A 70-storey City tower, a 2036 letting date, and the variables that decide whether it fills
A 70-storey tower in the City of London, due in 2033 and more realistically 2036. The building may well be right. The harder question is who will be working in it, and how many of them, after a decade of AI progress.
Peter Bill’s EG column this week carried a nice piece of City intelligence. AmTrust, the American insurer that bought Exchequer Court on St Mary Axe for £105m in 2017, wants to replace the seven-storey, 146,000 sq ft block with a WilkinsonEyre tower of around 70 storeys and 1.1m sq ft. Sir George Iacobescu has been hired to shepherd it through planning, and the stated aim is to finish and fill it by 2033. Peter’s verdict on that timetable: “Budget for 10?”
He is probably right. Which makes it a useful thought experiment, because a building conceived in 2026 and let in 2036 is a bet on what work looks like after ten years of AI. The building is in London, but the variables apply to any city with towers in the pipeline. So let’s run them.
WHERE AI WILL BE IN 2036
Start by separating capability from adoption, because they run on very different clocks.
Capability first. METR measures the length of task, in human working time, that AI agents can complete. Between 2019 and 2024 that length doubled roughly every seven months; its January 2026 update put the doubling time since 2023 nearer four. The best models now succeed, about half the time, on software tasks that take a skilled person more than a full working day. I don’t believe the curve stays that steep for ten years, and I don’t need to. Halve the rate and by 2036 agents are handling, end to end, digital tasks that today take a person weeks.
Cost tells the same story. The price of a given level of capability has been falling by something like an order of magnitude a year. By 2036 the model’s share of production work, the bounded and checkable tasks that fill so many office floors, will cost very little. Making a whole business process reliable is a separate, stickier cost.
The open variable is coordination. Doing a task, completing a long workflow and working together as a team are separate capabilities, improving at different speeds. Single agents are moving fast: on Mercor’s APEX-Agents benchmark, built from banking, consulting and legal work, the leading models now score above 80% (leaderboard as of early November). Teams of agents lag well behind; a 2025 study of popular multi-agent frameworks found them failing on 41–87% of tasks.
Alex Imas, Andrew Koh and Shane Legg argue that once AI coordination crosses a threshold, the economics of firm size can change very quickly, and the improvement that tips it may be too small to register on any benchmark.
It is my belief that the threshold is crossed for routine coordination (routing, review, status, handoffs) well inside the decade. Preference and commitment (the question, the relationship, the signature) stay with people, by choice and regulation as much as by capability. A human signature, though, says nothing about how many humans stand behind it.
Adoption is the slow clock. The economists surveyed by Karger, Tetlock and colleagues assign real probability to rapid AI progress by 2030, yet their central growth forecasts sit only modestly above conventional projections. The electric dynamo took thirty years to show up in the productivity statistics because factories had to be redesigned around it. Organisations are the bottleneck, and organisations are what sign leases.
For much routine digital work, I expect the technology to be ready well before the tenants are. A tower completing in 2033–36 will be let while its target occupiers are in the thick of redesigning themselves.
That letting campaign sits inside my 2030–35 danger zone.
FIVE VARIABLES
1. Location: strongly in its favour. Work has a postcode. My working hypothesis is that human-led strategic work, the dealmaking, complex advice, capital allocation and accountability, concentrates in prime HQs near clients, regulators and other firms. A new tower beside the Gherkin, a short walk from Lloyd’s, is about as well placed as an office gets.
2. Scale: the real problem. The same hypothesis says prime will house a narrower, more elite slice of each firm, used intermittently and sized for the peaks. Take a deliberately crude example: a 500-person firm in 60,000 sq ft that gains 40% output per person. If its output stands still it needs about 29% less space. At +10% output, about a fifth less. If output doubles, about 43% more. That assumes space per person holds steady, which it may not: a smaller, more senior workforce tends to want more meeting space and sharper peaks.
With a 40% productivity gain, a firm has to grow its output by 40% just to stand still on space.
So where does a million square feet of demand come from? Mostly from the City’s older stock. Two things then happen at once. Productivity stops business growth turning into new space, which is what I call missing absorption: growth that never becomes occupancy. And flight to quality concentrates the demand that remains into new buildings, leaving holes behind it. The tower can let perfectly well while the market around it weakens.
The optimistic case deserves its due. London could keep drawing new firms, consolidators could take market share, and international occupiers could fold scattered offices into fewer, larger headquarters. Each of those adds demand for exactly this kind of building. The underwriting question is how much expansion and share-taking the scheme needs to offset every existing occupier using less space per unit of output.
3. Work mix: a City split down the middle. The natural occupiers straddle what I call the exposure order: production work goes first, coordination next, preference and commitment last. Specialty underwriting, broking, complex law, restructuring and capital markets sit in finite, bespoke markets heavy on preference and commitment (choosing the question, owning the relationship, signing off the decision). Their own middle offices (claims handling, policy administration, bordereaux (the policy and claims reports agents send insurers)”, compliance, standardised legal work) are the textbook case of exposed work: executing someone else’s question, at volume, in a digital medium where the answer can be checked. Strip out production and coordination roles, keep the client-facing and accountable ones, and you get the senior-heavy column: smaller lettings at higher specification on shorter terms. That suits the quality of the building. It suits neither the quantum nor the lease lengths that finance a speculative tower. And finite markets cut both ways, mind. They may preserve human involvement in the work, but limited demand leaves little room to absorb productivity gains, so the same work may simply need fewer people.
4. Format: a stack of the riskiest unit. If better coordination favours consolidators, and cheaper contracting favours small specialist guilds, the occupier base polarises. My hypothesis is that the dedicated mid-sized floor gets squeezed, so a tower that depends on letting floor by floor to mid-sized occupiers is exposed to precisely that segment. The buildings that suit a polarised market are anchored by a big committed occupier or run as operator-led, shared space selling the moments that need a room.
5. Timing: the letting campaign lands in the danger zone. It also lands alongside another 300m City tower (One London), which is the rivalry Peter’s column is about. And the tenants it chases will be negotiating from strength. A financially healthy occupier restructuring around AI is a bigger threat to occupancy than a stressed one that still needs bodies in seats. The risk was never default. It was renewal.
TWO DECISIONS, NOT ONE
Which brings me to the part of the story that is easy to miss. AmTrust is making two decisions, and they deserve very different verdicts.
The first is securing consent, and it is a clever move. The 2017 price is sunk, although the site still carries an opportunity cost: holding, selling and redeveloping all remain live options. A permission that takes the site from 146,000 sq ft to around 1.1m sq ft, roughly seven times the floor area on the headline figures, creates substantial option value, provided the consented scheme is deliverable. Hiring the former Canary Wharf boss to get it through is money well spent. And on any realistic programme the decision to actually start building is two or three years away, by which point we will know a great deal more about what City occupiers are doing with AI.
The second decision, building a million square feet of largely speculative offices, is a bet on a particular future. AmTrust has said nothing yet about how it would fund or let the tower, so treat this as a scenario to test. It pays off if AI productivity gains in City financial and professional services are absorbed by growth, or if the tower simply takes share from older buildings. It struggles if demand from its target occupiers grows more slowly than their productivity.
If I were advising, I would want four things in place before the first pile went in:
-
An anchor. A large pre-let from an occupier that has already thought hard about its work mix in 2035, ideally AmTrust itself.
-
Floorplates and services that can switch to an operator-run, flexible model with limited disruption and capital spend.
-
Upper floors designed, and appraised, for a change of use if office demand comes in below plan.
-
A start date triggered by evidence, meaning what is actually happening to the work inside comparable occupiers, rather than by the programme.
A flexible, adaptable, well-specified building is the minimum stake for a 2036 office. Every scheme in the City pipeline will promise the same. The advantage will come from knowing what work your tenants will be doing inside it.
THE 2036 TEST
Exchequer Court is one building, but the questions apply to anything you are underwriting with a 2030s letting date. For each target occupier, use three questions as a diagnostic:
-
Is it executing work someone else specified, or deciding what work to do?
-
Is its market deep and standardised, or finite and bespoke?
-
Is its coordination done by people, or by plumbing that software is learning to run?
An occupier that answers “specified, deep, plumbing” may still be a fine tenant in 2028. It is unlikely to want the same floorplate in 2036.
The same variables travel. Run them over the tower pipelines in Midtown Manhattan, La Défense, Frankfurt or Riyadh and the answers will differ. A city whose occupiers serve fast-growing markets can absorb productivity gains that a mature London specialty insurer cannot. A city that has already sent its processing work to capability centres has less left to lose. The questions hold everywhere. The arithmetic is local.
None of us knows exactly where AI will be in ten years. I would bet heavily on the direction and lightly on the date, and a tower is a bet on both.
Peter says budget for ten years. I’d add: budget for a different tenant. And if you want the space let, start learning now what that tenant’s work will look like.