AI for Real Estate Agents: What Earns Commission

AI for real estate agents pays for itself in one place before all the others: the gap between an enquiry arriving and a human answering it. Everything else on the list is useful. That one is the difference between a viewing and a competitor's viewing.

Portal enquiries arrive at eleven at night, during a viewing, and on the Saturday you were not working. A solo agent cannot answer them all in minutes by hand. A drafted reply waiting for one approving tap means they can.

AI for real estate agents starts with lead response#

The workflow is simple and it is the first thing I build for agency clients.

1. Enquiry lands from the portal, the website or a form.
2. It is matched to the property and to any previous contact with that person.
3. A reply is drafted: acknowledges the specific property, answers the obvious question, proposes two viewing slots.
4. It waits in your phone for approval.

Approve, edit, send. Under a minute, at any hour, from anywhere. The reply is personal because it references the actual property and the actual question, not because a template inserted a first name.

Never let it send unsupervised. An automated reply that promises a price, a completion date or an availability you cannot honour is a complaint, not a lead.

Listing copy, from facts only#

Generated descriptions work when the model writes from a verified fact sheet and nothing else: dimensions, year built, rooms, features, restrictions, energy rating.

Given photos and a title, it invents. Confidently. It will add a "recently renovated kitchen", a school catchment, a permit or a square footage because those things usually appear in descriptions of similar properties. Each invention is a misrepresentation published under your licence.

Two rules keep this safe:

• Structured facts in, prose out. If a fact is not on the sheet, it does not appear in the copy.
• Read every line before it publishes. Including the ones you have read forty times, because the forty-first is where the invented detail is.

Advertising and fair housing rules apply to the model's phrasing exactly as they apply to yours. A model asked to make a description appealing will sometimes produce language implying a preferred type of buyer or neighbourhood character. That is your liability, not the vendor's.

Viewing follow-up, where commission is lost#

Most agents follow up well for two days and badly after that. The pipeline of people who viewed and did not decide is where deals quietly disappear.

Automate the memory, not the message:

• A summary of what each viewer said, written into the record straight after the viewing.
• A scheduled prompt at day three, day ten and day thirty with a suggested message referencing what they liked and what they hesitated about.
• A flag when a new listing matches something a past viewer said they wanted.

That last one is the closest thing to free commission in this business, and it is impossible to do by hand across a hundred past contacts.

Market notes and comparables#

Assembling comparables, summarising recent movement in a postcode and preparing a vendor conversation is research work, and research is what these tools are genuinely good at.

Use it to prepare. Do not use it to conclude. An automated figure presented to a vendor as a valuation is, in most markets, a regulated activity you are not performing correctly, and it is also frequently wrong because the model cannot see the condition of the property.

The safe framing: here is what similar properties did, here is what I think and why. The model produced the first sentence. You produced the second.

The stack for a small agency#

Small, deliberately:

• An assistant for drafting and research. See Claude for small business for the setup.
• n8n to connect the portal, the CRM and your phone, with a failure alert you actually see.
• Your existing CRM. Do not migrate. AI CRM covers adding the useful parts to what you have.

Under a hundred a month for most agencies, plus the build. AI automation for small business covers the ownership and maintenance question, which matters more than the tooling.

Vendor reporting, the job nobody enjoys#

Sellers want to know what happened this week. Agents dislike writing it, so it arrives late, thin, or only when the vendor chases.

This is an ideal automation because everything it needs already exists in your systems: portal views, saved searches, enquiries, viewings booked, feedback given. Pull those on a schedule, assemble them into the same layout every week, and add one paragraph of interpretation that you write yourself.

That last paragraph is the part clients pay for and the part a model should not produce. It is where you say the price is wrong, or the photographs are, or that three viewers mentioned the same thing. The numbers are assembly work; the judgement is the service.

The side effect is that difficult conversations happen earlier. A vendor who has seen the viewing count fall for three weeks is easier to talk to about price than one who is hearing all of it for the first time.

What to measure#

Median response time to a portal enquiry. Target minutes. This is the number that moves viewings.

Viewings booked per enquiry. The conversion that actually matters.

Re-contacted past viewers per month. The forgotten pipeline.

Do not measure listings published or messages sent. Both rise automatically when you automate and neither tells you whether you earned anything.

The short version#

Build lead response first and keep a human approval on every outbound message. Generate listing copy from a verified fact sheet only. Automate the memory of past viewers rather than the words you send them. Use market research to prepare a conversation, not to issue a valuation.

If you want this built on your portal and CRM with your team trained to run it, that is what my AI training packages cover, and the AI consultant for small business page is where I do this work for local businesses.