Why this matters
Buyers check property marketing in person, so a description or photo that flatters too far is found out at the viewing. The accuracy and review workflows guide sets out review tiers sized to the consequence of error. The disclosure and data privacy guide asks whether a reasonable person would feel misled, and what happens to the data that goes in. The approvals and compliance reviews guide turns both into triggers that route work to the right reviewer. Remarks on regulation here are general direction, not legal advice.
The agency and its AI uses
Elmcourt Estate Agents is an invented independent agency selling and letting homes in a market town. Negotiators write listings and take photos; a marketing coordinator runs the website, email and social channels. Client records sit in the agency's property CRM; there is no customer data platform (CDPUnified first-party customer profiles.). Nobody has written rules for AI.
Elmcourt: AI uses found in an afternoon's audit
| Use | Who | Tool | Volume |
|---|---|---|---|
| Listing descriptions drafted from viewing notes and the floor plan | Negotiators | Business AI assistant | About 30 a month |
| Photo edits: brighter rooms, bins and cars removed, grey skies made blue | Coordinator | Photo editing app | Every listing |
| Virtually staged furniture in empty rooms | Coordinator | Photo editing app | About 5 a month |
| Website chat assistant answering buyer questions, with a human-sounding name | Coordinator | Website plug-in | Always on |
| Property-match emails written from applicant records | Coordinator | Email platform's AI feature | Weekly |
| Monthly local market update with sale price figures | Coordinator | Business AI assistant | 1 a month |
| Vendor meeting notes summarised, including personal circumstances | Senior negotiator | Free consumer chat tool | Several a week |
| "Happy family in the garden" image generated for social posts | Coordinator | Image generator | Occasional |
| Testimonials "tidied up" from short reviews into longer quotes | Coordinator | Business AI assistant | Occasional |
Classifying uses by risk
Use the accuracy guide's tiers: internal and reversible; published with an editor; customer-facing and automated; regulated or sensitive. Assign the tier by what happens if the output is wrong. Disclosure and the data rule are separate decisions, and a use can pass one and fail another.
Task
- Place each use in a tier and give the reason in a sentence.
- Set the review rule for each: who checks, against what, and whether a second person is required.
- Set the disclosure rule for each: would a buyer, tenant or vendor feel misled?
- Mark any use that must stop now, and say what replaces it.
- Write the list of client data that must never enter an AI tool, and the rule for data that may.
- Choose the measure the owner should watch to know the rules are working.
Worked answer
| Use | Tier | Review rule | Disclosure rule |
|---|---|---|---|
| Listing descriptions | Published with an editor | Second person checks every feature against the viewing notes, floor plan and vendor's confirmation | None needed for copy; the agency owns the words |
| Light photo edits | Published with an editor | Coordinator may brighten and tidy; nothing permanent removed or added | None if the property is shown truthfully |
| Sky replacement | Published with an editor | Stop unless the agency is content to label it; it changes what the buyer expects | Label if used |
| Virtual staging | Regulated or sensitive | Manager approves each image; the unstaged photo stays in the listing | Label on every staged image |
| Chat assistant | Customer-facing and automated | Weekly sample of conversations read by the manager; an incident route | Rename it; say at the start it is automated; offer a person |
| Property-match emails | Customer-facing and automated | Sampled before each send; minimum fields only | None for the copy |
| Market update | Published with an editor | Every figure traced to the agency's own sales records; a valuer reviews and takes the byline | Disclose substantial AI drafting under the valuer's name |
| Vendor meeting summaries | Regulated or sensitive | Stop now | Not applicable |
| Generated family image | Regulated or sensitive | Stop for listings; label for general social posts | Visible label; never implied to be real residents |
| "Tidied" testimonials | Regulated or sensitive | Stop now; use only real clients' words, with permission | A rewritten quote is a fabricated endorsement |
Three points drive the table. First, the agency's product is the property, so any edit that changes what a buyer will find on arrival misleads, whatever the tool. Second, the valuer's local experience is the market update's whole value, its E-E-A-TExperience, Expertise, Authoritativeness, Trust., so the valuer must shape the piece, not merely approve it. Third, the approvals guide's matrix routes AI-assisted content to an editor plus any other trigger: a testimonial or a new claim still goes to whoever checks claims.
Client data that must never enter an AI tool
- Identity documents and anti-money-laundering check results.
- Proof of funds, mortgage details and other financial records.
- Vendors' or tenants' personal circumstances, such as health, bereavement, divorce or debt.
- Information about children.
- Key, alarm and access arrangements.
- Full client exports, in any tool.
Other personal data, such as an applicant's first name and search criteria, may go only into approved tools whose contract forbids training on inputs and sets retention, and only the fields the task needs. The meeting summaries fail every part of this rule: they contain special-category detail in a consumer tool.
The measure
Use the register as a KPIA metric tied directly to an objective. source: the share of live AI uses with an owner, a tier and a passed disclosure and data check at the last review.
Elmcourt's register after the rules
Uses audited: 9. Stopped: 2 (meeting summaries, tidied testimonials). Changed: 4 (chat assistant renamed and disclosed, virtual staging labelled, market update rebylined, family image labelled). Continuing as they were with a review rule: 3. Uses with owner, tier and passed checks at the first monthly review: 7 of 7. Chat sample in the first month: 40 conversations read, 3 handed to a negotiator late, so the hand-off wording changes.
Common mistakes
- Asking whether AI was used. It nearly always was; ask whether anyone would feel misled.
- Treating all photo edits alike. Brightening is fine; removing a fixture is not.
- A friendly name on a bot. That is the surprise disclosure exists to prevent.
- Rewriting real reviews. A longer quote the client never wrote is not their review.
- Self-review. The negotiator who wrote the notes is the wrong person to check the draft against them. The beginner level overview leads next to the capstone policy.
