AI in real estate brokerage: what to deploy, what it costs, and what the law says

Talk about AI in real estate brokerage tends to land on one of two extremes: either "robots will replace agents" or "this is all hype, we run fine on WhatsApp and a spreadsheet." The reality sits in between and is far less dramatic: a handful of specific tasks where a model is objectively faster and cheaper than a person, wired into a process a human still runs — because a real-estate deal is about trust, documents and money, not just text. Below: which kinds of AI already work in the niche, when it makes sense to bring them in, what it costs, and what Russian law says about it.

Which kinds of AI already work in real estate

• Generative language models (LLMs). First-touch chat bots, rewriting and generating listing descriptions, answering routine questions in a messenger, drafting contracts and proposals — drafts, not the final document.
• Computer vision. Sorting and ranking listing photos, flagging blurry or duplicate shots, virtual staging of an empty apartment, estimating renovation condition from a photo.
• Speech technology (ASR + NLP). Transcribing call-centre and agent calls, automatically tagging a conversation against a script (was budget discussed, was the area confirmed), script-adherence checks.
• Automated valuation models (AVM). Price prediction from features — area, floor, condition, location, comparable deals — used as a prompt for the agent, not as a public price.
• Recommendation and matching. Surfacing listings by semantic similarity to a buyer's request instead of rigid filters.
• Agentic workflows. Several models and rules chained together to carry a lead through the funnel — qualify, book a viewing, remind, hand off to a manager — running as a background CRM process, not a chat window.

What makes brokerage different from a typical AI rollout

• The cost of a trust error is high. A deal is worth months of a client's income; a bot stating something untrue costs more here than the same mistake in e-commerce.
• Data sources are scattered. Avito, Cian, Domclick, phone calls, WhatsApp, Telegram, paper documents — rarely a single CRM from day one, so the first stage of any rollout is almost always tidying up data, not the model itself.
• Seasonality and load spikes. Demand jumps on weekends and after a rate cut; a bot closes the gap exactly when people are asleep or overwhelmed.
• Sensitive personal data and documents up front. Passport details, a mortgage agreement, a property registry extract — legally sensitive information that immediately pulls in Russia's personal-data law.
• Legally binding steps stay with a human. No model signs a sale agreement, files paperwork with the property registry, or handles a notarised transaction — this bounds where automation can actually reach.

When to bring it in — and when it's too early

Readiness signals worth checking before starting anything:

• enquiry volume of roughly 100–150 a month per agency or agent group — below that, manual handling is almost always cheaper;
• scenarios repeat — the same five to seven questions on first contact, not a new conversation from scratch every time;
• some CRM already exists (Bitrix24, amoCRM, even a homegrown spreadsheet with an API) — the model needs somewhere to write its output;
• someone is available to read the model's output and correct it for the first few weeks — without that, nobody tunes the prompt after the first mistakes.

A typical rollout order:

1. A first-touch messenger bot — qualification and answers to frequent questions.
2. Call transcription and analytics — without it, nobody actually knows what clients and agents say to each other.
3. Generating and proofreading listing descriptions — saves time, but needs an editor, never published blind.
4. A pricing hint (AVM) for the agent — a second opinion before meeting the seller, not a replacement for an appraiser.
5. A full agentic workflow — only once the first three steps run without raising questions from the team.

Too early looks like: a one-to-three-person agency with no CRM, where leads are already handled personally and weekly volume is less than a day's worth of manual triage — here AI adds more setup work than it saves.

What it costs and what pays back

The numbers below are indicative ranges for projects of this scale, not a quote for a specific agency — an exact estimate always follows a proper review of the task (see how that works on the Bitrix page).

• A qualifying bot wired into a CRM (webhook + LLM + writing results back into the deal) — this scope usually runs a week and a half to two weeks of development, i.e. 60–80 hours. At a freelance rate, that is roughly 174,000–232,000 ₽ one-off. A similar architecture is broken down in the post about ChatGPT in Bitrix24.
• Support and prompt tuning after launch — from 29,000 ₽/month (10 hours), until the team stops sending weekly corrections.
• Call transcription and analytics — a ready-made speech-to-text service typically runs 3–8 ₽ per minute of audio depending on how deep the analysis goes; for 500 calls a month at five minutes each, that's 7,500–20,000 ₽/month.
• Listing description generation — the per-description API cost is a handful of roubles, but the real saving is measured in copywriter hours, not tokens: a card that used to cost 300–500 ₽ from a freelancer becomes a draft an editor cleans up.
• Ready-made no-code bots (SaaS builders) — an alternative to custom development, 15,000–50,000 ₽/month for a subscription with no deep CRM integration out of the box: faster to launch, less flexible, and the data goes to a third-party service — which brings personal-data law back into the picture.

On effectiveness, similar rollouts tend to report the same pattern: first-response time drops from 30–40 minutes to one or two; the share of enquiries answered at night or on weekends goes from near zero to 60–80%; enquiry-to-viewing conversion rises 10–20%. These are ranges seen in projects of comparable scale, not a guarantee for any specific agency — it depends heavily on how those same enquiries were handled before.

What a qualifying bot conversation looks like

What follows isn't a real client's transcript but a typical first-contact scenario assembled from similar rollouts — this is what it looks like in practice, not in a slide deck.

23:47, enquiry from Avito

Client: Hi, is the 2-bedroom on Lenina 45 still available?
Bot: Good evening! Yes, it's available. Are you looking at a mortgage
     or cash?
Client: mortgage, pre-approved for 6.5M
Bot: Great, that fits the budget. When works for a viewing — a weekday
     evening or a weekend?
Client: Saturday afternoon
Bot: Booked provisionally for Saturday, agent Irina will confirm the
     exact time during business hours. Want me to send 3 similar
     listings in the same area?
Client: yes, please

[09:14, next business day]
System → CRM: deal created, tags "mortgage pre-approved",
"viewing: Saturday", owner — Irina.
Irina sees an already-qualified lead, not "is it still available?"

What a call breakdown looks like

The second example isn't a dialogue but what a manager receives after an agent's call with a client. The model doesn't take part in the conversation, only reviews the recording afterwards.

Call: agent Dmitry → client, 14 min 20 sec

Transcript (excerpt):
"...we looked at a one-bedroom in Voroshilovsky, but it's shared
construction, we want to avoid the risk... budget up to 5 million,
though for something really good we could stretch to 5.5... need it
this year, tired of paying rent..."

Automatic breakdown:
— Deal type: purchase, resale, priority — not shared construction
— Budget: 5,000,000 ₽, flexible up to 5,500,000 ₽
— Timeline: by year end (high urgency)
— Area: Voroshilovsky mentioned, no strong preference otherwise
— Objection not handled: client raised the shared-construction
  concern themselves, agent never asked what specifically worries them
— Recommendation: pin down the actual fear (delayed handover? double
  sales?) on the next call instead of changing the area without
  finding out why

What a monthly case report looks like

The third example is a summary an agency gets at the end of a month, three months into running a qualifying bot. Again, this is a composite illustration, not a report for a real client — but this is roughly the shape and the kind of numbers that come up when reviewing a rollout.

Monthly summary — qualifying bot, month 3

Enquiries handled by the bot:          412
Fully qualified:                       356 (86%)
Routed to manual review (low confidence): 56 (14%)

Average first-response time:
  before rollout:  34 min (business hours), no response at night/weekends
  after:           48 sec, around the clock

Enquiries answered within 5 minutes:
  before: 22%
  after:  91%

Viewings booked through the bot:       198
Enquiry → viewing conversion:
  before: 27%
  after:  34%

Cost: 212,000 ₽ one-off (integration) + 29,000 ₽/month support
Manager time saved: ~55 hours/month on initial enquiry triage

The legal side, in Russia

This isn't legal advice for a specific situation — before launching, it's worth checking with a lawyer or running a 152-FZ compliance check.

• Personal data law (152-FZ). A bot and call recordings collect names, phone numbers, sometimes passport details and mortgage information — that's personal data, and the requirements are the same as for a website: consent before collection, not after; notifying the regulator before processing starts, not once it's already running; the minimum fine for a sole trader is 30,000 ₽ just for the missing notification (covered in more depth on the 152-FZ site check page — there's also a quick self-check, the RKN quiz).
• Call recording. Clients need to be told a call is recorded — a voice notice at the start of the call — otherwise the recording can't strictly be used as evidence, and that's a separate ground for complaint on its own.
• Data localisation (242-FZ). Personal data of Russian citizens collected by a bot — including through a third-party SaaS provider or a foreign model's API — must first be recorded on a server in Russia. That constrains which model or transcription vendor is usable; it isn't just a formality.
• Advertising and consumer-protection law. A model-generated listing description is still an advertisement or an offer, and liability for inaccurate details (area, floor, extra rooms) sits with the agency, not the model or its vendor — a model has no legal personality. The practical consequence: generated listing text goes through an editor, never publishes automatically.
• Disclosing that it's a bot. As of September 2026 there's no single federal law requiring a real-estate chat to state "you're talking to a bot," but if a client explicitly asks "is this a person?", a dodgy answer becomes a reputational and potentially unfair-practice problem, not just an ethics one. Best practice: identify as a bot in the first message.
• Legally binding steps stay outside automation. Signing a sale agreement, filing with the property registry, acting under a power of attorney — these require a human's personal involvement and a qualified signature; none of the scenarios above sign or confirm a deal.
• The contract with the AI vendor. Using a foreign API (OpenAI and similar) counts as transferring personal data abroad, which also falls under 152-FZ and needs its own legal basis — either the subject's consent to cross-border transfer, or stripping identifying data before it reaches the model.

Where to start

If this compresses to one rule: start with the most boring, most measurable scenario — a bot on inbound enquiries — and only move further once it stops needing daily corrections. AI in brokerage isn't about replacing the agent; it's about the agent spending time on viewings and negotiation instead of repeating the same answer for the hundredth time in a day.