How Estate Agents Use AI to Prepare Valuation Appointments

Use AI for the desk work before you knock on the door: turn the booking notes into a property brief, tidy the sold and listed comparables you've exported into a clean table, draft questions for this particular vendor, and assemble a tailored pre-valuation pack. The price stays yours. AI organises the evidence; it hasn't walked the rooms.

The trap is asking a chatbot to "find comparables" or "value 14 Elm Road". Without data you supply, a general assistant will produce plausible sold prices that may not exist, or quietly mix asking prices with achieved ones. Feed it evidence from sources you trust, such as your CRM, your sold-price data and portal exports, and treat every figure it returns as unverified until you've matched it to a row in the source.

Follow me on Instagram@sagnikteaches

What a strong appraisal needs from the desk beforehand

A valuation appointment is a pitch as much as a price. Vendors often invite more than one agent, and the one who arrives knowing the house, the street's recent sales and why the vendor is moving comes across as the safer pair of hands. Most of that knowledge is assembled at a desk, which is where AI earns its keep.

Connect on LinkedInSagnik Bhattacharya

Here's where the preparation time goes for a typical three-bedroom appraisal, and what changes when AI does the assembling. The minutes are illustrative; time your own for a week before and after.

Subscribe on YouTube@codingliquids

Notice what isn't in the table: the price. Everything AI produces feeds your judgement at the visit; none of it replaces the visit.

Two days before: turn the booking call into a property brief

The brief is only as good as the booking note. Whoever takes the call, whether a negotiator, the office manager or a chatbot that books valuations out of hours (the options are compared in whether an AI chatbot can book valuations for you), should capture the same fields every time: property type, bedrooms, any extensions or conversions, reason for moving, timescale, whether they've found an onward purchase, which other agents they've invited, and anything they mentioned about condition.

Paste that note, plus the text of any previous listing for the property, into the assistant with a prompt like this:

You are helping an estate agent prepare for a market appraisal.
Using ONLY the notes and documents below, write a one-page property brief
with these headings:
1. Property facts (type, bedrooms, bathrooms, floor area, tenure,
   parking, outside space)
2. History (previous listings, asking prices, dates, time on market)
3. What the vendor told us (reason for moving, timescale, onward
   purchase, other agents invited)
4. Gaps: facts we still need to confirm at the visit
Rules: if a fact is not in the notes, write "not known". Do not estimate
or fill anything in. Refer to the vendor as "the vendor", not by name.

NOTES:
[paste booking note and previous listing text]

That takes five to ten minutes including a read-through. For a three-bedroom semi, the reply looks something like this (illustrative):

1. Property facts: semi-detached, 3 bedrooms, 1 bathroom, floor area
   approx. 94 m2, freehold, driveway for one car, rear garden.
2. History: listed in 2019 at $339,000; withdrawn after 11 weeks.
3. What the vendor told us: moving closer to family; found an onward
   purchase, offer accepted; wants to complete within 4 months;
   two other agents invited.
4. Gaps: whether the loft conversion has sign-off; age of boiler;
   parking for a second car; not known whether the garden faces south.

Two things to fix before you rely on it. The 94 square metres came from the 2019 listing, which was before the loft conversion the vendor mentioned, so mark it "2019 figure, pre-conversion" and measure on the day. And "withdrawn after 11 weeks" is worth a question in the room rather than a line in the pack. The "Gaps" section is the most useful part: it becomes your checklist for the visit, so you don't leave without the loft sign-off or the age of the boiler.

If your team uses ChatGPT or Claude on a business plan, put this prompt, your pack template and a short tone guide into a shared Project, so every valuer starts from the same instructions instead of their own saved chats.

The day before: a comparables table built from evidence you trust

This is the step where AI does the most damage when it's used badly and saves the most time when it's used well. The rule is simple: you supply the rows, the spreadsheet does the arithmetic, and the AI writes the commentary.

1. Export the evidence. Pull 8 to 15 sold properties of the same type and similar size, sold in the last 6 to 12 months, as close to the subject as your data allows. Add 5 to 10 current listings and anything recently gone under offer.
2. Put it in one sheet with honest column headers. Address, type, bedrooms, floor area, sold price, sold date, launch asking price, days on market, condition notes. Keep asking and sold prices in separate columns so nothing can blur them.
3. Let a formula calculate price per square metre. One column, one formula. Language models are unreliable at arithmetic across a table, so don't ask them to divide.
4. Ask the AI to read the finished table. Its job is to summarise, spot outliers and suggest reasons from your notes.
5. Add your adjustments by hand. Condition, garden, parking, a better position on the street: these are the valuer's calls.

The current listings from step 1 deserve their own short summary, because the vendor will have looked at them on the portals before you arrive. Asked to summarise an export of current competition, an illustrative reply read: "Four similar 3-bed semis for sale, asking $399,000 to $435,000. Two have been reduced, one twice, after 10 and 14 weeks. The $435,000 house has a new kitchen and has had no reduction after 3 weeks." Check the reductions against the export's price history, because a relisted property can look like a new one, and keep this summary separate from the sold evidence so no asking price drifts into the comparables.

An illustrative table for a 94 square metre, three-bedroom semi in average condition might look like this:

The median price per square metre is $4,366, which puts the subject around $410,000 before adjustments. Comparable E sits almost 10% above the median, and the notes explain why. That's the kind of observation you want the AI to surface, using a prompt such as:

Here is a comparables table for a 3-bed, 94 m2 semi-detached house in
average condition.
1. Summarise the evidence in 5 bullet points a vendor would understand.
2. Identify any comparable more than 8% above or below the median price
   per m2 and suggest why, using only the Notes column.
3. Do not state a valuation. End with the questions I should answer at
   the visit to narrow the range.

An illustrative reply, trimmed:

- Five similar 3-bed homes sold in the last 9 months, between $372,000
  and $455,000.
- The middle of the range is about $4,366 per square metre.
- Comparable E is about 10% above the median, likely because of its
  corner plot and driveway.
- Comparable D sold lower and needed updating throughout.
- Prices in the area have risen around 5% this year.
Questions to answer at the visit: kitchen and bathroom condition;
parking; whether the loft room counts as a bedroom.

The fifth bullet is the one to delete. Nothing in the table supports a 5% rise; the model filled a gap with a generic-sounding market claim. That is exactly the sentence a sharp vendor will ask you to prove. Instruction 3 matters too: if you let the model propose a figure, it anchors you before you've seen the property.

The vendor pack: personalise three pages, keep the rest standard

A pack that reads like every other agent's pack wastes the prep. Keep fees, process, team and testimonials as fixed pages that only change when your business does. Let AI draft three pages from the brief, then edit them yourself:

• A covering letter that reflects what the vendor told you: their timescale, their onward purchase, the thing they're worried about.
• The evidence page: the five-bullet comparables summary in plain language, with the table attached.
• How you'd market this home: the likely buyers, the features to lead with, launch timing, photography and floor plan. The listing copy itself comes later, and writing property descriptions with AI has its own checks.

The difference shows in the first paragraph of the covering letter. A generic draft reads: "Thank you for inviting us to value your home. We are a leading local agent with an excellent track record, and we look forward to helping you achieve the best possible price." The personalised version, written from the brief: "Thank you for showing me round on Thursday. Because your onward purchase is agreed and you'd like to complete within four months, I've planned a launch for the first week of next month, with viewings concentrated on two Saturdays so you can keep packing around them." Only the second could have been written by someone who was listening.

Strike any claim you can't back with your own figures. AI drafts love lines like "we sell homes faster than anyone in the area". Unless you have the data and can show it, that sentence is a liability.

The pre-visit email goes out the day before: confirm the time, say who's coming, and list what's useful to have ready, such as the energy certificate, guarantees for recent works, and service charge details if the property is leasehold. AI can personalise it from the brief in under a minute.

Questions for the room, matched to why they're moving

The same house needs a different conversation depending on why it's being sold. Ask the AI to turn the "What the vendor told us" section of the brief into eight questions, then cut it to the four or five that matter. As a guide:

• Moving up with an onward purchase: have they found somewhere, how firm is that timescale, would they consider a rental gap if the chain breaks?
• Downsizing: what are they keeping, what does "the right buyer" look like to them, how flexible is completion?
• Relocating for work: the hard deadline, and whether they'll be around for viewings.
• Selling after a death in the family: who has authority to sell, whether other family members need to be consulted, and a gentler pace in the conversation.
• Separating owners: whether both owners agree to the sale and how each wants to be kept informed.

Keep the AI to what the vendor actually said. It should never speculate about someone's finances or personal life from indirect clues, and anything sensitive belongs in your CRM rather than in a chat window.

The 24-hour follow-up that turns a visit into an instruction

Record a two-minute voice note in the car straight after the appointment: condition, layout, what surprised you, and your price reasoning while it's fresh. Transcribe it with your phone's dictation or a meeting note-taker, then have the AI draft the follow-up letter: thanks, your recommended asking price with the reasoning tied to named comparables, the marketing plan, fees and next steps.

An illustrative run for the semi from earlier. The transcribed voice note: "Kitchen original, same sort of standard as comparable A. Loft room is decent, roof windows, proper staircase, vendor thinks sign-off is in place and will send it. Garden faces west, not south. Driveway fits one car. I'd put it at 405 to 415 and launch at 415." The AI's first draft of the price paragraph:

Based on the strong demand we're seeing locally, we're confident your home
will achieve $425,000 or more, and we recommend launching at $430,000 to
attract maximum interest.

Every number in it is wrong for this file. The valuer said 415, the draft said 430, and "strong demand" appears nowhere in the evidence. The rewrite ties the figure to the comparables: "We recommend an asking price of $415,000. The closest evidence is comparable A, a similar house with an original kitchen that sold for $398,000 four months ago. Your loft room gives you space A didn't have, and buyers should pay for it once the sign-off is confirmed, which is why we'd ask for the documents before launch." That paragraph can be defended in the room if a competitor quotes $440,000.

Check every figure against your table before it goes. Then read the price paragraph once more with one question in mind: is this the price I believe, or the price I think wins the instruction? A draft written to persuade will lean towards the second. Valuations pitched high to win business are a familiar cause of stale listings and unhappy vendors, and a clearly reasoned figure is easier to defend when a competitor quotes higher.

If the vendor goes quiet, the ongoing sequence is a separate job; automating buyer and vendor follow-up covers the timing and wording.

Where AI prep goes wrong, and how you'd notice

• A sale that doesn't exist. Every comparable the AI mentions must trace to a row in your export. If you can't find it, delete it and tighten the "use only" wording in your prompt. The mechanism behind invented details is explained in why AI makes things up.
• Asking prices treated as sold prices. This happens when both sit in one column or in pasted listing text. In one illustrative case, a valuer pasted a page of live listings alongside the sold data, and the evidence page described a house "sold at $455,000" that was in fact still for sale at that price. The vendor had seen the listing and knew. Separate columns and explicit headers fix it.
• Three different floor areas. The old floor plan, the energy certificate and your tape measure often disagree. Ask the AI to list every area figure it found and where it came from, then measure on the day.
• Market commentary from the model's memory. A general assistant's sense of "the market" is months or years old. Commentary should come only from your recent data.
• Packs that read alike. Put two recent packs side by side. If the covering letters are interchangeable, the brief isn't feeding through.
• Promises in the draft. "Will sell within a fortnight" or "guaranteed interest" should never survive editing. The same discipline applies later to listings, as AI listing mistakes that mislead buyers shows.

Two offices, three valuers: the numbers

Take an illustrative agency with two offices and three valuers running 32 appraisals a month. Before AI, preparation averages 75 minutes per appraisal, about 40 hours a month across the team. After two weeks of adjusting the prompts and pack template, the same preparation averages 35 minutes: around 19 hours a month, so roughly 21 hours back, most of it on assembling comparables and writing covering letters.

The cost is modest. Three seats of ChatGPT Business run $25 each a month on monthly billing ($20 on annual), and Claude Team is priced the same way; either keeps business content out of model training by default. Budget a morning to set up the Project and templates, and another hour a week for the first month to refine them.

Time is the easy measure. The ones that tell you whether the preparation is better, not just faster, are these:

• Instruction rate: instructions won divided by appraisals attended, tracked monthly.
• Asking price versus agreed price: if the gap widens, your evidence pages may be persuading rather than informing.
• Time to first offer on instructions won after the change.
• What vendors say: one question in your post-instruction call, "what made you choose us?", logged word for word.

If the instruction rate doesn't move, you still have 21 hours a month back. If it drops, look at the packs first: a generic pack is usually the reason.

Questions valuers ask about AI prep

Can I show the vendor the figure an automated valuation model gives?

You can show it as one reference point, labelled as an automated estimate with its source and date, but don't present it as your valuation. Automated models can't see condition, layout, light or the quality of an extension, and they tend to lag on renovated or unusual homes. Your recommended asking price should come from comparables you checked and what you saw in the room, and you should be ready to explain why your figure differs.

Should I tell vendors I used AI to prepare?

Being straightforward costs nothing. If a vendor asks, a simple line works: we use software to gather and organise the sales evidence, and the valuer checks every figure and sets the price. What you must avoid is passing off AI-written market commentary as local knowledge you don't hold. Separately, if a chatbot on your website talks to visitors in the EU, they must be told it is an AI.

Is it safe to put the vendor's circumstances into a prompt?

Not in a consumer account with model training switched on. The reasons people sell, such as separation, bereavement, debt or illness, are exactly the details that shouldn't sit in a personal chat history. Use a business plan that doesn't train on your content by default, refer to the vendor by role rather than name, and keep the full notes in your CRM rather than in the AI tool.

Further reads

• AI Virtual Staging for Property Listings: Costs and Disclosure — What to disclose if AI imagery goes into the listing after you win it.
• Can AI Analyse My Sales Spreadsheet? What to Upload and Check — How to upload a sales table so the AI reads it correctly.
• A Five-Minute Fact-Check Routine for AI Output Before It Goes Out — A quick routine for checking every figure before the pack goes out.
• How to Use ChatGPT Projects to Keep Client Work Separate — Keep each valuer's prep in a shared, tidy workspace.
• Is It Safe to Put Customer Data Into ChatGPT? — Which plan settings matter when vendor details are involved.
• How to Set Up Human Review for AI Work Without Slowing Down — Set up sign-off on packs without slowing the team down.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business and Claude Team plan pages (pricing and business-data defaults), checked September 2026.