The buyer’s question looked routine: “Could I rent this unit later if my plans change?”
The listing assistant answered yes. It cited the condominium declaration in the property record and explained the lease requirements. The answer sounded careful because it was grounded in a real document.
It was also wrong.
The declaration had been amended three years later. The current rule imposed a rental restriction that materially changed the answer, but the amendment was not in the document set. The system had not invented a policy. It had relied on authentic evidence that was no longer complete.
Who is accountable?
The model generated the answer. The technology provider designed how sources were retrieved and how uncertainty was expressed. The agent supplied or approved the property material. The brokerage chose where the system could speak to consumers. The association maintained the governing documents. The buyer relied on the result.
Several parties touched the failure. That does not mean responsibility can be left between them.
Accountability in real-estate AI is not the search for one person to blame after something goes wrong. It is the prior assignment of who must verify the evidence, approve the use, monitor the system, make the consequential decision and repair the outcome.
The scenario is illustrative. This article addresses operational accountability, not a legal conclusion about liability in a particular case.
A shared system still needs individual owners
AI rarely acts alone. A consumer-facing answer may depend on a foundation model, a retrieval system, listing data, uploaded documents, brokerage rules and an agent’s approval. A pricing recommendation may combine public records, comparable sales, vendor models and a seller’s objectives. An underwriting workflow may involve applicant data, a scoring system, policy rules and a human reviewer.
When responsibility is described only as “shared,” each participant can point to another:
- The agent trusted the product
- The brokerage trusted the vendor
- The vendor relied on the supplied data
- The model produced a probabilistic output
- The consumer accepted a disclaimer
Every statement may contain part of the truth. Together, they can still produce a workflow in which no one owns the answer.
The better question is not, “Who controls the whole system?” Few participants do. It is, “Who controls this part of the risk?”
Are the sources authoritative, current and complete enough for this use?
Owner: agent, data steward or content owner
Does the system cite, abstain, escalate and preserve records as designed?
Owner: technology provider or internal product team
Is this an appropriate use, with the right permissions, limits and supervision?
Owner: brokerage or operating organization
Has a consequential recommendation or action received the required review?
Owner: the licensed or authorized professional
Can a person question, correct or appeal an answer and reach someone accountable?
Owner: the organization presenting the system
Will the specific failure be contained, corrected and prevented from recurring?
Owner: a named incident owner across the participating teams
In regulated decisions, opacity is not an excuse
There is no single rule that allocates every possible real-estate AI error. Existing law and professional standards instead attach duties to particular activities and actors.
A creditor “cannot justify noncompliance… based on the mere fact that the technology it employs to evaluate applications is too complicated or opaque to understand.” Applicants are still owed specific and accurate reasons for an adverse action.
Primary actor: creditorsThe Justice Department challenged Meta’s housing-ad targeting and delivery algorithms under the Fair Housing Act. Meta agreed to discontinue its “Special Ad Audience” tool for housing ads and to develop a system reducing the variance, measured by sex and estimated race or ethnicity, between the eligible audience for a housing ad and the audience that actually received it. The parties settled those compliance metrics in January 2023.
Primary actor: the advertising platformBuying a credible tool is not enough. Covered institutions need policies and control systems to ensure confidence in estimates, protect against data manipulation, avoid conflicts of interest, require random sample testing and reviews, and comply with nondiscrimination laws.
Primary actors: covered institutions using AVMs in covered mortgage decisionsArticle 2 requires avoiding “exaggeration, misrepresentation, or concealment of pertinent facts relating to the property or the transaction.” Article 12 requires presenting “a true picture in their advertising, marketing, and other representations.”
Professional standard for: REALTORS®, whoever or whatever drafted the wordsA lack of understanding of your own methods is not a defense.
Human review needs a job description
“A human remains in the loop” is often offered as the complete answer to accountability. It is not.
A person cannot provide meaningful oversight if the answer arrives after the decision, if the underlying evidence is hidden, if review is expected but no time is allocated, or if the reviewer lacks authority to stop the workflow. A name beside an approval box does not create judgment.
Useful human oversight answers five practical questions:
- What triggers review? The threshold might be financial consequence, fair-housing risk, conflicting evidence, low confidence, missing documents or a request for professional advice.
- What can the reviewer see? At minimum, the material sources, their dates, relevant conflicts, the system’s limitations and any earlier intervention.
- What can the reviewer do? The reviewer needs authority to approve, revise, refuse, escalate or disable the action, not merely acknowledge it.
- When does review occur? Before a consequential answer, recommendation or automated action reaches the point at which it is difficult to reverse.
- What is recorded? What was presented, what changed, who decided and why.
This is consistent with NIST’s AI Risk Management Framework, which asks that “roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented and are clear to individuals and teams throughout the organization,” that human-AI configurations be differentiated, that third-party systems and data carry their own controls, and that organizations plan to respond to, recover from and communicate about incidents. Its generative-AI profile narrows to four considerations: governance, content provenance, pre-deployment testing and incident disclosure.
The lesson for a brokerage is practical. “Agents must review AI output” is not a sufficient policy unless the product routes the right output to the right agent, shows the evidence needed for review and prevents the system from bypassing that review when the stakes rise.
The vendor is responsible, but not for everything
Technology providers should be accountable for the parts of the system they design and operate. That includes testing known failure modes, documenting material limitations, protecting data, monitoring performance, managing model or prompt changes, preserving useful logs and making escalation possible.
The deployer still owns decisions the vendor cannot make: which consumers may use the system, which documents are approved, which questions require a licensed professional, what risk is acceptable and what happens when a person disputes the output.
The professional using the result owns a different boundary. An agent does not need to inspect model weights, but should understand whether an answer is supported, whether the subject requires specialized advice and whether the recommendation fits the client’s objectives. A lender cannot outsource an explanation it is legally required to provide. A property operator cannot treat an automated diagnosis as a completed repair.
Executive responsibility sits above these individual decisions. Someone in the brokerage or operating organization must decide that the benefits justify the residual risk, fund the necessary controls and suspend the system when its performance falls outside the intended use.
Contracts can allocate obligations among the parties. Disclaimers can clarify that an answer is informational or subject to verification. Neither substitutes for a functioning control. Telling a consumer to “verify everything” after presenting a confident, property-specific answer transfers work. It does not create accountability.
A correction is part of the product
Accountability becomes most visible after the system is wrong. A dependable incident path should allow the organization to:
- Contain the problem. Pause the affected answer, campaign, recommendation or automation when continued use could compound the harm.
- Reconstruct what happened. Preserve the question, the response, source documents and versions, system configuration, human review and downstream action.
- Correct the record. Fix the underlying property information or rule, not only the wording of one response, and tell affected people when a material correction is necessary.
- Test the surrounding failure. Determine whether similar properties, users or workflows could receive the same result.
- Verify the repair. Assign one person to confirm that the change reached production, the affected workflow was retested and the incident was closed.
NIST’s generative-AI profile emphasizes logging, version history, change records and the involvement of relevant actors because incidents often cross organizational boundaries. A model provider may need to investigate one layer while the deployer corrects data, communicates with a consumer and changes an approval rule.
The objective is not a perfect audit trail for every generated sentence. The depth of evidence should match the consequence. A social-media caption and a credit decision do not require identical controls. A general neighborhood question and an answer about eligibility, protected characteristics, property condition or contractual obligations should not be treated as equivalent.
Accountability can become a product capability
The industry often treats accountability as a compliance cost added after the useful product has been built. In practice, many of the same controls also make the product better.
Source dates help an agent spot stale documents. Citations help a buyer understand an answer. Explicit uncertainty prevents a plausible guess from becoming a false fact. Escalation preserves the question and its context for the professional who must respond. Corrections improve the property record for the next user. Incident analysis reveals where the workflow, not only the model, needs to change.
This is also how we think about the problem. Grounding an answer in property information is only the beginning. The system should preserve where the answer came from, distinguish evidence from inference, recognize questions that require professional judgment and carry the buyer’s context into the agent handoff.
Those mechanisms do not remove responsibility from the agent, brokerage or technology provider. They make each party’s responsibility easier to exercise.
Accountability, in that sense, is not a brake on useful AI. It is part of what makes the intelligence usable.
The answer needs an owner before the question is asked
Return to the rental question.
A dependable system would not need to know the final rule in advance to behave responsibly. It could identify that the available declaration was dated, state that later amendments might control, avoid a definitive answer and route the question to the agent with the cited document and a missing-information warning attached. The agent or the appropriate association contact could verify the current restriction. The confirmed answer and current document could then update the property record.
In that version, the model does not become infallible. The workflow becomes accountable.
Across pricing, underwriting, listing promotion, buyer engagement and property operations, the same principle holds. AI can collect evidence, identify patterns, prepare explanations and recommend actions. The more consequential the outcome, the more clearly the system must show its evidence, respect its boundary and identify the person authorized to decide.
The final question is therefore not whether the vendor, broker, agent or model is accountable in the abstract. It is whether the workflow has assigned each responsibility while there is still time to act.
Accountable AI is not AI that never makes a mistake.
It is AI used inside a system where every consequential answer has an owner, an evidence trail, a boundary and a path to repair.
Sources
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023. GOVERN 2.1 is the source of the quoted passage on roles and lines of communication.
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024. Its four primary considerations are governance, content provenance, pre-deployment testing and incident disclosure.
- Consumer Financial Protection Bureau, Circular 2022-03: Adverse Action Notification Requirements in Connection with Credit Decisions Based on Complex Algorithms, 26 May 2022.
- U.S. Department of Justice, Civil Rights Division, United States v. Meta Platforms, Inc. (S.D.N.Y. 1:22-cv-05187), settlement agreement filed 21 June 2022, with the VRS Compliance Metrics settled January 2023.
- Office of the Comptroller of the Currency, Federal Reserve, FDIC, NCUA, CFPB and FHFA, Quality Control Standards for Automated Valuation Models, 7 August 2024, effective 1 October 2025.
- National Association of REALTORS®, 2026 Code of Ethics & Standards of Practice, effective 1 January 2026.
- National Association of REALTORS®, You’ve Tried AI, But Can You Trust It?, 12 February 2026, reporting a Realtors Property Resource® survey of 225 NAR members.
Reactions
No reaction from you
