Property AI Summit guides

How UK property businesses put AI to work in 2026

Two numbers in that list explain most of what is happening in UK property this year. Half the industry intends to adopt AI within the year, and six in ten teams are already reporting burnout. The interest is not really about the technology. It is about workload that keeps rising while the team stays the same size.

That is also why so many people try AI, get a mediocre result and quietly stop. They start with a tool instead of a task, ask it something a colleague would need context to answer, get something bland, and conclude the technology is oversold. The order below is the one that avoids that, and it is the order we use for the build sessions at the Property AI Summit on 17 October 2026.

1. Audit your week before you choose a tool

Step one

Spend one normal week writing down what you actually did, in fifteen-minute blocks, and mark each entry with one of four letters.

• R, reading. Legal packs, leases, surveys, EPCs, management reports, tenancy files, planning documents, long email threads.
• W, writing. Vendor and landlord letters, viewing feedback, contractor briefs, tenant correspondence, listing copy, offer summaries, meeting notes.
• J, judgement. What to offer, which deal to walk away from, whether to take a client on, how to handle a difficult person.
• D, data entry. Moving the same figures between a portal, a spreadsheet and a CRM.

AI is dependable on R and W, useful on D once someone has set it up properly, and no substitute for J. If your week is mostly J, a tool will not give you your evenings back and you should stop reading here. In most property businesses we see, R and W together are the larger share, which is the whole reason this pays.

The audit matters for a second reason. It gives you a baseline. Without one you will never know whether the tool helped, and you will end up arguing about it on instinct.

2. Write your context down once

Step two

The most common complaint is that the output is generic. It is a fair description and a wrong diagnosis. An assistant that has never been told anything about your business answers from the average of everybody else's, because that is all it has.

The fix takes twenty minutes and is reusable forever. Write one page, in plain language, covering:

• What your business does, and what it does not do.
• Your patch, in postcodes, and the stock types you actually deal in.
• Your numbers: typical prices, rents, yields, fees, the figures a new colleague would need on day one.
• The rules you work to, including the ones specific to your council.
• How you write. Formal or plain, long or short, what you never say.

Paste that page at the top of any task that matters. If your tool supports saved instructions, project files or a persistent workspace, put it there instead so you stop pasting. That single page is the difference between advice about property in general and a draft you can send after one read.

3. Supply the UK rules it does not hold

Step three

Most property material these models learned from is American. Left alone they drift towards American defaults, which is how you end up with output referring to escrow, realtors and closing costs in a letter about a Birmingham HMO. The drift is easy to spot when the words are wrong, and expensive when the mechanics are wrong.

These are the UK specifics that decide whether an answer is worth anything, and none of them can be assumed:

The practical rule: never accept a rule from memory. Ask for the source, then check the source. Which is step four.

Three of those mechanics moved recently, and the models learned the old answer. Section 21 is gone, Making Tax Digital is live, and the EPC C deadline is no longer 2028. The three dates AI currently gets wrong, with sources, are here. And in planning, the trap is Article 4: the answer almost every assistant gives about small HMO conversions is wrong in over a hundred council areas.

4. Verify anything that carries a penalty or reaches a client

Step four

This is the step people skip, and it is the one that decides whether AI is an asset or a liability in a regulated business. Three failures account for nearly everything that goes wrong.

Invented sources

A model asked for a rule, a case or a regulation will sometimes produce one that does not exist, with a convincing name and a plausible date. Through 2025 and into 2026 the UK property tribunals saw a run of cases where parties submitted AI-generated case law that either did not exist or did not say what was claimed. Being represented was not protection. Checking was.

Invented numbers

Ask what a property is worth and you will get a number, because you asked for one. It is an average dressed as a valuation. Supply comparable evidence or do not ask the question.

Confident wrong dates

Rules change, and a model has no reliable sense of where its own knowledge stops, so it will not warn you. Anything with a deadline attached, and 2026 is full of them, has to be checked against the source rather than the summary.

A protocol that works, and takes less time than it sounds:

• Source or silence. Any factual claim about a rule comes with a link or a document reference, or it does not get used.
• Two sources for anything with a penalty. Possession, licensing, deposits, anti-money-laundering checks, tax filings. One of the two should be the primary source: HMRC, the legislation itself, the council, or your own professional adviser.
• Human sign-off on anything outbound. Nothing reaches a tenant, a vendor, a lender or HMRC without a person reading it. This is also the only version of the workflow that survives a complaint.

5. Turn it into a system, not a good afternoon

Step five

Most people who try AI have one good afternoon and nothing to show for it a fortnight later. The reason is that a good afternoon is a demonstration, and a system is a habit with the setup written down.

What makes it stick, in our experience of running build sessions:

• One task, not ten. Pick the highest-frequency writing task in your audit. Frequency beats importance here, because repetition is what compounds.
• Run it the same way for two weeks. Same brief, same instructions, same checks.
• Write down the version that worked. The exact wording, and where the output still needs you. This becomes a process another person in your team can follow.
• Then add the second task. Not before.
• Measure against your baseline. The week you logged in step one is what makes the answer a fact rather than a feeling.

That is the whole method, and there is nothing to buy in it. If you would rather do it in a room with people who are doing the same thing, and walk out with the first system already running, that is what the Summit is for.

How to leave any AI day with a system rather than a notebook covers the before, during and after in detail, including the two-week rule that decides whether a training day was worth being out of the business for.

Common questions