Before you share client files with an AI tool

The first risky AI moment in an agency is often quiet. An agent has 12 minutes before a valuation, a vendor update to draft, and a tool that can turn rough notes into a neat paragraph. So they paste the live file: buyer names, viewing feedback, offer position, perhaps a sensitive reason the seller wants a quick move.

Before the agent pastes anything, the agency needs to have approved the tool for that information.

For a newer agent, this can feel abstract. A client file is simply the collection of information your agency holds about a person, property, tenancy, viewing, offer, or instruction. Some of it is ordinary working context. Some of it is personal, commercially sensitive, or only meant for certain people in the branch.

Before any real estate AI assistant touches that information, the agency needs a short permission check. Keep it practical: a working rule staff can use before live buyer, seller, landlord, or tenant records are pasted into or connected to a new tool.

Start with dummy data before client data

Every new AI tool should pass a dummy-data test before it sees a real file.

Dummy data means realistic but invented details: a fake buyer, a fake property, a fake viewing note, a fake complaint, a fake landlord instruction. The point is to test the task without exposing a real person.

This matters most with free AI tools for real estate agents, because the price tag can make the tool feel harmless. Approval still matters. The agency still needs to know who can access the input, whether prompts are stored, how long data is kept, and whether staff can retrieve an audit trail later.

Run the first test like this:

1. Create a fake contact with a believable search brief.
2. Add a fake property, viewing note, and next action.
3. Ask the tool to summarise the position or draft a follow-up.
4. Check whether the output invents facts, changes the client position, or hides uncertainty.
5. Record whether live data is still blocked, conditionally allowed, or approved for a specific use.

Record the approval explicitly. Keep a tool in testing until it has a recorded approval, even if one draft looked good.

The ICO guidance on AI and data protection gives a useful grounding here: if personal data is being processed, organisations need to understand it and account for it. In agency terms, that starts before a staff member pastes client notes into a box.

Decide what the tool is allowed to see

The permission check should separate types of information instead of relying on gut feel.

An agent may be allowed to use AI to tidy a public-facing listing paragraph built from approved property facts. Live negotiation notes, complaints, vulnerable client notes, access arrangements, and tenant arrears histories need a stricter rule.

Use this table before approving any AI real estate tools for branch work:

This is deliberately cautious. It is easier to widen a rule after testing than to claw it back after staff have built a habit.

NAR’s piece on why brokerages need an AI use policy makes the same operational point from a US brokerage angle: teams need approved tools, data standards, and human oversight before usage spreads informally.

Ask the retention question in plain English

If the answer is vague, the tool is not ready for live client files.

Ask:

• What happens to the text, files, and prompts staff enter?
• Is the information used to train or improve models?
• How long is it stored?
• Who can access it?
• Can the agency delete it?
• Can the agency prove what was entered and what came back?

These questions decide whether a vendor update, offer note, or landlord instruction can be traced later.

Real estate work often becomes contentious after the calm moment has passed. A buyer says they were told one thing. A seller asks why a stronger offer was not recommended. A landlord disputes what was said about a repair.

If AI touched the work, the agency should still be able to show the source, reviewer, and final action.

NIST’s AI Risk Management Framework is broader than real estate, but its emphasis on governance, mapping risks, measuring issues, and managing them is useful for principals. Translate that into branch language: know the use, know the risk, assign an owner, and keep a record.

Give every AI tool an owner

Unowned tools spread through teams quickly. One agent uses a browser extension. Another tries a mobile app. An admin tests a document helper. A manager hears about it only after client data has already moved.

Make someone responsible for each tool before staff use it on live files.

The owner can stay out of everyday prompting, but they need to know what the tool is used for, which types of information are allowed, who can use it, what training staff received, and when the decision comes back for review.

This is where a CRM for real estate work should do more than store contacts. In AvaroAI, the permission check can become an assigned task with a review date, linked to the people responsible for the decision.

Keep that approval where staff can find it when they need to use the tool.

A simple owner record should include:

• Tool name and purpose
• Owner
• Approved users or roles
• Allowed types of information
• Blocked types of information
• Dummy-data test result
• Retention answer
• Review owner
• Staff training date
• Next review date

That record gives staff a practical answer when they ask, “Can I use this real estate AI assistant for this file?”

Keep approved AI inside permission boundaries

There is a major difference between an approved assistant working with agency records and an external tool where staff paste fragments of live files.

AvaroAI’s AI chat assistant is designed to work against the agency’s own listings, contacts, offers, viewings, and tasks. That matters because permission, ownership, and context can be handled where the work already lives. Staff can refer to the existing records without copying client information into a separate notebook.

The agency still needs someone responsible for these decisions. Instead of asking each agent to remember a dozen ad hoc rules, the agency can decide which roles see which records, which tasks need review, and where sensitive context belongs.

When comparing tools, check whether they enforce access permissions or rely on staff to decide what they can paste.

The second option is fragile on a busy day.

Train the first use as well as the policy

Staff training should cover the first live use of a tool. That is where habits form.

Give agents 3 examples:

1. A safe test using dummy data.
2. A conditional use with ordinary working context in an approved tool.
3. A blocked use involving sensitive or disputed client information.

Then ask them to sort 5 real branch scenarios into allowed, ask first, or blocked. Keep the examples close to daily work: viewing feedback, a price reduction email, a landlord update, an offer summary, a complaint note, a tenant repair issue.

This takes 20 minutes and exposes the gaps fast. If 3 experienced staff members sort the same scenario 3 different ways, the rule is not clear enough.

A 30-minute permission check for this week

Pick one AI tool your agency is already testing or likely to test. Run this check before any more live client information goes into it.

If the team cannot establish where the data goes or who can access it, keep live files out until the person responsible can confirm those details.

Related reading

• The data habits that make AI safer for agency teams
• The AI tools worth using are the ones that fit the work
• Where AI helps real estate agents, and where it gets dangerous