A deal that used to take me an evening now takes about a minute. I take an address off a portal, and before I have finished my coffee I have a yield, a rent range, a refurb estimate, a read on void risk and a list of comparable sales going back years. Two years ago that was a spreadsheet, three phone calls and a drive past. Now it is a single query.
That speed is real and it is worth having. But I want to be careful about what it actually means, because the story being sold to investors right now is that AI hands you an edge. In a market like student HMOs, it mostly does not. It hands everyone the same edge at the same moment, which is another way of saying it hands no one an edge at all. What separates a good buy from a bad one is the same thing it always was. The boring part just got cheaper.
What genuinely changed
Let me be specific about what AI took off my plate, because it is a lot.
The analysis collapsed. Pulling fourteen rental comparables and turning them into a yield I would defend used to be an afternoon. An automated valuation model now does it in under a second, and the better ones run at error rates below three per cent, tighter than most surveyors will put in writing. I can screen every listed property in a city against my buy-box, filter on yield floor, EPC band and tenure, and only ever look at the handful worth looking at.
Void modelling got honest. The single biggest mistake I see new HMO investors make is underwriting at full occupancy. Student houses empty over summer. Rooms turn over. If you model a seven-bed let at 100 per cent occupancy you are not modelling reality, you are modelling a brochure. AI lets me cost voids room by room rather than house by house, which is how the cash actually behaves.
The refurb maths is instant. Most of my stock needs an EPC lift. Working out what it costs to move a house from an E to a C, and what rent that unlocks, used to mean a builder's gut and a wet finger in the air. Now it is a range I can sanity-check before I have even offered.
And the admin got automated, which matters more than it sounds. From April 2026 any landlord earning over £50,000 has to keep digital records and file quarterly under Making Tax Digital. The threshold drops to £30,000 in 2027. That is a genuine time cost, and software absorbs most of it.
The rules AI has to price in, but cannot decide for you
This is where the student HMO market pulls away from the brochure version of “AI is transforming property”. The supply side of this market is now defined by planning, not by demand. Article 4 directions, which strip out the permitted right to convert a family home into a small HMO, are in force across most major English student cities. Manchester, Leeds, Birmingham, Bristol, Nottingham, Southampton, large parts of London. In those areas you cannot create a new HMO without full planning permission, and permission is frequently refused.
Sit with what that does. It puts a ceiling on new supply in exactly the places students want to live. Existing licensed HMOs in those zones do not just hold value, they become difficult to replace, which is a better thing. HMOs already make up around half of all student accommodation in this country. A planning regime that throttles new ones while demand holds is, for an owner of the existing stock, close to a moat.
An AI tool will tell you a property sits in an Article 4 zone. It will not tell you whether the line moved last month, whether the council is minded to grant on that specific street, or whether the scarcity is worth paying up for. I have walked away from deals the model loved, because I knew the ward was saturated and the council had stopped granting. I have paid over the odds for tired houses the model was lukewarm on, because they were licensed, sat inside an Article 4 area, and would never be reproduced. The model cannot make that call. It does not know what it does not know.
Then there is tenancy law. The Renters' Rights Act came into force on 1 May 2026. Section 21 is gone. Fixed terms are gone, replaced by rolling periodic tenancies. For student HMOs there is a carve-out worth understanding, a new Ground 4A that lets you recover a student property at the end of the academic year, provided you gave notice at the outset. Whether your model is built to handle that, or whether it is quietly assuming a tenancy structure that no longer exists, will never show up in a yield figure.
And compliance has teeth now. Operating an unlicensed HMO carries civil penalties of up to £30,000 per offence and rent repayment orders of up to twelve months' rent. A model can flag that a licence is required. It cannot run the house to standard once you own it. That part is on you.
What still wins
After twenty years the pattern is clear to me. AI is excellent at the parts of investing that are really information retrieval. It is no use at all at the parts that are judgement.
Reading a street is judgement. Knowing that one side of a road lets in a week while the other sits empty is judgement. Pricing the operator risk in a large HMO, the management drag, the fire compliance, the wear, is judgement. Deciding what you would actually pay, and having the discipline to walk when the number is not there, is the whole game, and no model will do it for you, because no model has your money on the line.
What changed is that I can now apply that judgement across far more deals. I used to properly assess maybe five a week. Now I screen a hundred and hand-judge the ten that survive. The funnel got wider. The decision at the end of it is still mine.
The part that actually matters
If there is one thing I would push investors to think harder about, it is not which AI tool to use. They are converging, and most of them are fine. It is the data underneath the tool.
A clever model on stale or thin data gives you a confident wrong answer, and confident wrong answers are how people overpay. What is worth caring about is whether the yield, the comparables, the ownership history and the EPC position are current and complete, refreshed against the registers rather than scraped off a portal six weeks ago. That is the layer I spend my time on now, and it is why we built Propalt the way we did. But the principle holds whatever you use. Judge the data before you judge the deal.
The institutions have worked this out, by the way. Brookfield bought around 1,300 HMO beds and an operating platform in a single transaction last year. When money of that size moves into a fragmented market, it is not because of clever software. It is because supply is constrained, demand is structural, and the data finally lets them underwrite at scale. The same logic is available to a one-portfolio investor. The tools are cheap now. The judgement is still the job.
Sources and further reading
— HMO supply, Article 4 and institutional entry — Allsop
— Article 4 directions across UK student cities — HMO Market Report 2026
— Renters' Rights Act, Making Tax Digital and the 2026 regulatory shift — NRLA
— HMO licensing penalties and occupancy modelling — Quartico
— AVM accuracy and AI underwriting tools in 2026 — V7 Labs
Co-founder of Propalt and a property investor and developer for 20 years
