AI in a Real Estate CRM: Lead Scoring, Summaries, Follow-Ups

AI earns its place in a real estate CRM in five jobs: triaging new leads, drafting replies an agent approves, writing listing descriptions, setting follow-up reminders and summarising calls. It should not text leads on its own or score them before you have outcome data. On published API prices, the model cost is roughly $2–8 per agent a month at moderate volume.

If you would rather have it built for you, see how RAITHub would build this below.

This post is about the AI layer only. Whether to buy or build the CRM itself, and the order to build its core in, is covered in real estate CRM for agents: build vs buy. The wider PropTech picture is in the PropTech software development guide.

What can AI actually do inside a real estate agent's CRM?

It can read and write text quickly, which covers most of an agent's admin. It cannot know which lead will buy, and it cannot take legal responsibility for a message. The useful jobs, and the ones that disappoint:

Job	What the AI does	Who decides	Verdict
Lead triage	Reads a new inquiry and extracts intent (buy, rent, sell, value), budget, area, timeline and urgency into fields	Your routing rules, using those fields	Worth it from day one
Reply drafting	Writes a first reply from the inquiry and the listing record	The agent, who edits and sends	Worth it, with a human approval step
Listing descriptions	Turns structured listing data and agent notes into copy	The agent, who checks every fact	Worth it; facts must come from the record
Follow-up reminders	Spots promises in emails and notes ("I'll call Friday") and proposes a task	The agent accepts or dismisses it	Worth it; small and safe
Call summaries	Summarises a call transcript into notes, next steps and changed details	The agent confirms before it overwrites fields	Worth it if calls are already transcribed with consent
Predictive lead scoring	Guesses which leads will close	Nobody should rely on it yet	Wait until you have months of outcome data
Autonomous texting	Sends messages without a person	Nobody, which is the problem	Avoid; consent and liability sit with you

The pattern is the same in every "worth it" row: the model produces a suggestion or a structured field, and a rule or a person makes the decision. That keeps the CRM predictable when the model is wrong, which it will sometimes be.

Should AI score real estate leads?

Not at first. "Lead scoring" sold as AI is usually one of two things. The first is extraction: reading "pre-approved, need to move by March, two kids, near Lincoln Elementary" and filling in budget status, timeline and area. Language models are good at this, and it works on your first lead. The second is prediction: a number saying this lead is 72% likely to close. That needs your own history of leads, response times and outcomes, and it is a classic statistics problem rather than a chatbot one.

So start with extraction feeding plain rules: a lead with a timeline under 30 days and a pre-approval goes to the front of the queue. Record every lead's source, first-response time and result from the first day. After six to twelve months you will know whether a predictive model would beat your rules. The rules themselves are covered in real estate lead routing.

One limit applies to both. In the US, the Fair Housing Act prohibits discrimination based on race, color, religion, sex, national origin, familial status and disability, including steering people towards or away from areas (US Department of Justice). A model that reads free text will pick up details like "two kids" or a name. Do not let those fields drive routing, scoring or which listings a lead is shown. Strip them from the triage output, and test that they cannot change the result. This is general information; confirm with your adviser.

Can AI write and send follow-up texts to leads?

It can write them. Whether it should send them is a consent question first. Under the US Telephone Consumer Protection Act rules, calls made with an autodialer or an artificial or prerecorded voice to a mobile number need the "prior express consent of the called party", and telemarketing or advertising calls need "prior express written consent" (47 CFR 64.1200). The FCC has long treated text messages as calls for these rules. The same section bars telephone solicitations before 8 a.m. or after 9 p.m. local time and requires honouring the national do-not-call registry. In February 2024 the FCC also ruled that AI-generated voices count as "artificial" voices under the TCPA (FCC declaratory ruling 24-17).

Consent rules have changed several times since 2023 and state laws add their own, so treat this as general information and confirm with your adviser. For the software, the design consequences are clear enough:

• Store consent as data. Who consented, to which channel, with what wording, when, from which form, and when they revoked it.
• Check consent at send time, not when the draft is created. People reply STOP between the draft and the send.
• Respect the recipient's local time, which means storing the lead's time zone, not the agent's.
• Keep a person on the send button for anything the model wrote.

A minimal send guard in TypeScript, where the model never calls the messaging API directly:

type Channel = 'sms' | 'email'

interface Draft {
  leadId: string
  channel: Channel
  body: string
  approvedBy: string | null // agent who reviewed the AI draft
}

interface Db {
  smsConsent(leadId: string): Promise<{ grantedAt: Date; revokedAt: Date | null } | null>
  leadLocalHour(leadId: string): Promise<number> // 0-23 in the lead's time zone
  scheduleAtNextWindow(draft: Draft): Promise<void>
  send(draft: Draft): Promise<void>
}

export async function sendDraft(draft: Draft, db: Db): Promise<'sent' | 'scheduled'> {
  if (!draft.approvedBy) throw new Error('AI draft needs human approval')

  if (draft.channel === 'sms') {
    const consent = await db.smsConsent(draft.leadId)
    if (!consent || consent.revokedAt) throw new Error('No SMS consent on record')

    const hour = await db.leadLocalHour(draft.leadId)
    if (hour < 8 || hour >= 21) {
      await db.scheduleAtNextWindow(draft)
      return 'scheduled'
    }
  }

  await db.send(draft)
  return 'sent'
}

Each branch gets a test: no approval, no consent, revoked consent, 7:59 and 21:00 in the lead's time zone. How to test the model side, including drafts that invent facts, is in how to test LLM features.

How do you keep client data private when a CRM uses AI?

Send the model only what the task needs, from your own server, under an API agreement whose data terms you have read. In practice:

• Minimise the prompt. Triage needs the inquiry text, not the lead's full history or ID documents.
• Call the model from the server, never from the browser, so keys and other clients' data are never exposed.
• Keep the tenant boundary. In a brokerage CRM, a prompt must never mix two brokerages' records. Build the prompt from queries that are already filtered by tenant.
• Log prompts and outputs with the lead and agent IDs, with a retention period, so you can answer "why did it say that?"
• Tell clients. Your privacy notice should say that AI tools help process inquiries. Confirm the wording with your adviser.

How much does AI cost per agent in a real estate CRM?

Less than the CRM seat, on most workloads. Anthropic lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, Claude Sonnet 5.5 at $2 and $10, and Claude Opus 5.5 at $4 and $20 (Anthropic API pricing, checked 2 October 2026). OpenAI and Google publish similar per-token tiers (OpenAI API pricing, Gemini API pricing). An illustrative month for one busy agent:

Task (per agent a month)	Volume and size	All on Haiku 4.5	All on Sonnet 5.5
Lead triage	200 leads × 1,500 in / 200 out tokens	$0.50	$1.00
Reply drafts	300 drafts × 2,000 in / 300 out	$1.05	$2.10
Listing descriptions	10 listings × 1,500 in / 500 out	$0.04	$0.08
Call summaries	40 calls × 6,000 in / 400 out	$0.32	$0.64
Total		about $1.90	about $3.80

The same mix on Opus 5.5 is about $7.60. These are model costs only. They exclude transcribing calls to text, which is a separate speech-to-text charge, plus hosting, monitoring and the engineering time to build it. Anthropic also notes that its newer models' tokenizer produces about 30% more tokens for the same text, so keep a margin. The practical saving comes from routing: send triage and reminders to a small model, and only listing copy or tricky replies to a larger one. That is the idea behind PadhAI's 70/20/10 router, covered in LLM model routing and cost.

Buy, build or hire: which AI CRM route fits?

Route	Example	Cost signal	Choose this when
Off-the-shelf real estate CRM, using its built-in AI features	Follow Up Boss lists Grow at $69 per user a month (Follow Up Boss pricing, checked 29 September 2026); check each vendor's current AI features and add-ons	Per seat, monthly	You are a solo agent or a small team with a standard process. This is the right answer for most agents.
No-code automation around your CRM	Zapier or Make connecting your CRM to an LLM API (Zapier pricing)	Per task volume, plus API tokens	You want one or two AI steps, such as drafting a reply into the CRM, and can live without tests or approval workflows.
Custom AI layer, beside a bought CRM or inside your own	A server-side service that triages, drafts and summarises through the CRM's API	A fixed build, then model tokens	You run a brokerage or PropTech product, need tenant isolation, consent checks and audit logs, or AI is part of what you sell.

Why RAITHub for this

• The CRM core has been built. BlockEstate, a multi-tenant listing and inquiry platform RAITHub built, has lead routing and agent dashboards with each brokerage isolated. Its MVP shipped in 6 weeks. It has no MLS integration, and RAITHub will not claim one.
• The AI plumbing has been built. PadhAI, an AI tutoring platform RAITHub built, runs 11 services with a 70/20/10 LLM router and retrieval-augmented generation, the same cost and grounding problems an AI CRM has.
• QA-first. Consent checks, quiet hours and approval gates are written as tested branches, not left to the prompt. PropDesk, RAITHub's property management build, ships with 1,024 automated tests.
• Data handling. RAITHub signs NDAs and DPAs and works inside your controls; production data stays in your own cloud account and development uses synthetic data.

When you don't need us

• You are one agent or a small team. Use the AI features in your CRM subscription, and add a no-code step if one is missing.
• You want a fully automated texting bot. RAITHub will not build one that sends without consent checks and a human on the send button.
• You need a native mobile app. RAITHub builds web apps and PWAs that work on agents' phones, not native iOS or Android apps.
• You need a live MLS feed as the core. Pick a vendor with shipped integrations on your MLS.

How RAITHub would build this

• Scope: a server-side AI service beside your CRM, or inside a custom one, covering lead triage into structured fields, reply drafts with an approval queue, listing descriptions from the listing record, follow-up suggestions and call summaries.
• Guardrails: consent records and send-time checks, quiet hours in the lead's time zone, protected-characteristic fields excluded from routing, and per-tenant prompt building.
• Cost control: a model router, token budgets per brokerage and a usage dashboard.
• Evaluation: a test set of real-looking (synthetic) inquiries with expected fields, run in CI on every prompt change.

Timeline: 4–6 weeks for a fixed-scope MVP of the AI layer on an existing CRM; 6–12 weeks if it ships as a backend service with a full API for a PropTech product. You receive: automated tests and CI, handover docs and runbooks, and full IP under NDA. See SaaS development and the real estate industry page, or past work on /work. Next step: book the free 15-minute technical audit, then get a written fixed quote.

Last reviewed: 2 October 2026. API pricing checked on 2 October 2026.

Frequently asked questions

What is an AI CRM for real estate agents?

A CRM where language models handle text-heavy admin: extracting intent and budget from inquiries, drafting replies, writing listing copy, proposing follow-up tasks and summarising calls. The agent still approves what goes out and owns the client relationship.

Is AI lead scoring accurate for real estate?

Extraction is reliable from day one: reading budget, timeline and area out of an inquiry. Predicting who will close needs months of your own outcome data, so start with extracted fields feeding plain rules and record outcomes from the start.

Can an AI CRM text leads automatically?

It can, but US TCPA rules require prior express consent, and prior express written consent for telemarketing, for automated texts to mobile numbers, plus calling-hour limits. Store consent, check it at send time and keep a human approving AI drafts. Confirm with your adviser.

How much does AI add to a real estate CRM per agent?

On Anthropic's published prices, an illustrative busy month of triage, drafts, listings and call summaries costs about $1.90 on Claude Haiku 4.5 or $3.80 on Claude Sonnet 5.5 per agent, excluding transcription, hosting and build cost.

Is client data safe when a CRM uses AI?

It can be, if prompts include only what the task needs, calls go from your server, each brokerage's data stays separate, and prompts and outputs are logged. Read your model provider's API data terms and update your privacy notice.

Can I add AI to the CRM I already use?

Usually. Check the AI features your vendor already offers first. If they fall short, a separate service can read leads through the CRM's API, write fields and drafts back, and leave your agents in the tool they know.