Inside the AI Market Appraisal: three models, one median, and a figure the write-up cannot move

There is no house price model inside the AI Market Appraisal. There is a comparable finder, three frontier language models that are each handed the same evidence and asked for a range, an arithmetic step that takes the middle answer, and a set of checks that throw away any answer which does not stand up. For a tool that produces a figure an agent will say out loud in a vendor's kitchen, the path to the figure is the product. Here is the whole path.

The comparables are assembled before a model is asked anything

Nothing is put to a model until there is evidence to reason from, and that evidence is gathered on a ladder rather than in a single query.

On the live side, the tool searches listings by distance from the subject property at half a mile, then one mile, then three, then five. At each rung, if fewer than six suitable properties come back, it loosens the property type first and the bedroom range second before widening the radius again. It stops as soon as six are found, and keeps the best six.

The sold side runs the same ladder against Land Registry price paid records with an eighteen-month lookback, widening the bedroom band by one only as a last resort. Six are kept there too. That dataset is published in full by HM Land Registry as price paid data, so any completed sale cited in an appraisal can be checked against the public source by anyone who cares to.

Twelve comparables reach the models: six asking prices from a market that is moving now, six completions from the eighteen months behind it.

Similarity is scored, not assumed

Every comparable carries a similarity score, and that score is arithmetic an agent could redo on paper. Each one starts at 100 and loses points for how far it sits from the subject property:

Recency earns points back on the sold side. A completion under three months old gains five points, one under six months gains three.

Nothing is filtered out by the score, and that is deliberate. A weak comparable that is the only thing sold on the street still appears in the report, carrying the score that says so. The agent sees what the models saw, thin evidence included.

What the models are told about the house

The wizard captures the property in structured form: twelve property types, twelve key features, and special features grouped into six categories.

Around that, the property record fills itself in. EPC floor area and energy efficiency score arrive automatically, along with the construction age band and the attributes we hold on garden and parking, plus roof and windows. Tenure and plot area come from Land Registry title polygons, with the plot measured from the polygon geometry rather than typed in by hand.

Photographs are optional. Up to five can be added, and a vision model returns a condition read on what it can see, which helps most where a refurbishment has not yet reached the comparables. Images that our own image service serves as anti-scrape decoys are recognised and skipped rather than described, so a poisoned picture never becomes evidence.

Three models, asked at the same moment

The council is three frontier models called in parallel in one pooled request, each given identical comparables and identical property facts.

The whole council sits inside a forty-five second budget, because an agent in a car park will not wait longer. Parallel also means no model sees another's answer, and nobody anchors on anybody. Three independent readings of one body of evidence is a different thing from one reading reviewed twice.

Each model returns a base price, an expected price and a go-to-market price for a sale appraisal, or an achievable monthly rent for a lettings one, together with its reasoning. The prompts carry explicit prohibitions: a model may not invent a predicted price or estimated value that the evidence does not support, and it may not state tenure, floor level or ex-local-authority status that is absent from the data in front of it.

A verdict has to pass before it counts

Every returned verdict goes through a sanity check before it is allowed near the consensus. A verdict is thrown out if a price is not numeric, if a sale figure comes in under £20,000, if a rent falls outside £100 to £100,000 a month, or if the ordering breaks: the base has to sit at or below the expected, which has to sit at or below the go-to-market figure. A model that returns a go-to-market price beneath its own base price has misunderstood the question, whatever the numbers look like.

Surviving figures are rounded to sensible steps rather than quoted to the pound: £5,000 steps below £100,000, £10,000 steps below £500,000. An appraisal presented as £412,637 would be false precision dressed as rigour, and the rounding stops the tool implying a resolution it does not hold.

The consensus is a median, and the label is the spread

The consensus is the median of the verdicts. Never the mean.

That single choice does most of the work. With a mean, one model that has read a converted flat as a five-bed house drags the whole figure with it, in proportion to how badly it is wrong. With a median, the outlier is outvoted and the answer lands where the other two bracket it. The cost of a model having a bad day is capped.

The agreement label is the other half of the output. It is the largest deviation between the models, reported as strong, good, moderate or weak, and it travels with the figure. Weak agreement is not a failure. It is a signal that the evidence is thin or the property is unusual, and it is precisely the appraisal where an agent should lean hardest on what they know about the street. A confidence score sits alongside it.

A stand-in never enters a consensus that has answers in it

Models time out. Keys expire. Rather than leave an empty slot, the tool substitutes a synthetic stand-in worked out from the comparable average, or from a bedrooms formula where there is nothing better, so the interface stays whole.

What happens next is the rule that matters. Those stand-ins are excluded from the median. A consensus is computed only from verdicts that came back from a model and passed the sanity check, so a failed model cannot vote. Only when no model answered at all does the result fall back on the stand-ins, and in that case it is tagged as demo data rather than served up as a valuation. No agent should be handed a number that no model produced without being told so plainly.

The write-up explains the figure, it cannot move it

The justification a vendor reads is written afterwards, by a separate and smaller model in the Inigo voice, once the consensus has been fixed. Its instruction states that the figure is final. Its job is confined to explaining the comparables and the reasoning behind the range, on a much shorter time budget than the council itself.

Valuing and explaining are different jobs, and handing both to one call is how a tool ends up with prose that quietly re-anchors the number it was meant to describe. Keeping them apart means the write-up can be improved or regenerated without the valuation moving a pound.

Three prices, and the agent's edit is the last word

The output is a range with three points on it, which is how appraisals are discussed on a doorstep anyway:

• Base: the defensible floor, the figure supported by the weight of the comparables.
• Expected: the realistic achievable price.
• Go-to-market: the asking price, with negotiating room in it.

All three are editable before the report is saved, and the agent's adjustments are stored in their own fields, separate from what the council proposed. The record keeps both, which is the point. If you know about the extension mid-build or the buyer pool that has thinned since March, you change the number, and your number is what the vendor sees.

The report, the link and the expiry

A saved appraisal becomes a branded report with its own identifier, tied to the branch and the person who created it. It carries a share link, a one-click PDF rendered through a headless browser, and an email that sends it straight to the client. Expiry is set at seven, thirty or ninety days, or left open. View count and the date it was last opened come back to the agent, so a vendor who has read the report four times is visible before the follow-up call. The same report structure sits behind the embeddable valuation widget and behind valuations that start in the CRM.

The honest limit is that the handoff stops there. There is no automated push from a saved appraisal into the letter tools, the e-signature flow or the valuation lead pipeline. If you want a fuller leave-behind for the appointment, that is the job of the branded market intelligence report, a separate document.

What it costs

Every account gets five free valuations. After that a credit is £4.99, bought through Stripe, or valuations are unlimited on the Professional plan at £199 a month alongside the rest of the agent toolkit. A credit is deducted only after a successful council call, so a generation that fails or times out does not cost anything.

What it is not

Three things this tool does not claim.

It is not a formal valuation. A market appraisal produced by software is an automated estimate to support professional judgement, not a valuation carried out by a registered valuer under professional standards, and it should never be handed to a client as one.

It is not the same engine as the consumer estimate. The instant valuation on the public site is an algorithmic comparables calculation with no language model in it whatsoever. The council described here exists only in the agent tool. Two products, two methods, and it is better to say so than to let the word "AI" blur them together.

It does not publish an accuracy figure. There is no backtest percentage in this post because there is no backtest behind the tool today, and a measured hit rate is the only kind worth printing. What the appraisal offers instead is legible: the twelve comparables, the score against each one, the three verdicts, the spread between them and the arithmetic that turned them into a figure. That is a stronger claim than a percentage nobody can check, and it is the standard we hold ourselves to in Patch, where the scoring formula is published rather than described.

Walk into the appointment with the range and the evidence under it. Change the number if you know better. The tool is built on the assumption that you often will.