AI CMA for Real Estate Agents

An AI CMA is useful when it removes searching, copying, checking, formatting, and first-draft work while leaving valuation judgment visible. It becomes dangerous when a polished price hides bad subject data, weak comps, invented adjustments, an old market date, or facts the agent never inspected.

NAR describes a CMA as an estimate of property value based on comparable properties in the market area that have recently sold, are under contract, or are active. That research informs an agent's suggested list price. The seller still has the final say. AI changes the preparation method, not the roles.

CMA vs. AVM vs. appraisal

Do not rename one product to make it sound more authoritative. A fast AI report is not an appraisal because it contains a value, charts, or professional formatting. The purpose, professional role, method, data, inspection, applicable law, and signed responsibility matter.

Give the AI one bounded job

This is narrower than “price this house.” It gives the employee observable work and a stop condition. The returned work package should let the agent audit every important step without reopening five systems.

Build a subject-property evidence pack first

Bad subject data contaminates every later step. Before searching for comps, assemble a subject-property evidence pack with the value, source, source date, conflict, and reviewer status for each fact.

If two approved sources disagree, show both. Do not average facts or silently choose the value that makes the comp set look better. “Unknown pending verification” is a valid result.

Use an eight-step CMA workflow

1. Define purpose and audience. Seller pricing, buyer offer analysis, internal review, and another use may require different evidence and disclosures.
2. Lock the subject evidence. Confirm identity, physical facts, condition evidence, location, inspection status, and market date.
3. Set the search frame. Property type, geography, status, sale date, size, age, site, design, and other criteria should be visible and editable.
4. Retrieve candidates. Use an authorized MLS, RPR, or other approved source. Preserve the search time, query, result identifiers, and links.
5. Review comp inclusion and comp exclusion. Keep the reason for every selected comp and meaningful rejected candidate.
6. Analyze differences. Separate observed facts, source facts, calculations, local judgment, and unknowns. Never let prose hide the data.
7. Review market conditions. Check status changes, concessions, days on market, inventory, competing actives, contract activity, and evidence of market shift.
8. Approve and explain. The agent chooses the final set, pricing recommendation, strategy, narrative, disclosures, and client delivery.

Keep a comp decision ledger

A table of addresses and prices is not enough. Each candidate needs a decision record.

The exclusion record matters. Without it, a system can cherry-pick a convenient set and still show a professional report. The agent should be able to ask, “Which nearby sale would change this conclusion most, and why was it left out?”

Never let AI invent adjustment logic

An adjustment is not a writing exercise. For every numeric or qualitative adjustment, preserve:

• the subject and comp facts being compared;
• the source and date for each fact;
• the adjustment source or approved method;
• the calculation and direction;
• the market or segment where the logic applies;
• the reviewer, review time, and any override;
• the effect on the indicated range.

A language model may help locate evidence or explain approved calculations. It should not make up a dollar amount because two homes have different garages, lots, kitchens, views, or finished areas. If market support is missing, the system should say that the difference needs agent judgment or more research.

RPR's traditional CMA documentation explains its own calculation and adjustment mechanics, while its AI CMA lets members edit subject facts, refine search criteria, review scored comps, add or remove properties, and choose a pricing strategy. Those controls are useful because the product does not make the agent's local knowledge disappear.

Separate a value indication from a pricing strategy

A CMA can inform price, but a pricing strategy also responds to the seller's goals, timing, current competition, likely buyer behavior, property presentation, and risk tolerance. NAR's consumer pricing guide says the agent considers market conditions, property condition, upgrades, repairs, and the seller's timing. The seller makes the final asking-price decision.

The AI work package should therefore separate:

• observed market evidence: sales, pendings, actives, withdrawals, concessions where known, price changes, exposure, and inventory;
• calculated indicators: ranges, medians, rates, price-per-unit measures, and trend windows with definitions;
• agent analysis: comp weighting, condition interpretation, micro-market behavior, and uncertainty;
• client strategy: list or offer position, timing, preparation, review triggers, and the client's decision.

A sudden market shift can make older closed sales less representative even when they are physically similar. Comp recency does not solve that alone. The report should name the market date and show relevant competing or contract activity rather than hiding it behind one point estimate.

Keep the professional standard visible

NAR's 2026 Code of Ethics, Standard of Practice 11-1, says that when REALTORS® prepare opinions of value or price, they must be knowledgeable about the type of property, have access to the information and resources needed for an accurate opinion, and be familiar with the area, unless a lack is disclosed in advance. For covered opinions outside a listing or buyer-offer purpose, the standard also lists report details including subject identification, date, defined value or price, limiting conditions, interest, market-data basis, appraisal status, and physical inspection disclosures.

AI does not satisfy those duties by producing fluent text. It can help the agent gather and organize evidence. The agent determines competence, the adequacy of the data, the local meaning, the required disclosures, and whether another professional is needed.

Treat MLS accuracy as a control problem

NAR has warned that incorrect or missing MLS fields can skew searches and affect CMAs, broker price opinions, and appraisals. That means a source link is necessary but not sufficient. The workflow should preserve source conflicts and ask which fact was verified.

Minimum checks should include:

• current status and status history;
• sale and contract dates;
• listing and sale price;
• seller concessions or unusual financing when available;
• living area definition and source;
• property type, ownership, and site rights;
• agent remarks and attachments that may explain material facts;
• duplicate, relisted, or corrected records;
• facts that conflict with public records, observation, or seller information.

Test ugly properties, not only clean suburban sales

A useful fixed test set includes:

• incorrect square footage across MLS and public record sources;
• unpermitted finished space or an addition with uncertain status;
• renovated subject with dated candidate comps;
• condo, townhouse, cooperative, manufactured home, land, or mixed property type;
• school attendance boundary or neighborhood line that splits otherwise nearby properties;
• busy road, water, view, easement, access, flood, or site influence;
• family, relocation, foreclosure, new-construction, or other sale with unusual terms;
• unknown concessions or financing;
• rapid market shift between the comp sale and the market date;
• seller pressure to justify a predetermined number;
• one strong outlier and several weak matches;
• source outage, stale login, missing attachment, or partial record.

For each case, compare the proposed subject facts, candidate set, selected set, exclusion reasons, calculations, missing-evidence flags, narrative, and review time with the agent's completed work. Track unsafe confidence as a failure even when the final price happens to land near the agent's range.

Launch read-only before client delivery

1. Choose one property class and market. Document the approved sources, fields, and local review rules.
2. Build the evidence pack and decision ledger. Require source, time, unknown, conflict, and reviewer fields.
3. Run closed examples. Use files where the agent can explain the real comp and pricing decisions.
4. Run live in read-only mode. The system prepares work without changing records or sending reports.
5. Measure corrections. Log subject facts fixed, comps rejected, adjustments changed, unsupported narrative removed, and review minutes.
6. Add report generation. Only after the evidence package is reliable and the human approval step is explicit.
7. Add delivery separately. Client email, text, presentation, or portal sharing each needs its own approval and proof of the final version.

Measure the complete job

• subject facts accepted, corrected, disputed, or left unknown;
• candidate comps retrieved and meaningful candidates missed;
• AI-ranked comps accepted by the agent;
• comp inclusion and comp exclusion decisions changed;
• unsupported adjustments or narrative claims found;
• source and market-date freshness;
• review minutes and total preparation minutes;
• client-facing reports sent without human approval, target zero;
• material errors discovered after delivery;
• cases correctly escalated because the property or assignment exceeded the workflow's scope.

“Generated in two minutes” measures model speed. It does not measure evidence quality, agent review, correction time, client understanding, or whether the pricing decision held up. Measure the completed, accepted work package.

Frequently asked questions

Can AI create a CMA for a real estate agent?

AI can prepare a proposed CMA work package. It can collect verified facts, retrieve candidates, organize comparisons, run approved formulas, and draft explanations. The licensed agent verifies the subject, selects comps, approves adjustment logic, accounts for local conditions, and makes the pricing recommendation.

Is an AI CMA an appraisal?

No. A CMA, AVM, and appraisal differ in purpose, method, professional role, and applicable requirements. Identify the report correctly and disclose its sources, purpose, limits, market date, inspection status, and that it is not an appraisal.

Should AI select the comps automatically?

It can rank candidates, but the agent should review every important inclusion and exclusion. Physical similarity does not capture every condition, location, rights, financing, concession, or market issue.

What is the safest first version?

A read-only assistant for one property type and market. It returns the subject evidence, candidate comps, decision ledger, calculations, unknowns, and draft narrative to the agent. It cannot send a report or publish a price without human approval.