A small brokerage can use AI for eight practical jobs: listing drafts, enquiry sorting, viewing coordination, follow-up, valuation preparation, transaction checklists, document indexing and performance reporting. Start with listing drafts or agent-approved follow-up when you have reliable source data and a named reviewer. Leave pricing, access decisions, legal interpretation and client promises with people.
“Use AI” is not a purchase decision. For each job below, specify the source of truth, the tool category, the person who checks the output and the failure that stops it. A writing assistant, existing CRM, calendar and spreadsheet may be enough for a pilot. The busiest workflow is not always the best first one if the source data is poor.
Choose the first job with a simple brokerage score
Ask four questions of each candidate workflow. How many times does it occur in a normal week? Can the output be checked against a trusted record? Would an error mislead a buyer, seller or colleague? Does the current system already capture the result? Give each a score from one to five, with five meaning high volume, easy checking, low consequence and easy recording. Add the scores only as a discussion aid. The final selection should still be made by the responsible broker.
For an illustrative three-agent brokerage, listing drafts score 4, 5, 3 and 4, totalling 16. Buyer valuation advice scores 3, 2, 1 and 3, totalling 9. The latter may consume more senior time, but an unsupported price claim has a larger consequence and is harder to verify. The first pilot is therefore listing drafts, with an agent checking every claim against the property sheet. This is a priority rule, not a claim that a particular score predicts savings.
Before copying any client data into ChatGPT or another assistant, check the exact plan, privacy settings, retention and access rules against your brokerage policy. The official ChatGPT plan page changes over time; it does not replace checking the specific service agreement for the way your team will use it. A CRM function may be preferable when it preserves the record and permissions already in place.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy now1. Draft descriptions from a verified property sheet
Tool: a secure writing assistant attached to a structured listing form, or a CRM writing function if the brokerage already has one. Supply property type, room counts, measured area, confirmed amenities, access details and wording restrictions. Ask for two plain-English versions that use only these fields. This can remove the blank-page problem for an agent without asking the model to invent a selling point.
An illustrative sheet says “two bedrooms; one bathroom; balcony; parking space unverified; 82 square metres per measurement record”. The assistant drafts “private parking and generous 90-square-metre accommodation”. The agent rejects both claims, checks the measurement source and leaves parking out. The corrected description states the verified area and balcony without upgrading an unknown into a fact. Keep the rejected draft with the approved version during the pilot so recurring errors are visible.
Human check: compare every physical, financial and access claim with the approved record, photographs and current seller instructions. Avoid language that could imply a preference for or exclusion of a type of person. Wording is only half the risk: ad platforms let a sender steer a listing towards some people and away from others, so check who an advert actually reached. A regulator withdrew its 2024 guidance on housing adverts delivered through digital platforms in September 2025, yet its withdrawal notice says conduct that breaches the underlying fair housing law remains subject to enforcement. Get qualified advice on the rules governing your listings. For the writing details, see AI property descriptions for estate agents.
Give the writer a field labelled “unknown” rather than leaving uncertain facts blank. A blank parking field may invite the model to fill it from a photograph; “parking: unverified, omit” is a clearer instruction. The agent should also read the final copy next to the selected photographs. A description can be technically accurate yet misleading when a photograph shows a feature outside the property boundary. Record the approved source date, because a service charge or availability claim can change between draft and publication.
2. Sort enquiries without deciding who deserves a response
Tool: a CRM queue or inbox assistant that labels incoming messages by request type: viewing, property details, seller enquiry, complaint or other. Classification helps the team route work; it should not decide a person's eligibility or silently discard a message. Use simple confidence thresholds: if the assistant is unsure, mark “agent review” and keep the original text visible.
Suppose the brokerage receives 38 enquiries in a day. An illustrative assistant labels 19 as viewing requests, 11 as property questions, five as seller enquiries and three as uncertain. An assistant at the desk checks the three uncertain cases and samples five from the other groups. One “seller enquiry” is actually a complaint about inaccurate marketing and gets escalated immediately. The useful result is a queue with accountable routing, not a claim that 35 messages were solved automatically.
Human check: ensure every enquiry has a visible owner and timestamp, test the categories against messages written in different styles and languages, and review whether response times vary unfairly across groups. Do not use demographic guesses, names or inferred protected characteristics as lead scores. A brokerage chatbot can collect and route the same requests, provided the full enquiry reaches a person when the answer is uncertain.
Configure a “no silent closure” rule for the queue. If a contact asks about two properties, a single label may lose one request; the record should show both property IDs. If a message includes a complaint and a viewing request, the complaint should not be buried under the routine viewing category. Review a short daily exception list and compare it with the unfiltered inbox. That comparison finds classification errors that a dashboard of assigned messages will never reveal.
3. Coordinate viewings around real calendar constraints
Tool: a calendar linked to the property record, with AI drafting proposed time choices and confirmation wording. The actual availability should come from the agent diary, property access instructions and occupants' consent. The assistant can propose slots; a booking is real only after the calendar and all necessary people confirm it. For out-of-hours chat, connect the response to a human hand-off rather than implying a guaranteed viewing.
For example, a buyer asks for Saturday morning. The diary appears free at 10:30, but the property note requires 24 hours' notice and the occupant has not confirmed access. The AI suggests “Your viewing is booked for 10:30”. The coordinator replaces that with “We have requested 10:30 and will confirm once access is agreed.” That single status distinction prevents a disappointed arrival. The viewing request remains pending in the CRM until the occupant and agent are confirmed.
Human check: verify the address, property status, access rules, travel time, date format and recipient before sending. A cancelled instruction must remove the slot from every channel. Track proposed, confirmed, changed and cancelled viewings separately; otherwise a dashboard can count tentative requests as completed appointments. For a narrow out-of-hours use, read AI chatbot valuation booking and apply the same confirmation discipline to viewings.
Before connecting a chatbot to the diary, test overlapping appointments and last-minute changes with dummy records. One agent may appear free while travelling between properties, or the property may have restricted access hours. The bot should offer a request window, not expose the entire agent calendar or an occupant's personal schedule. Design a fallback message for an unavailable slot: offer two alternatives and say a coordinator will confirm. Measure confirmed viewings and failed confirmations separately during the pilot.
4. Write follow-up that reflects the actual conversation
Tool: a CRM draft queue using the agent's approved notes, not a guessed profile. After a viewing, AI can turn terse notes into a short message asking for feedback or offering the next factual step. It can also highlight contacts who were promised a reply but have no task assigned. An agent decides what to send and when; automated persistence is a poor substitute for a useful answer.
An illustrative note reads “liked layout; asked about service charge; no figure confirmed”. A bad draft says “The service charge is low and should be fine for your budget.” A checked version says “You asked about the current service charge. We're confirming the latest figure and will send it when verified. Was there anything else you wanted us to check?” This is more useful than an upbeat message with an invented number.
Human check: compare the draft with the agent's notes, check permissions to contact the person, remove stale property links and make a stop rule for replies. If the buyer says no, an AI sequence must not keep chasing them. A seller's negotiation position is not suitable for an automatic template. The deeper buyer and vendor follow-up workflow covers the message sequence rather than the catalogue decision here.
Use the CRM's actual status fields to decide who is due a reply. “No activity for seven days” is not enough: a buyer may be awaiting a document from the brokerage, or the seller may have asked for no contact until a specific date. Put the promised next action into the task record and let AI draft from that. For quality control, sample messages from contacts who replied, contacts who opted out and contacts who never responded. A high send count can coexist with poor service.
5. Prepare valuation meetings without generating a price
Tool: a research and note-organising assistant fed with verified property information, previous client questions and the brokerage's own approved comparable-property records. Ask it to produce an appointment brief: missing facts, questions to ask, property features to inspect and sources requiring confirmation. It may suggest a structure for the discussion. It should not present a market value as fact or invent comparable sales.
In an illustrative pre-appointment pack, the model spots that the form says “converted loft” while the floor plan labels the space “storage”. It drafts a question asking how the area is used and what supporting documentation exists. The agent resolves the discrepancy during the visit before treating the area as living space. The useful AI contribution is locating the contradiction early, not turning it into a higher asking price.
Human check: the valuer validates every comparable, adjustment, property condition and seller objective. If source data is old or incomplete, the brief should say so. Document the difference between an asking price discussion and an evidence-backed valuation. Any regulated or contractual representation should be handled by a qualified person. This is why valuation preparation belongs ahead of automated valuation claims in the brokerage priority list.
A good appointment brief also has an evidence request list: current floor plan, known works, occupancy details, recurring charges and any restrictions that the client has disclosed. Do not ask AI to infer these from photos or similar properties. After the visit, the agent marks which questions were answered and which remain open. This turns the brief into an auditable preparation aid. It prevents a polished generated narrative from hiding the fact that essential information is still missing.
6. Track transaction milestones and missing actions
Tool: a transaction checklist inside the CRM, with AI suggesting which items may need attention from approved status notes. It can point out that a promised document has not been logged or a party has not received an update. The checklist must be built from the brokerage's actual workflow and transaction type; a generic model-generated list can omit a critical step or add one that does not apply.
Imagine a sale file with 14 internal milestones. The assistant reports that 11 are complete and three remain. One “complete” item was checked because a draft document was uploaded, although the signed version is still missing. The transaction coordinator reopens it and records “awaiting executed copy”. The broker sees a truthful status rather than a green dashboard. The AI did not verify the legal adequacy of the document.
Human check: reconcile milestones with correspondence, version status and the responsible professional's advice. Never let the assistant declare that a transaction is legally ready, interpret a contract or tell a client that a condition has been satisfied. Use neutral statuses such as requested, received, checked and approved. Assign an owner to each overdue action and set a manual escalation path when a deadline approaches.
For each milestone, store the trigger and the evidence that closes it. “Document received” needs the file identifier; “buyer updated” needs the sent message and date. A model should not mark an item complete merely because it sees the word “done” in a note. When a deadline is changed, keep the old date and the reason rather than overwriting the record. The transaction coordinator can then explain what changed to colleagues without trusting the model's memory of a chat.
7. Index documents and spot inconsistent property facts
Tool: document extraction or search across an access-controlled property folder. AI can list file type, date, apparent version and questions raised by inconsistent numbers. This is useful when a team inherits a file with photographs, forms, certificates and correspondence in different folders. The output should include a link to the original document, not merely a confident summary.
An illustrative file has a floor area of 82 square metres on a current measurement record, 78 in an older listing and 80 in an email. The AI index surfaces all three. The agent records which measurement is approved for marketing, checks its method and corrects the listing and brochure. Without the index, a copywriter might simply choose the largest number. The example shows a mismatch to investigate, not a rule for which figure is legally correct.
Human check: inspect original versions, ownership and access permissions. Confirm that the document relates to the right property; similarly named developments or unit numbers can lead to cross-file contamination. Remove unnecessary personal information before sharing a summary with another party. If the file is incomplete, the index must display “not provided” instead of implying that a document does not exist.
Try a controlled extraction test before rolling this out. Give the tool ten files whose types and dates are already known, including two deliberately similar names and an outdated version. Record how many it classifies correctly and whether its source links open the right file. If it struggles with scanned pages, change the scanning process or keep those documents in manual review. An index that cannot reliably connect an answer to its original is more dangerous than a slower folder search.
8. Explain brokerage performance from a checked export
Tool: a spreadsheet or reporting assistant working from a dated CRM export. It can group enquiries by source, count confirmed viewings and draft plain-English notes for a manager. State the definitions before the model analyses anything. An enquiry is not a viewing; a proposed viewing is not a confirmed one; a closed deal is not a lead. Keep source IDs so a surprising result can be traced.
For a worked illustrative month, the brokerage records 120 enquiries, 48 confirmed viewings, 12 offers and three completed transactions. That gives a 40% enquiry-to-viewing rate, a 25% viewing-to-offer rate and a 2.5% enquiry-to-completion rate. The assistant writes “performance improved because social ads brought better buyers”, but the export has no previous-month comparison or buyer-quality measure. The manager removes the causal claim and instead reports the four counts and three clearly defined ratios.
Human check: look for duplicate contacts, reopened leads, cancellations and missing source tags before drawing conclusions. Compare like periods and the same definition of completion. A low conversion rate may reflect unavailable stock or delayed recording, not poor agent work. The manager should annotate uncertainty, assign a specific follow-up question and keep the raw export. AI is helpful at finding a pattern; responsibility for its meaning stays with the brokerage.
Give the assistant a data dictionary before asking for prose: one row per enquiry, unique enquiry ID, first received date, property ID, source, confirmed viewing date, offer date and completion date. If those fields are absent, the right output is a data-quality question. For example, 18 of 120 enquiries may have no source tag; the report should show “15% unknown source” instead of attributing them to the largest channel. A manager can then fix data capture before changing marketing spend.
A two-week pilot that makes the eight jobs comparable
Set one rule across all eight jobs: AI may organise and draft from approved data, but it must not decide who receives housing information, which audience sees an advertisement or whether someone is a suitable applicant. A targeting system can produce uneven access even without an explicitly discriminatory instruction. Review the actual delivery and response outcomes, not just the wording of a prompt. The responsible broker should obtain qualified advice on the housing rules that apply before deploying automated advertising or screening. This catalogue does not supply a legal test for those decisions.
Pick one job and one owner. Write the allowed input fields, sample output, reviewer and stopping conditions on a page. Run five past cases through the proposed workflow without sending anything. Count factual errors, unsupported statements, missing records and minutes of human correction. Then run a small live trial with every output checked. If the job touches client information, verify tool permissions and retention before any upload. If it touches housing access or advertising, seek qualified advice on the applicable rules.
Use a simple acceptance record: “job; source version; output owner; errors found; decision; date”. For a listing, the decision might be “publish after area correction”. For an enquiry queue, it might be “do not automate routing until mixed requests are captured”. This prevents a successful demonstration on one clean example from being mistaken for a working brokerage process. Test a difficult case as well: an incomplete property sheet, a changed viewing instruction, a duplicate contact or a transaction file with two document versions. The review should reveal whether the system stops safely when it lacks a dependable answer.
A filled illustrative scorecard for the listing-draft pilot might say: 12 listings processed; average first draft 7 minutes including form entry; average checking 11 minutes; four unsupported amenities removed; two floor-area discrepancies escalated; zero unreviewed descriptions published. Compare this with the team's previous preparation time measured the same way. The figure that matters most is zero unreviewed publication, because a shorter drafting step is worthless if inaccurate claims reach the market.
At the fortnight review, decide whether to expand the chosen job, repair its source data or select another of the eight. If the listing form is often missing measurements, solve that first. If follow-up is slow because agents do not write notes, an automated message sequence will amplify the gap. A small brokerage gets the most value when AI makes a well-defined hand-off faster and leaves an agent able to explain every outward-facing claim.
Questions before a brokerage pilot
Which AI job should a two-agent brokerage try first?
Start with a repeatable, source-backed job such as drafting a listing description from a verified property sheet or preparing a viewing follow-up for agent approval. Count the work and corrections for two weeks. Avoid automating valuation claims, applicant decisions or legal documents before you have a reliable review process.
Can AI decide which buyer should see a property?
It should not make opaque decisions that restrict access or change service based on a person's characteristics. Use objective, published property requirements for administrative sorting, let an agent review every response, and obtain qualified advice on applicable housing and advertising rules. Keep an audit trail of what was offered and why.
Does a brokerage need eight separate AI subscriptions?
No. Many of these jobs can be piloted with a secure writing assistant, existing CRM functions, a calendar and a spreadsheet. Choose the workflow before buying software. Check whether your current tools support the necessary permissions, review queue, data export and human hand-off, then price only the gaps.
Further reads
- AI Tools for Real Estate Brokers: From Transactions to Agent Support — Compare tools across the whole brokerage.
- AI for Real Estate Brokers: Compliance File Review Step by Step — Design a deeper transaction-file review.
- AI Policy for a Real Estate Brokerage: What Agents May and May Not Do — Set permissions for agents and assistants.
- Real Estate AI Chatbot Features: A Checklist Before You Buy — Check a chatbot's actual hand-off features.
- How Estate Agents Use AI to Prepare Valuation Appointments — Prepare appointments without delegating value judgement.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: housing regulator's notice withdrawing its 2024 guidance on advertising through digital platforms (published 6 April 2026, withdrawal effective 17 September 2025); OpenAI ChatGPT plans and pricing page (checked 4 November 2026).