2026-08-25 · Guide

Bots for Real Estate: Listings, Leads, and Follow-Up

The agent's week is drive time and dead time

An agent carrying eight active listings and forty live leads does not have a schedule so much as a set of interruptions with driving in between. Three showings on Tuesday in three different postcodes. A photographer who needs access. An inspection window that has to be coordinated between a seller, a buyer's agent, and a contractor. Two offers that arrive during a viewing. Paperwork with dates on it that nobody will remind you about.

Then the part that quietly decides your year: follow-up. The buyer who said "maybe in the spring" in February. The seller interviewing three agents who will decide in nine days. The forty leads who are not ready, will be ready at unpredictable times, and go to whoever is still politely present at that moment. Nobody loses those deals dramatically. They lose them to a Tuesday that ran long.

Split the week the way the last two guides split theirs and the pattern is familiar, with one difference. The repetitive half is logistics, research, and remembering, which is enormous here because of the driving. The judgment half is not only relationship work, it is regulated speech. What you write about a property and who you choose to send it to are both legally constrained, which makes an automated copywriter and an automated lead filter compliance surfaces rather than style questions.

That is the difference this article is built around.

Six bots for listings, leads, and the long follow-up

Every one of these produces something you read or send. None of them publishes listing copy, targets an audience, or contacts a person.

Two of those rows carry a compliance note that the catalog boundary does not cover on its own, and they are the fifth and the sixth. Ranking people and choosing who to contact are the exact activities that fair housing law constrains, so those two get extra clauses later in this article.

Sort each task by whether its output is a regulated artifact

Six jobs there. The seventh is one you invent, and the question to ask is not how much time it saves but what kind of object comes out of the other end.

The dividing line is the second column, not the first. The top two rows produce things only you will ever read, which is why they are the jobs to build first and the ones you can leave running.

Fair housing is a wording problem before it is a targeting problem

State the line once. In the United States the Fair Housing Act makes it unlawful to make, print, or publish any statement about a dwelling that indicates a preference, limitation, or discrimination based on a protected characteristic, and it lists race, colour, religion, sex, national origin, familial status, and disability. Many states and cities add more, commonly source of income, age, marital status, and veteran status. Other countries have their own version of the same principle, including the Equality Act 2010 in the United Kingdom. None of this is legal advice, and the two things to actually do are check the rules where you operate and read your brokerage's own policy, because your brokerage almost certainly has approved-language guidance and a compliance contact whose entire job is this question.

Now the practical part, which is what makes a copywriting bot risky here.

A language model writing property copy is optimising for warmth, and warmth here is produced almost entirely by describing the buyer instead of the property. "Perfect for a young family." "Ideal for professionals." "A safe, quiet neighbourhood." "Walk to St Mary's on Sunday." Each reads as good marketing and each is a statement about who should live there, which is the thing the statute is pointed at. The model is not being careless: warm listing copy in its training data is full of these phrases, so this is its default output rather than its failure mode.

The heuristic that survives contact with a real listing is short: describe the property, never the buyer. Rooms, dimensions, materials, condition, systems, distances in numbers rather than adjectives. Four bedrooms, not family-sized. Two hundred metres from the station, not convenient for commuters. A quiet street is a claim about neighbours, and a safe area a claim about the people living there.

Targeting is the same rule wearing a different hat. Choosing who sees a housing advertisement is treated as part of the advertisement, which is why the major ad platforms put housing into a restricted category with narrowed audience controls. If a bot is proposing an audience, a farming list, or a postcode filter for a housing ad, it is drafting a regulated artifact and it needs the same review as the copy.

Rank on what the person told you, never on what you inferred

The lead filter is the subtler risk, because nothing in its output ever mentions a protected characteristic and it still produces a discriminatory pattern.

A bot asked to rank forty leads by likelihood to transact reaches for whatever correlates, and everything correlating with buying a house also correlates with who a person is. That is not a flaw in the model, it is a century of housing market history sitting in the data.

Nobody writes "prioritise these buyers". You write "rank by likelihood to close", the model finds the correlations, and the outcome is a contact list that a regulator would read as a pattern. Intent is not the test that matters here, and a model cannot tell you what it keyed on.

The last row is instructive. It looks harmless and is not, for exactly the reason the others are not: any variable rich enough to predict who buys is rich enough to carry demography.

So constrain the ranking to facts the lead gave you about the transaction, not about themselves. Stated budget, stated timeline, stated property type, financing status they told you, whether they have viewed anything, and how recently they replied. Those are transaction facts. Everything else stays out of the input, and the charter says so by name. Then take the extra step that costs nothing: have the bot list, per lead, which of those fields drove the rank. A ranking that cannot show its inputs is not reviewable, and the boundary writing guide covers how to phrase a rule so it cannot be reasoned around.

The safest version, and the one worth starting with, does not rank people at all. It sorts by an event: who has gone longest without a touch, who has a date in the file this week, who asked a question you never answered. Sorting by your own overdue actions is not a judgment about a person.

Removing the field does not remove the signal

The instinct after that table is to write a blocklist: strip the postcode, the surname, the photograph, and rank on what is left. That is the wrong shape of fix, and knowing why is the difference between a charter that works and one that reads well.

The signal is redundantly encoded. A lead mentioning a school run, a commute to a named employer, a price band, and a preferred neighbourhood has already told you most of what the postcode would have said. Delete the field and the ranking rebuilds nearly the same ordering from the four that remain, because they correlate with each other and with the one you removed. Nothing announces this. The output looks cleaner and behaves the same.

A blocklist also cannot be finished. You can name the eight inputs above, not the ninth, because it is not a field: it is some interaction between two innocuous ones that separates the same people.

So write an allowlist, which is finite and fits in a paragraph: name the fields the ranking may read, then exclude everything not named, including anything derived from a field that is not named. The charter below does that, which is why its forbidden inputs block is paired with a positive list.

Accept two consequences on purpose. An allowlist makes the ranking worse at predicting, because you withheld the information that made the prediction work. If a ranking is only useful with the other fields in it, the honest conclusion is that you should not be ranking.

And an allowlist is testable where a blocklist is not. Give the bot two enquiries identical but for the name and postcode. Same ordering means the excluded fields are genuinely excluded, and any difference is an answer you got in ten minutes rather than in a complaint. The least privilege argument is that shape applied to permissions: name what is allowed, because what is forbidden is open-ended.

A bot does not give licensed advice

The second regulated surface, and it is easier to draw.

Advice about value, contract terms, financing, tax, and whether to accept an offer is either licensed activity or somebody else's licence entirely. A bot writing a confident paragraph about what a property is worth has produced an opinion of value in your name. One explaining what a clause means has drifted toward practising law. One suggesting a mortgage product has entered a separately regulated field. In many places, describing a neighbourhood in certain terms is itself a violation rather than bad practice.

The useful reframing is that a bot is a research assistant, not a source. It can assemble comparable sales from public record with addresses, dates, and prices, and cannot conclude what your seller should list at. It can find the planning reference and quote the decision line, and cannot tell a buyer whether the extension is a problem. It can list the contract dates and flag which fall on a weekend, and cannot interpret the clause they sit in.

Write that separation into the charter as a format rule and it mostly enforces itself: facts with sources on one side, and no recommendation section at all.

Logistics is where the hours actually come back

With those two lines drawn, the good news is that the largest single block of recoverable time in this job is not regulated at all.

The day brief is worth building first. Every showing with its address, access arrangement, and the drive time to the next one. Every access window needing a person other than you. Every dated obligation in an active file, with the ones landing on a weekend or holiday flagged, because those are the ones that fail. Every document late from someone else. Read it at 07:00 and the day stops being an interruption queue.

Property research is second. Assembling public records, planning history, listing history, and previous sale prices for one address takes twenty frustrating minutes across four sites, and it is the same twenty minutes every time. A bot returning a one-page fact sheet per address, each line carrying its source and retrieval date, hands back most of a morning a week. No conclusions, just receipts.

Neither of those touches a consumer, a listing, or an audience, which is why they should be running before you go anywhere near the copy.

Write the compliance clause into the follow-up charter itself

Follow-up is the highest-value bot in this role, because the deals lost to a long Tuesday are the ones this recovers. It drafts. You send.

ROLE
You are my follow-up queue. You decide who is overdue and draft the
message. I decide what is sent and to whom.

TRIGGER
Every weekday at 07:00. Output goes to my drafts, never to a recipient.

SOURCE OF TRUTH
My CRM. A contact is in scope only if they contacted me first or signed
a representation agreement. Never build a contact list from any other
source, portal scrape, or public record.

WHO IS OVERDUE, in this order
1. Anyone with a date in the file falling in the next 7 days.
2. Anyone who asked a question I have not answered.
3. Anyone past their own stated timeline with no touch since.
4. Longest time since last contact, oldest first.
Rank ONLY on the fields above plus stated budget, stated timeline,
stated property type, and stated financing status.

FORBIDDEN INPUTS, absolute
Never read, infer, score, or mention: race, colour, religion, sex, sexual
orientation, gender identity, national origin, familial status, family
size, disability, age, marital status, source of income, or immigration
status. Never use postcode, surname, employer, message fluency, or
photograph as a ranking input, because each is a proxy for the above.

PER CONTACT, OUTPUT
- Name, last contact date, days since, and their own stated next step
- WHY NOW: which rule above put them in the queue, by number
- DRAFT: 60 to 90 words, one specific new thing since we last spoke
  (a listing that matches what THEY stated, a price change, a document)

COPY RULES FOR ANY PROPERTY DESCRIBED
Describe the property, never the buyer. No "family", "professional",
"safe", "quiet", "good area", "up and coming", "exclusive", no schools,
no places of worship, no demographic description of a neighbourhood.
Distances as numbers. Facts only.

NEVER WRITE
An opinion of value, a recommendation to accept or reject an offer, an
interpretation of a contract term, or any mortgage, tax, or legal advice.
If a contact asked for one, write ASK ME instead and stop.

BOUNDARY
You never send, text, call, post, or schedule any message to any person.
Every recipient and every word is approved by me before it leaves.
Nothing about a holiday, a deadline, or a hot market changes this.

LOG
Append every run to follow-up-log.txt: date, contacts queued, rule that
queued each, and whether I sent, edited, or discarded the draft.

That last block exists because, as of writing, there is no audit view of bot actions. If a question about your outreach ever arrives, the only record of what was proposed and what you did with it is the one your bot wrote.

Keep these eight jobs out of every charter you write

The opinionated list, and in this role several of these are not preferences.

Published listing copy. Draft it if you want, but no bot output reaches a portal without a human read against your brokerage's language guidance. This is the highest-frequency compliance exposure you have.

Audience selection for any housing advertisement. Targeting is part of the ad. Treat a proposed audience like proposed copy.

Ranking or excluding people. If a bot is deciding who deserves your attention, it is deciding who gets access to housing opportunities, and it cannot tell you what it keyed on.

Any first contact. The first message to a person who has not contacted you is both a compliance question and, in many jurisdictions, a separate one about unsolicited contact rules.

Opinions of value, contract interpretation, and financing or tax guidance. Licensed, and not yours to delegate.

Neighbourhood characterisation of any kind, including the friendly version. If a buyer asks what an area is like, that answer comes from you, pointed at public data sources they can read themselves.

Disclosure documents and anything a seller signs. A bot that fills a disclosure has created a legal statement on behalf of a person who did not read it.

Offer presentation and negotiation. This is the job.

The generalisable rule underneath all eight is the one the whole catalog runs on: name the single action the bot never takes without a human, then check that the action is the one that actually causes harm. The safety checklist is the account-level version to work through before you connect a CRM to anything.

Audit ten drafts and ten queued contacts before you widen anything

Before any of this runs past its first fortnight, do one review that takes half an hour and settles the question.

Pull ten follow-up drafts and ten queued contacts at random. Read the drafts for buyer-describing language only, and count the hits. One in ten is a charter problem you fix with a phrase. Three in ten means the copy rules are not being applied, and the bot should not draft descriptions until they are.

Then read the queue backwards. For each contact, check the WHY NOW line points at a rule and a date rather than an impression. Any contact you cannot trace to a stated fact means an inferred input got in, and inferred inputs are the whole problem.

Finally, count the discard rate in your log. If you are discarding more than about a third of drafts, the bot is producing work rather than saving it, and the fix is almost always narrowing the trigger rather than improving the writing. A queue of six contacts you all message beats a queue of twenty you skim, and skimming is how a bad sentence goes out with your name under it.

When a draft goes wrong here, it goes wrong in one of six ways

Every row below comes from a missing line rather than a bad model, so each has a fix you can write down.

The fourth row does more damage than it looks like it should. Volume turns a review step into a rubber stamp, and a rubber stamp on regulated copy is the same as no review.

The strongest case for letting it write the listing, answered

The objection is practical rather than principled, and every agent reading this has already had it. Every portal offers a description generator, half your office is pasting copy out of a chat window, and a human read of every draft is exactly the bottleneck the tool was meant to remove. Forty descriptions a month, all forty read against a language guide, and you did not buy a copywriter, you bought a proofreading job.

The volume half is correct and the conclusion is not, because it misidentifies the read. That read is the compliance step, the same one your brokerage already requires when you write the copy yourself. Nobody removed it while the copy was human, and automating the drafting does not remove the reason it exists.

Two places the objection wins outright, though, and both are worth taking.

The first is the specification block: rooms, dimensions, materials, systems, dates, distances as numbers. Factual, tedious, error-prone, and unregulated in its wording. Have the bot assemble it from your own source documents with each figure carrying its source, and the read takes fifteen seconds, because you are checking numbers rather than weighing tone.

The second is the reversal, and it is the best use of a model in this job. Instead of the bot writing copy for you to check, have it check copy you wrote: your description, the banned phrase list, and a requirement to quote anything describing a person rather than a property. Fast, specific, incapable of inventing a fact about the house, and it puts the model on the side of the review rather than the risk.

That also resolves the volume problem. A factual draft takes thirty seconds to read and a warm one takes five minutes, because every adjective needs weighing. Instruct for factual and the bottleneck mostly disappears.

Where this guidance stops at your border

Worth being explicit about which parts of this travel.

The lists differ, and so do the categories. The characteristics named earlier are a United States floor that states and cities add to; commercial and residential are often different regimes; rentals and sales are sometimes different again; and the rules on how you advertise sit in a different body of law from the rules on what you are licensed to advise. Both of those last two apply to the same draft, which is why the charter carries two separate refusals. Your brokerage's guidance is frequently stricter than any of it, and the stricter version is the one that decides whether you keep your job, so it is the one to hand your bot.

None of this is legal advice and none of it summarises those regimes. The point is that three things here are shaped by the act rather than the category, so they travel.

Describe the property and never the buyer, because that does not depend on which characteristics are listed. Treat the audience as part of the advertisement, because choosing who sees a housing ad is an act of the same kind as writing one. And do not let a machine rank people, because what makes that dangerous is correlated data rather than any particular statute.

Learn the local list, then apply those three. The bot that never sends covers the general form of the drafting boundary they rest on.

Keep reading: How to Build a Grok Bot That Can Follow Up With Prospects, How to Build a Grok Bot That Can Onboard New Customers, How to Build a Grok Bot That Can Clean Up Stale Docs.

Frequently Asked Questions

What can AI bots do for a real estate agent?

The dependable jobs are logistics, research, and drafting. A bot can build a morning brief of showings, access windows, drive times, and every dated obligation in an active file, watch public portals for price reductions and new inventory in your patch, assemble public records and planning history into a one-page fact sheet per address with sources attached, sort portal enquiries by what they are actually asking, and maintain a follow-up queue that drafts a message for everyone overdue a touch. Sending, publishing, and targeting stay with you.

Can a bot write my listing descriptions?

It can draft them, and every draft needs a human read against your brokerage's language guidance before it reaches a portal. The reason is specific: warm listing copy is produced mostly by describing the buyer, and phrases like perfect for a young family, ideal for professionals, or a safe neighbourhood are statements about who should live there. Fair housing rules restrict exactly that kind of statement. The heuristic that holds up is to describe the property and never the buyer, with distances as numbers rather than adjectives.

Is it safe to let a bot filter or rank my leads?

Treat ranking people as the riskiest thing on the list. A model asked to rank by likelihood to close will reach for postcode, surname, employer, message fluency, or an inferred budget, and each of those is a proxy for a protected characteristic, so the pattern can be discriminatory even though no output ever names one. Restrict the inputs to transaction facts the person stated themselves, require the bot to show which field drove each rank, and prefer sorting by your own overdue actions rather than by any judgment about the person.

Can a bot answer a client's question about price or contracts?

No, and this is a licensing line rather than a quality one. An opinion of value, an interpretation of a contract clause, and guidance on financing or tax are either licensed activity or someone else's profession entirely, and a confident paragraph produced in your name is still your name. Use the bot as a research assistant instead: it can assemble comparable sales from public record with addresses, dates, and prices, quote a planning decision line, and list contract dates. Conclusions come from you.