Ferdinand, which makes the case concrete in a way no statistic can. Founded in January 2024 with no offices and no associates in the US, it has grown from 130 partners at launch to over 270 across 26 markets in just under two years. It uses Harvey AI to do most of the work that, at a traditional firm, associates would do. Asked why, the co-chairman Michael Pierson said the firm could deliver legal practice without ‘having to train the associates on our clients’ pound’.

This is the cleanest articulation I have seen of the structural argument. The clients are no longer subsidising the apprenticeship pipeline because the firm has decided not to run one. The economic logic is impeccable. The developmental logic - what happens when these partners retire and there is no one inside the firm who has spent ten years learning how to actually practise law - is left to the next generation to solve.

A legal recruiter quoted in the same piece worried that junior lawyers might find it hard to develop professionally without being in an office setting. He was identifying a problem. He was missing that, for Pierson Ferdinand, it is not a problem. They have decided, deliberately, that other firms will handle development and they will hire mid-level laterals when needed. They have a ‘junior partner’ tier - but it starts at year five, meaning someone else trained the lawyer for the first half-decade.

This is the buy-not-build strategy applied to human capability. It works as long as enough other firms are still building. The moment they stop, and Pierson Ferdinand’s success will accelerate that moment, the strategy collapses. We will return to that.

THE BIGGER FRAME

The original piece established the supply-side case: how firms preserve and develop human capability under AI pressure. That case stands. What I want to add now is the demand-side case, which sits underneath it and changes the purpose of the exercise entirely.

This is where the argument stops being merely defensive. If AI commoditises routine cognitive production, then the premium shifts toward the work where human presence, trust, taste, care and consequence still matter. The economist Alex Imas makes this argument through what he calls the relational sector: the parts of the economy where the human element is inseparable from the value. It is also the economic foundation of my #HumanIsTheNewLuxury thesis.

In compressed form, Imas argues that as AI drives down the cost of commodity cognitive work, real incomes rise. As real incomes rise, people do not simply buy more of the same. They shift toward different categories of value: hospitality, care, education, therapy, craftsmanship, curated experience and trust-mediated services. Imas calls this the ‘relational sector’. His argument, backed by the structural-change literature and by behavioural research on mimetic preferences, is that as AI automates the commodity economy, this is where employment and expenditure migrate. Not because anyone designs it that way, but because human preferences are nonhomothetic: rich people want different things, and what they want most is precisely what automation cannot fully provide.

If you have followed any of my writing on #HumanIsTheNewLuxury and #SpaceAsAService, this will sound familiar. The CRE-specific version of the relational-sector argument is exactly what I have been writing about for years. The buildings and places that thrive will be the ones that deliver human intensity, experience premium, trust and care, identity and status, and durable community gravity. The asset classes built around these properties - hospitality, members’ clubs, curated workplace, healthcare, experiential retail, premium hospitality-led residential - are the relational sector translated into bricks and mortar.

Putting the Human + Machine architecture next to the relational-sector argument changes the purpose of the exercise. You are developing human capability because that is where the next decade’s economic value is going to live. Pipeline preservation is the by-product, not the goal. The juniors you are training are being positioned for a market that values them.

This connects supply and demand into a single argument: build people who can deliver what AI cannot, in places designed around what AI does not replace.

PHRONESIS

The original piece used the word ‘judgement’ a lot. I want to swap it out for a better one: ‘phronesis’. It is Aristotle’s term for practical wisdom: the kind of capability that cannot be taught from a manual, that lives in pattern-matching against accumulated experience, that is built case by case, mistake by mistake, consequence by consequence. The doctor who departs from protocol because something feels wrong. The investor who senses a deal start to break in a way the spreadsheet does not yet show. The negotiator who reads the moment to push or concede.

I am borrowing the term from a recent article by Harvey Lewis (’From Hierarchy to Judgement’, LinkedIn, April 2026), which is worth your time. Lewis’s argument is that hierarchy in knowledge-intensive organisations did more than route information: it was the institutional technology for producing phronesis. Apprenticeship in medicine, the Inns of Court, military command, the trading floor - all of these were ways of putting capable people into supervised contact with consequential decisions, repeatedly, until the pattern library was built.

The cognitive science backs this up. Gary Klein’s research on expert decision-making, military commanders, ICU nurses, chess grandmasters, firefighters, shows that experts do not compare options exhaustively. They pattern-match against a library of remembered cases, run a quick mental simulation of whether the first plausible option will work, and act. Phronesis is a case library that gets built over time, accumulated through real exposure to particulars under conditions where mistakes have visible consequences. There is no shortcut to building it.

This matters because it tells us what we are actually trying to develop, and what we lose if we do not. If your apprentice spends two years validating AI outputs without ever bearing real consequences, without being on the wrong side of a failed deal, without having to defend a recommendation that turned out to be wrong, they will not develop phronesis. They will develop AI-supervision skills. The aviation evidence is sobering: pilots who spend most of their time supervising autopilots get materially worse at manual flight, by margins large enough to raise risk significantly. Nicholas Carr’s The Glass Cage covers this in detail. The skills you do not use, you lose.

The implication for the architecture is direct. Junior development must include consequential decision-making from day one. AI supervision and quality-checking matter, but they are auxiliary skills. The flight-simulator analogy still holds, but only if the simulator includes scenarios where bad decisions produce visibly bad outcomes that the apprentice has to live with.

THE THREE-LAYER ARCHITECTURE, SHARPENED

The architecture stands. Three layers: an Execution Engine where AI handles systematisable routine work; a Phronesis Development layer where humans are placed into dense, supervised, consequential decision environments; and a System Stewardship layer where humans design AI workflows, externalise expertise and continuously improve the system. What changes in v2.0 is the framing of Layer 2.

Layer 2 is where phronesis is deliberately cultivated. The original framing of ‘humans focus on non-routine work’ is correct as far as it goes, but it under-specifies the design. Cultivating phronesis requires three properties the original piece did not quite name:

• Artificial Density: Apprentices in a traditional firm see perhaps twenty serious deals in three years. A well-designed simulation library combined with structured deal exposure can put them through two hundred. Higher case-volume per unit of time compensates for lower consequence-weight per case. This is the flight-simulator element done seriously: a deliberately denser environment than the natural workflow ever produced.

• Supervised Real Stakes: Density alone produces pattern-matchers who have never been wrong. Real-stakes work, actual deals, real clients, real consequences, under close supervision is what builds phronesis rather than knowledge. The mentor’s job is to intervene before consequences become terminal, while still letting the apprentice be wrong, find out, and feel it.

• Deliberate Friction: Some decisions in your firm are slow on purpose. The slowness is load-bearing: it creates space for challenge, dissent, second opinions, and the formation of the case library that becomes phronesis. Automate those decisions and you remove the development substrate, even if the organisation tells itself oversight remains. The discipline of v2.0 is deciding which decisions stay slow on purpose, and protecting them from optimisation pressure.

The other two layers are largely as before. Layer 1 (The Execution Engine) should be transparent, showing reasoning to preserve learning opportunities. Layer 3 (System Stewardship) becomes more important as the firm matures, because the AI workflows themselves require curation and continuous improvement. Competitive advantage will accrue through creation; it will last through curation.

THE OBJECTION I’VE TAKEN SERIOUSLY

The strongest objection to v1.0 came implicitly through the aviation literature and explicitly through Lewis’s piece: not all human capability can be effectively cultivated in the presence of automation that is almost always right. The experience of being wrong, finding out, and bearing the consequence is the substrate. If AI handles the cases where being wrong is detectable, what is left is the cases where being wrong is undetectable until catastrophic. That is exactly the wrong domain in which to develop phronesis.

There is also what Lewis calls the ‘Mythos effect’: the human tendency for bounded competence (’this model is better than me on this’) to slide irrationally into general deference (’this model is better than me’). The mechanism is psychological. It kicks in long before any actual general superiority is established. Once your team starts treating AI output as a firm conclusion rather than a draft, the development substrate is gone, regardless of what the org chart says.

These are real challenges to v2.0; the architecture has to confront them rather than wave them away. The three responses above - density, real-stakes supervised practice, deliberate slowness - are the answer, but they require active design and constant maintenance. The default path of any organisation is for the slow decisions to get faster, the supervision to get thinner, and the deference to get more general. Resisting that requires explicit institutional commitment.

The honest claim about the 4–6 year timeline I proposed for senior capability is this: it is a hypothesis. Traditional development took 8–12 years partly because the bottleneck was the economics of running a junior tier, not the inherent difficulty of the learning. Compress the development through density, real-stakes mentorship, and good simulation, and the timeline can plausibly halve. Whether it does is an empirical question, and the firms running this experiment over the next decade will answer it.

BEYOND CRE

I write about commercial real estate, but everything in this piece applies, with local variation, to every profession and every form of knowledge work. Law (Pierson Ferdinand), accounting, consulting, banking, surveying, architecture, engineering, medicine: structurally similar pyramids, structurally similar phronesis-development requirements, structurally similar substitution pressure at the junior tier. The CRE-specific cut, investment, asset management, development, brokerage, property management, is one application of a general framework.

If you are reading this from a different profession, the question to ask is the same. What is the phronesis your seniors actually possess that the firm depends on? Where does it currently come from? And what happens to that source under AI substitution? If you cannot answer those three questions concretely, you are running on accumulated capital that nothing is replenishing.

THE SOCIAL CONTRACT

This is where firm-level strategy hits its limit. Even if individual firms read this argument and act on it, the aggregate dynamic is a textbook prisoner’s dilemma. The firms that invest in capability development bear the cost. The firms that cut juniors and rely on AI plus mid-level laterals can buy the benefits later. Pierson Ferdinand’s strategy works precisely because other firms are still developing the lawyers it may eventually hire. If enough firms adopt the same model, the lateral market does not become more efficient. It starts to empty.

That is why this cannot be treated purely as an organisational-design problem. It is also a social-contract problem. Capability development has always been partly subsidised by clients, firms, professions and the wider economy. AI makes that subsidy visible by giving firms a way to stop paying it. Left to itself, the market will let AI strip-mine the labour market the way unregulated industrial revolutions have stripped other resources - efficiently, profitably, and with the bill arriving a decade later.

The interesting fixes are mostly private. Two ideas worth taking seriously: teaching-firm models, where clients knowingly retain a ‘junior academy’ at a discount in explicit exchange for training capacity, with the lower productivity acknowledged in the fee structure rather than smuggled into the hourly rate; and modern guilds - cross-firm consortia that pool the cost of intensive simulation environments, share mentorship infrastructure, and set membership standards that signal something real to clients, peers and capital. Both have working precedents. The Investment Property Forum, INREV, ULI and the Better Buildings Partnership show that voluntary, peer-set, industry-led bodies can shape CRE practice more effectively than regulators ever have. What is missing is one tuned specifically to the capability question.

The state’s role is to tilt the economics rather than to inspect outputs. The lever is to make training the more profitable choice, not the more virtuous one: a reshaped Apprenticeship Levy that returns more than firms put in when they train juniors under genuine supervision; enhanced tax relief for capability-development spending that meets defined criteria; capability reporting alongside ESG so that boards and capital allocators can see who is investing in the next generation and who is harvesting the last. None of this is regulation in the inspection-and-punishment sense. All of it works through price. Whether nudges of this kind are strong enough to close the gradient against firms that capture very large margin gains by avoiding training entirely is a real question; if they are not, harder instruments will eventually come into view, but they should be the third move, not the first.

The policy architecture, and the business models, belong in a separate piece, and I will write one. The principle is simple enough: if AI allows firms to extract productivity from a professional system without replenishing the human capability on which that system depends, the market will not self-correct in time. The framework is firm-level. The fix cannot be only firm-level.

THE CHOICE

Pierson Ferdinand has made one. They have decided, openly, that they will not train the next generation, and that someone else will. The choice is rational at the firm level and corrosive at the system level. Every firm that follows them is making the same choice, including, importantly, every firm doing it by default rather than by design. If you are quietly cutting your graduate intake, slowing your trainee programme, or relying on AI to do work that juniors used to do without thinking about what comes next, you are echoing Pierson Ferdinand. You have simply not named the choice you are making.

The Human + Machine architecture is the alternative: an offensive positioning for an economy in which human capability, properly developed, properly deployed, properly stewarded, is going to be the scarce input that commands the premium. You build the people. You design the system around them. You take the harder route because the easier one ends in a place no one wants to be.

The pyramid has already broken. What you build in its place is, for now, still a choice.

CRE AI Is a Layer Cake

Why most CRE AI programmes are building the wrong layer first - and what the right architecture actually looks like in 2026.

There is a view that keeps surfacing in CRE: in boardrooms, at conferences, and in the LinkedIn comments under my posts. It says the useful AI work in property is the analytical kind. Forecasting rents. Predicting prices. Scoring risk. Optimising portfolios. The kind of work that takes a data science team, an eighteen-month build and a great deal of patience. That view is not wrong exactly. It is just, for most firms and most people, extraordinarily beside the point. And its practical effect is to make the industry wait for a perfect model, a clean dataset, a specialist team, while the tools that could be improving your work today sit unused.

Executive summary

Modern AI in CRE is a layered architecture, not a single system. Five layers: data foundation, reasoning substrate, grounded retrieval, workflow automation, and bespoke analytical AI. Value compounds from the bottom up, and the largest gains for most firms sit in the middle, in the layers most often dismissed as ‘just chatbots’. The mistake the industry keeps making is to start at the top, ignore the middle, and then wonder why two years of programme spend has produced nothing useful. This piece walks through each layer: what it is, what good looks like, and how to sequence the stack so that value shows up in weeks rather than years.

THE OVERVALUED LAYER

Go to almost any CRE AI discussion and the conversation drifts, within minutes, towards forecasting, scoring and prediction. This is Layer 4 of the modern stack. It is real. It is occasionally necessary. It is also the smallest, most selective, and least commonly needed layer for the vast majority of firms.

Most people in commercial real estate do not spend their days building predictive models. They spend their days doing something else entirely.

WHAT YOU ACTUALLY DO ALL DAY

Look at your own diary this week. Your time has been going into:

• reading documents

• extracting facts

• comparing clauses

• drafting memos

• reviewing evidence

• checking compliance

• finding precedents

• assembling reports

• moving information from one place to another

That is your working life. It is not Layer 4 work. It is Layer 1 and Layer 2 work. It is exactly where modern AI systems are already excellent. And it is exactly where most of the value is currently being left on the table.

‘JUST CHATBOTS’ IS THE EXPENSIVE MISTAKE

A vocal faction in CRE still dismisses LLMs as ‘just chatbots’. The phrase gets repeated at conferences, on analyst calls, and in the mouths of executives who have never seriously tried to use one. That dismissal is the single most expensive AI mistake currently being made in our industry.

It mistakes the interface for the system. A conversational prompt is how you talk to Layer 1. It is not what Layer 1 is. What Layer 1 is, when deployed properly, is a reasoning engine with access to the firm’s templates, the firm’s house style, and the firm’s codified approach to work like covenant analysis, lease abstraction and tenant financial assessment. Set inside a well-configured Project, with the relevant documents in context, it drafts. It compares. It extracts. It reviews. It challenges. It does in ninety minutes what a competent analyst would take three days to do on the current deal, and the analyst can see which documents in the Project it drew on. And that is Layer 1 alone, before you connect it to the firm’s wider archive or build an agent on top of it.

Dismissing that as ‘chatbots’ is like dismissing a modern steel mill as ‘just a furnace’. The interface is not the point.

SO WHAT IS THE POINT?

The point is that CRE AI is a layered architecture, and most of the value sits below the predictive layer that gets all the attention. Five layers. Each one compounds on the one below it. Skip the foundations and the upper layers wobble. Build the foundations properly and the upper layers become cheap to add.

And there is a second architectural shift that matters even more. A few years ago, analytical AI was the primary system, with natural language interfaces added on top. In modern architecture, that has been inverted. The reasoning layer is primary. The analytical AI is a tool the reasoning layer calls when it needs specialist computation. Claude does not forecast rental growth. It calls a forecasting model that does, interprets the output, contextualises it against other evidence, and drafts the narrative for the analyst. This is not a cosmetic difference. It changes what the system can do, what it costs to build, and where the value lives.

What follows is the architecture, layer by layer. What each is. What good looks like. Where it is genuinely needed. Where firms systematically over-invest. And how to sequence the whole thing so that you get working capability in weeks rather than years.

LAYER 0: DATA FOUNDATION (THE PREREQUISITE)

Start here, or everything above it fails.

What it is. The data hygiene, indexing and access infrastructure that makes everything else in this piece possible. It is not AI. It is the precondition for AI.

What good looks like. People across the firm can locate documents quickly. Reference systems are consistent. Deal rooms are structured. Historical records are digitised and searchable. There is a clear ownership model for data quality. Lease data is not scattered across seventeen PM systems and a shared drive called FINAL_FINAL_v2. Deal documents have consistent metadata. ESG data flows into a system rather than being re-keyed from PDFs each quarter.

What it enables. Everything. Without Layer 0, the rest of the stack either doesn’t work or produces unreliable output with no audit trail.

The reality check. For many institutional CRE firms, the honest starting position is that Layer 0 is the biggest single constraint on AI ambition. Closing the gap is tedious, unglamorous and expensive. It is also unavoidable. The firms that pretend they can skip it produce demos that impress the board and fall apart under real use. The firms that take it seriously spend the first six to twelve months of their AI programme doing the work nobody wants to do, and then their subsequent layers compound properly. This is the single largest determinant of whether an AI programme delivers real value or productivity theatre.

One important qualifier. Layer 0 is a prerequisite for firm-wide capability: the kind of institutional memory and portfolio-level grounding that Layer 2 depends on. It is not a prerequisite for individual practitioners getting on with specific pieces of work. An analyst with the documents for a specific deal in a specific Project can produce real Layer 1 value tomorrow morning, regardless of the state of the firm’s wider data estate. The waiting game that Layer 0 concerns sometimes invite is a mistake. Your firm’s data might be a mess. The documents you need for today’s work are in front of you. Start there.

And a second nuance. Layer 0 work is itself increasingly AI-assisted. LLMs are now genuinely good at extracting structured data from unstructured documents, normalising inconsistent records, and reconciling references across systems. The Layer 0 cleanup and the Layer 1 deployment can proceed in parallel, with the Layer 1 tooling accelerating the Layer 0 work. But the principle holds: you cannot build reliable firm-wide capability on top of unreliable data.

LAYER 1: THE REASONING SUBSTRATE

This is the layer most firms underestimate. It is also the layer where most of the value is.

What it is. A frontier LLM environment (Claude, ChatGPT, Gemini) deployed as the default thinking environment for your analysts, portfolio managers, and operations staff. Not a chatbot bolted onto existing workflows. The reasoning layer that sits underneath everything your people actually do.

A note on tooling. The language of Projects, Skills and configured coworkers is currently most developed in Claude, which leads the market on this architecture as of early 2026. Equivalent or near-equivalent capability is arriving rapidly across the frontier - OpenAI, Google and others - and by the end of this year the architectural pattern will be general rather than vendor-specific. The principles below apply across all frontier providers; the current naming reflects where the tooling is furthest along.

The building blocks at this layer.

Projects hold the context for a specific piece of work. A Project for an acquisition might contain the offering memorandum, the data room index, the comparable transactions, the underwriting template, and the draft IC memo. Everything the analyst is working on lives inside the Project, and every conversation with the reasoning layer is grounded in that context. Projects are the unit of work: not the individual prompt, and not the whole firm’s knowledge base, but the specific thing being done right now.

Skills are codified workflows that encode how your firm does a specific task. A lease abstraction skill knows what fields to extract, what schema to produce, what red flags to surface, and how your templates are structured. A covenant analysis skill knows your firm’s approach to tenant financial assessment. An IC memo skill knows your house style and the required structure of each section. Skills turn institutional know-how into repeatable cognitive workflows that any analyst can invoke. They are the single most underused capability in modern AI systems, and they are where most of the firm-specific value actually lives.

Configured coworkers are persistent personas set up for a specific role: a covenant analyst, a compliance reviewer, a climate risk scout, an investor relations drafter. A covenant analyst coworker is not a workflow you run: it is a colleague you ask. It carries its own instructions, reference materials and behaviours, and it can be called on by anyone in the firm who needs that kind of thinking applied to their current problem. The point of configured coworkers is to turn expertise that currently lives in one person’s head into something the whole firm can access.

What Layer 1 actually delivers. Most of the value most firms need, most of the time. This is the part people systematically under-estimate, because it sounds too simple.

An analyst working inside a well-configured Project, with access to the right skills and coworkers, can draft an IC memo in an afternoon that would previously have taken a week. A covenant review that required three days of manual reading can be compressed to ninety minutes of structured interaction. Compliance flagging becomes a background process rather than a quarterly fire drill. None of this requires custom agents. None of it requires analytical AI. None of it requires a knowledge graph. It requires Layer 0 data and a well-deployed Layer 1 substrate.

Why this matters. Layer 1 is where the compounding value lives. Every new skill, every new coworker, and every new Project template captures institutional knowledge in a form the whole firm can use tomorrow. It is also where firms that ‘get it’ start to pull away from firms that don’t. The gap between the two is already visible inside the firms running my courses. It will become obvious in the market within eighteen months.

LAYER 2: GROUNDED RETRIEVAL OVER FIRM DATA

Now the reasoning layer stops answering from its training data and starts answering from yours.

What it is. A structured retrieval layer that lets the Layer 1 reasoning substrate access the firm’s actual data with proper grounding and evidence chains. This is what RAG (retrieval-augmented generation) does when it is built properly. Your deal rooms, historical documents, lease archives and reporting systems become queryable through the reasoning substrate, with every answer traceable back to source.

Why this is separate from Layer 1. Layer 1 can work with the context you put into a Project, but a single Project cannot contain the whole firm’s history. Layer 2 extends the reasoning layer’s reach to the firm’s institutional memory, while maintaining the evidence-linking that makes the output defensible. When an analyst asks ‘have we ever underwritten a deal with a similar covenant structure?’, the answer comes from Layer 2 retrieval, with citations back to the actual historical documents.

What good looks like. Every answer produced at this layer has an evidence chain. The analyst can see which documents were retrieved, which passages were cited, and how the reasoning layer used them. Nothing is hallucinated, because everything is grounded in retrievable source material. The underlying retrieval technology (vector databases, structured extraction, semantic search) has matured to the point where building this layer is engineering work rather than research.

A note on knowledge graphs. If your firm already has a well-structured knowledge graph, that is a significant advantage at this layer. A knowledge graph captures explicit relationships between entities rather than just semantic proximity, and it can answer queries that vector search struggles with. ‘Show me all assets where the tenant covenant has weakened since acquisition and the CRREM pathway is red’ is a natural query for a knowledge graph and a hard one for pure vector search.

So: are knowledge graphs a good thing if you have them? Emphatically yes. A knowledge graph is a Layer 2 asset that should absolutely be connected into the retrieval fabric and used as the highest-quality grounding source available. The argument is not that knowledge graphs are unnecessary. It is that a firm starting from scratch today should not begin by building one. The modern sequencing is: deploy Layers 0 and 1, build grounded retrieval at Layer 2 using whatever structured data is available, and add a knowledge graph when the limits of simpler retrieval become binding. Not that long ago, a knowledge graph was the only way to get grounded cognition, because there was no reasoning engine to pair it with. In 2026, the reasoning engine exists, and the knowledge graph is one of several options for grounding it. Powerful, but no longer foundational in the way it had to be a decade ago.

LAYER 3: CUSTOM AGENTS AND WORKFLOW AUTOMATION

Now the reasoning substrate stops being something people talk to, and starts being something that runs by itself.

What it is. Task-specific automated workflows built on top of Layers 1 and 2. Where Layer 1 supports an analyst doing their work interactively, Layer 3 runs the workflow end-to-end with defined inputs, defined outputs and defined quality checks. A Layer 3 agent for quarterly rent roll reconciliation takes the source files from the PM system, checks them against the GL, flags exceptions, produces the reconciliation report, and surfaces anything that needs human attention, all without an analyst needing to drive the process.

The tooling. Claude Code is currently the leading example of a development environment for this layer. It is a terminal-based agentic environment in which the reasoning layer can execute code, call APIs, manipulate files, interact with external systems, and run multi-step workflows autonomously under human oversight. Equivalent tooling exists from other providers. The important thing is that Layer 3 work is genuinely development work. It requires thinking about task decomposition, error handling, exception routing, audit trails, and integration with source systems. This is not point-and-click, and it is where the engineering effort of an AI programme starts to become meaningful.

What Layer 3 is for. Repeatable, high-volume, rules-adjacent work where the firm currently spends meaningful human hours on tasks that follow a consistent pattern:

• rent roll ingestion and reconciliation

• covenant compliance checking

• CRREM pathway monitoring

• standardised report generation

• lease abstraction at scale

• due diligence document review against checklists

These are the Quadrant A and easier Quadrant B tasks from the RIRA CRE Automation Matrix (see other newsletters). They are where efficiency gains genuinely compound, because every run of a Layer 3 agent is work that no longer requires human time.

The sequencing trap. Most firms try to start here. They identify a painful workflow, commission an AI project to automate it, and build a custom agent before they have Layers 0, 1 or 2 in place. The results are predictable: the agent works in the demo, fails on edge cases, requires constant re-engineering, and ultimately delivers a small fraction of its promised value. Layer 3 only works when it sits on top of clean data, a capable reasoning substrate, and grounded retrieval. Build the foundations first.

LAYER 4: BESPOKE ANALYTICAL AI

And finally, the layer most people imagine is the whole thing.

What it is. Classical machine learning, statistical modelling, time-series forecasting, numerical optimisation, and other specialist techniques, for the problems where the frontier LLM approach is genuinely insufficient. This is the layer most people imagine when they hear ‘AI in real estate’. It is actually the smallest, most selective, and least commonly necessary layer of the stack.

Where Layer 4 is genuinely needed.

• CRREM-aligned pathway modelling at portfolio level, where regulatory defensibility and reproducibility genuinely matter

• time-series forecasting of operational metrics (energy, footfall, NOI under varied scenarios)

• large-scale portfolio optimisation under complex constraints

• satellite and drone computer vision for physical due diligence and ESG verification at scale, of the kind Kayrros and the space-intelligence cohort are doing for asset-level transition risk

• AML and fraud pattern detection in transaction flows, where false-positive rates need to be tight and the audit trail needs to withstand regulatory scrutiny

• graph-native queries across millions of entities with latency constraints that LLMs cannot meet

• specialist risk models that require reproducibility and regulatory defensibility

Where Layer 4 is not needed, even though firms often assume it is.

• drafting IC memos

• summarising deal rooms

• flagging unusual lease clauses

• checking compliance against a policy document

• producing quarterly reports

• answering questions about the portfolio in natural language

• generating first-draft analysis

All of those are Layer 1 and Layer 2 work. A surprising amount of what is currently branded ‘analytical AI’ in vendor pitches is actually Layer 1 capability dressed up in analytical language. Ask the vendors which layer their product sits in. The good ones will be able to tell you.

A word on the moat question

The most serious institutional objection to everything above is not “but we don’t have the data” or “but this is just chatbots”. It is the moat argument: yes, but Layer 4 is where proprietary advantage lives. Anyone can deploy Claude in a Project. Defensible moats come from specialist models, proprietary data, and analytical depth that competitors cannot replicate. And beside the moat, there is the money argument: Layer 4 is where the prestige fee income sits, where the alpha is meant to live, and where the industry’s quantitative firepower has always been concentrated.

Both arguments deserve a direct answer. In commercial real estate, neither is as strong as it looks.

First, most CRE firms lack the data density that makes serious machine learning worthwhile. Property data is fragmented across PM systems, transaction records and market reports, with limited history, inconsistent taxonomies, and thin comparability across assets. Classical statistics and good old-fashioned data science take you most of the way. ML at scale requires signal depth that most firms simply do not have. The data gap is the reason we have seen so little genuinely predictive AI in real estate despite the tooling having been widely available for a decade. It is not that the industry has been slow to notice. It is that the underlying substrate does not support the ambition.

Second, the share of CRE work-quantum that is predictive-modelling-shaped is small. Rent forecasting, yield prediction and portfolio optimisation matter, but they are a minority of what the industry actually does with its time. The bulk of the work - the reading, extracting, comparing, drafting and reviewing - is Layer 1 and Layer 2 territory. A moat in a small corner of the work is a small moat.

Third, and this is the point that should give any Layer 4 enthusiast pause: there is no Jim Simons of real estate. Renaissance Technologies exists because public equity markets have deep historical data, high liquidity, and genuine arbitrageable inefficiencies. Real estate has none of these at scale. If Layer 4 were where the alpha lived in CRE, someone would have extracted it by now. A decade of widely available ML tooling has not produced a Rentec of property, because the asset class does not support one. This is a structural feature of real estate, not a temporary tooling gap. The assumption that it will change because the models get better is, so far, unsupported by evidence.

Add to this the structural shift of real estate from financial engineering towards operational performance, where the competitive edge is increasingly about running buildings well, serving occupiers well, and reading markets through lived operational contact, and the case that future CRE moats live at Layer 4 gets weaker still. The moats of the next decade are more likely to live in the quality of a firm’s Layer 1 and Layer 2 infrastructure: the depth of its skill library, the calibre of its configured coworkers, the cleanliness of its institutional memory, and the velocity at which it can turn that infrastructure into decisions. That is a different kind of moat. It is also one that more firms can actually build.

How Layer 4 connects to the rest of the stack

Bespoke analytical tools at Layer 4 are called by the reasoning layer, not alongside it. A forecasting model does not produce the IC memo. It produces a forecast. The reasoning layer interprets the output, contextualises it against other evidence, calls additional tools if needed, and drafts the narrative for the analyst. Layer 4 sits as a callable resource inside the Layer 1 substrate, not as a parallel system the analyst has to manually consult and then translate.

SEQUENCING: BUILD FROM THE BOTTOM

A firm starting today should assemble this stack in roughly the following order, with significant overlap between phases.

Months 0–6. Layer 0 work begins and Layer 1 is deployed in parallel. Data hygiene projects start. Projects are set up for current live deals. The firm’s first skills are written, typically covering the two or three most common document-heavy workflows. Initial coworkers are configured. Quick wins start appearing within weeks, not months. This is the phase where the early evidence that the programme is working gets generated, and it matters, because it builds the institutional trust that makes the later phases possible.

Months 3–12. Layer 2 grounded retrieval starts to come online as Layer 0 data becomes usable. Existing knowledge graph assets, if any, are connected. The reasoning substrate begins to reach into the firm’s institutional memory, and the early agents become meaningfully more capable because they have access to firm history rather than just the documents in the current Project. This is where the verifiability story becomes real: every output now has evidence chains back to source material.

Months 9–18. Layer 3 custom agents begin to be commissioned for specific high-value repeatable workflows. These agents are built on top of the Layer 1 substrate and the Layer 2 retrieval fabric, so they inherit the reasoning capability and the grounding infrastructure rather than having to rebuild them. Development effort at this phase is genuine, but it is targeted at specific workflows with clear ROI, not at foundational infrastructure.

Months 18+. Layer 4 bespoke analytical work is commissioned only where specific, identified capability gaps cannot be addressed by the lower layers. Most firms will find that the surface area of genuine Layer 4 need is much smaller than they expected, because Layers 1 through 3 cover most of what they thought they needed specialist analytical AI for.

The critical difference from the AI strategies of five years ago is that this sequence generates value at every stage. A modern deployment generates working capability in the first weeks and compounds from there. The underlying principle (that verifiability, grounding and evidence chains matter) is identical. The path to achieving it is completely different.

WHAT THIS DOES NOT CLAIM

Three honest caveats, before anyone reads more into this than I am actually saying.

It does not claim Layer 1 is sufficient for all use cases. Some firms genuinely need Layer 4 analytical work, and pretending they don’t is as wrong as pretending they need it for everything. The point is that Layer 4 should be a selective, targeted commitment based on specific identified capability gaps, not a default starting point.

It does not claim this sequencing works without Layer 0 data work. The data prerequisite is the most important line in this whole piece. A firm that tries to deploy firm-wide capability on top of chaotic unindexed document storage will produce unreliable output with no audit trail, and the programme will stall within months. Individual practitioners can still make progress on specific deals with specific documents, but the firm-wide compounding does not happen without Layer 0.

It does not claim the human layer is unimportant. None of the architectural sequencing above replaces the need for investment professionals who know how to interpret AI output, challenge it when it is wrong, and make the judgement calls that remain irreducibly human. The entire point of verifiable cognition is that it supports human decision-making rather than replacing it. The governance, cultural and skills work that surrounds the technical stack is at least as important as the stack itself, and arguably more important. Firms that deploy the technology without investing in the human layer end up with sophisticated tools their people don’t know how to use.

AND FINALLY

This is a working artefact. It reflects the current state of frontier AI tooling as of April 2026 and the architectural patterns now emerging as best practice for institutional CRE firms. It will age, some parts faster than others, and should be revisited as the tooling improves.

But the core argument is not going to age. The value in modern CRE AI sits below the predictive layer. It is accessible now. It is layered, and it compounds from the bottom up. The firms that understand that in 2026 will be the firms that still matter in 2030.

You do not need to wait for a model to be trained. You do not need a CTO to green-light a platform. You do not need a data science team. You need to look at your diary this week, identify the five things you do that look like reading, extracting, comparing, drafting or reviewing, and start doing them with Layer 1 tooling tomorrow morning. Open a Project. Drop the documents in. Ask it to do the first task. Twenty minutes, not twenty months.

That is the shortest path from current capability to better capability in commercial real estate. It is not the glamorous path. It is the one that works.