What to check before sending an AI-written client message

The risky AI message is not always dramatic. It is often the polite vendor update that slightly overstates interest, the viewing follow-up that sounds warmer than the buyer felt, or the listing paragraph that turns “near transport links” into more than the file supports.

Most agents can spot a ridiculous AI draft. The harder problem is the one that is almost right.

Agencies need a consistent way to review those drafts. If AI helps write, summarise, or tidy client communication, the branch needs a clear way to decide when a message can go straight out, when it needs a peer check, and when a manager should review it.

One unverified sentence can become the agency’s official position once it reaches a client.

NAR’s article on whether agents can trust AI in real estate work reported that agents are already using AI while worrying about accuracy, compliance, and client-facing use. Managers should expect staff to use AI for daily correspondence and agree how to review it.

Not every AI draft deserves the same review

A branch will reject a review process if it treats a viewing confirmation like an offer recommendation. The first job is to classify the message before arguing about wording.

Consider what happens if the message is wrong. If a wrong answer causes a diary correction, the agent can usually review it alone. If a wrong sentence changes a client’s understanding of the property, the negotiation, the level of interest, or the agency’s promise, the message needs stronger sign-off.

The same principle applies whether the draft came from a standalone chatbot, an AI agent for real estate workflows, or an assistant inside the agency system. Consequence matters more than origin.

The source-record check comes before the wording check

Most AI review discussions start too late. People ask whether the draft “sounds good”. The review should start with whether the draft can be supported by the agency’s own records.

For listing claims, the reviewer should check the property record, approved photos, floorplan, brochure notes, vendor or landlord instructions, and any local disclosure notes. If the draft says “recently renovated kitchen”, where is that supported? If it says “ideal investment”, is that a marketing phrase the agency is comfortable with, or does it imply advice the agent should not be giving?

For viewing follow-ups, the reviewer should compare the draft with attendance, feedback, applicant requirements, objections, next action, and previous communication. A buyer who said “nice but too small” should not receive a message that treats them as strongly interested because the AI draft sounded upbeat.

For vendor updates, the source check is wider: enquiry volume, viewing outcomes, feedback themes, price conversations, comparable activity, open tasks, and previous promises. The draft should not create confidence the branch cannot justify.

The FTC’s consumer guidance on AI chatbots is blunt about verifying chatbot answers and avoiding reliance on them for high-stakes advice. Estate agency communication has its own version of that rule: verify the record before polishing the sentence.

AvaroAI’s listing management is designed around this review because property data, photos, documents, notes, and custom fields sit against the listing rather than in separate folders. When AI-assisted copy mentions a feature, condition, access arrangement, approval, or campaign decision, the reviewer should be able to check the source without rebuilding the file from inbox fragments.

Five checks before sending

A useful AI review checklist should be short enough to use under pressure. If it becomes a policy document nobody opens, it has failed.

Check the following:

1. Facts: Is every factual claim supported by a current record?
2. Authority: Is the message making a decision, recommendation, or promise the sender is authorised to make?
3. Tone: Does the message fit the relationship, recent history, and emotional temperature of the client?
4. Fairness: Could the wording mislead, exclude, pressure, or treat someone differently without a defensible reason?
5. Record: Is the final message, reviewer, and reason for any change visible in the agency record?

Pay particular attention to what the sender is authorised to say. AI can make a draft sound more confident than the agent intended. “We recommend accepting this offer” is not the same as “This offer meets the price and timescale points we discussed; shall we talk through your options?” The first recommends a decision; the second asks the client to discuss it.

AI often smooths away details that should affect the tone. A landlord waiting three days for a repair update needs acknowledgement, not a cheerful template. A seller who rejected a previous price reduction may need careful framing, not another generic “market feedback suggests” paragraph.

Unfair wording can damage a client’s trust as well as raise compliance concerns. If a draft describes applicants, tenants, buyers, sellers, or neighbourhoods in a way that would sound wrong if read aloud in a team meeting, it should not leave the branch.

Save who reviewed the message and which facts supported it.

Managers should define triggers, not approve everything

Managers do not need to become the bottleneck for every AI-assisted message. They need to define review triggers clearly enough that agents know when to escalate.

Price, rent, offer strength, negotiation position, and campaign strategy should usually trigger stronger review because the wording can affect client decisions. Property features, condition claims, legal-sounding phrases, and investment wording need source records behind them. Complaints, challenged advice, sensitive third-party context, new staff, cover work, incomplete notes, and reusable templates deserve the same caution.

Team visibility matters here. A manager cannot sign off well if the draft is isolated from the work it describes. A peer cannot review a viewing follow-up if the viewing record, applicant requirements, and earlier conversation are hidden elsewhere.

AvaroAI’s team collaboration and role-based access are built for that tension: enough visibility for managers and colleagues to review the work, with permissions that stop sensitive information being casually exposed. Review should happen where the work lives. If approval requires screenshots, forwarded emails, and “can you just check this?” messages, the process will drift.

Attach the draft and records to the review task

The weakest review workflow is a private chat message that says, “Can I send this?” The reviewer then has to ask which property, which client, what happened before, whether the wording was AI-assisted, and what decision is being made. The reviewer has to reconstruct the file before they can help.

A better review task names the message type, linked records, review trigger, source materials, decision needed, deadline, and final approved wording. In AvaroAI, tasks and reminders can link to contacts, listings, properties, and events. That is the difference between “please review” and “review this vendor update against the latest viewing feedback and price conversation before 4pm”.

It also protects the agent. When a manager changes a phrase, removes a claim, or decides the message should be a call instead, the reason should not disappear.

Staff may use AI for several different tasks. One agent may use AI for first drafts, another to summarise viewing feedback, and a manager to prepare a weekly client update. Without a sign-off workflow, every person invents their own risk standard.

The branch standard should fit on one page

Keep the communication standard short enough for staff to consult while working.

At minimum, it should answer five questions:

1. Which client-facing messages can agents review themselves?
2. Which messages need peer review?
3. Which messages need manager approval?
4. Which source records must be checked before approval?
5. Where is the final approved message recorded?

For agencies operating under formal professional standards, this should sit beside the normal rules on accuracy, complaints, record-keeping, and client care. The Property Ombudsman codes of practice are a useful reference point for UK agents because they focus on communication, evidence, and accountability.

The standard should also say what AI must not do. It should not provide legal, conveyancing, mortgage, valuation, or financial advice. It should not decide the negotiation position, invent facts to make a message smoother, or turn incomplete notes into certainty.

That still leaves useful work. AI can turn rough notes into a cleaner draft, summarise the last five updates, or help an agent prepare options for a manager to review. It can make a busy branch faster without making the branch careless.

Keep enough detail to explain what you sent

Clients rarely ask whether an email was AI-assisted. They care whether it is accurate, fair, timely, and aligned with what they told the agent.

A review should leave the agency able to explain the message it sent.

The practical test is simple. If a client challenges the wording tomorrow, can the branch show where the facts came from, who reviewed the message, what changed, and why it was sent?

Related reading

• Where AI helps real estate agents, and where it gets dangerous
• The AI tools worth using are the ones that fit the work
• Client communication fails when the thread has no owner