Not a blog. These are the positions the whole method rests on, written out in full so you can disagree with them before spending anything.
A deal dies in week five. The restriction that killed it — an Article 4 direction, a break clause, a licensing condition — was findable in week one, sitting in a document nobody had time to read closely. The legal spend is already gone.
A quarter closes and the position gets rebuilt by hand from a managing agent's export, an accountant's ledger and two spreadsheets that disagree about voids. It takes four days and describes a business that has already moved on.
Neither of those is a shortage of intelligence. Both are failures of sequence: the information existed, and it arrived in the wrong order or too late to change a decision. Point the most capable model in the world at either and it will produce a fluent summary of a problem that has already cost you money.
This matters commercially, not philosophically. Information-arrival problems are cheap to fix and the fix is measurable — you can count the days between practical completion and drawdown before and after. Genuine intelligence problems are neither. A firm that cannot tell them apart buys the expensive fix for the cheap problem, which is most of what the last three years of proptech AI has consisted of.
Which is why nothing gets built until the machine is drawn. The map is not a preliminary; it is the thing that tells you whether you have an AI problem at all.
Every firm bidding against you has access to the same models, at the same price, on the same morning. The intelligence is a commodity and it is getting cheaper. Anyone who tells you their model is the advantage is selling you the part that has no moat.
What is not available to them is the accumulated judgement of your firm. The yield you will accept in one postcode and refuse two streets away. The refurb assumptions you have corrected four times because the first three were wrong. How a partner weighs a difficult freeholder against a good number. And the four hundred deals you passed on, each with a reason that never got written down.
That is layer 03 in Institutional Context Architecture, and in almost every firm we look at it does not exist in any readable form. It is held by two people, it is inconsistent between them, and it leaves the building when they do. A firm can be twenty years old and have no written record of how it actually decides.
The commercial consequence is specific. A model with no access to layer 03 gives you general market advice — the same advice everyone else gets. A model wrapped around it produces judgements that look like your best partner's, at a volume your best partner cannot reach. The first is a subscription. The second is an asset on your balance sheet.
Writing it down is also the uncomfortable part, because it surfaces where two partners have been deciding differently for years. Every firm that has done it has found at least one of those.
Most AI business cases in this sector are built on cost. Analyst hours saved, abortive diligence avoided, a number of days of admin recovered. Those savings are real and they are also the least interesting thing about the work, because payroll is a rounding error against debt service. No principal deploying £30m is losing sleep over a few hundred thousand pounds of internal time.
The saving matters for exactly one reason: it pays for the system, which means the decision does not have to be made on faith.
The return is elsewhere. Screening is mechanical work — reading, extracting, checking against thresholds — and it scales five to ten times. Deep underwriting does not, because verification and judgement do not compress; two and a half to three times per analyst is honest. We separate those two numbers deliberately, because most vendors quote the first and deliver the second.
What that multiple actually buys is selection. Examining five thousand opportunities instead of a thousand means buying from the top fraction of a percent of the market rather than the top one percent, with the same team and the same number of closings. On a large deployment, improving average selection by thirty basis points is a permanent annual gain — and unlike a cost saving it does not stop, because better assets lift refinance valuations, which recycles capital faster, which compounds.
The ordering matters and we hold to it in that direction: you deploy more capital, from a better set, faster — and the back office does not grow to do it. A firm that leads with cost is telling you which of those it can actually deliver.
What is actually working in AI for UK property investment firms, and what isn't — written for principals rather than for a marketing funnel.
No sequence, no upsell, no webinar invitation. Unsubscribe removes you immediately and permanently.
Roughly 700 words, every other Tuesday.
Thirty minutes on a call, no deck. The first call is more useful when you arrive with an objection than when you arrive with a brief.
Book the Foundation Session — from £2,000