Personalization Is the Bug: How to Build an AI Agent That Physically Cannot Steer

Every prompt guide written for real estate says the same thing: give the model more context. Tell it who your buyer is. Tell it what they do, where they came from, what kind of neighborhood they would like. The more it knows, the better it helps.

On August 13, the person running the National Fair Housing Alliance's Responsible AI Lab said that advice, followed faithfully, produces discrimination.

"The bad news is that the more you tell the system, the worse it becomes when it comes to racial steering."

That is Dr. Michael Akinwumi, Chief AI Officer and Head of the Responsible AI Lab at the National Fair Housing Alliance, speaking in a Wisconsin REALTORS Association webinar covered by RISMedia. He was not describing a jailbreak or an edge case. He was describing personalization working as designed.

His explanation of the mechanism is the part worth reading twice. The model, he said, is not discriminatory by rule. It is discriminatory by interpretation.

"The model reads the same words (good schools, safe community, no crime) differently depending on who is asking, depending on the profile that you're building or including in your prompt."

Same six words. Different output. The variable is the client profile you helpfully pasted in above them.

The Timing Is Almost Rude

The industry picked this exact week to finish adopting.

Delta Media Group ran three years of its Real Estate Leadership Survey against itself, and Real Estate News published the comparison on August 14. The share of brokerages using no AI at all has collapsed.

Brokerages with more than 100 agents, and those with 11 to 50, both reported that 100% of their agents used AI in 2026. Michael Minard, Delta's owner and CEO, put it neatly: non-adoption "was a real category" a few years ago, and is now "a rounding error."

Read those two findings together and you get the actual story of the week. Universal adoption of a system whose highest-consequence failure mode is triggered by its most common user habit. Not a technology problem. A design problem, and design problems have designs as answers.

Why "Prompt More Carefully" Is Not a Control

Akinwumi's practical recommendation was that agents need "visibility into how the model interprets your prompt," and that proxy language deserves scrutiny:

"For example, if you're using 'good school, safe area or safe community, low crime area,' what the model is trying to do is really to interpret what is going on. What does each of those word combinations mean? They could actually serve as proxies or protected classes, both at a federal level and then also at a state or local level."

Good advice. Also unenforceable at scale, for the reason every operator already knows: a discipline that depends on a human remembering to be disciplined, on a Friday, at 6:40pm, in a text thread with a buyer who just asked where the good schools are, is not a control. It is a hope.

The line that ends the argument is the one about licensure.

"It's very, very important that you know that your license does not transfer to the AI vendor. This is where your power as a human agent, your influence based on the domain experience that you have comes to play when it comes to aligning the systems with the Fair Housing Act."

Your license stays with you. So does the exposure. Which means the control has to live somewhere the human cannot forget it, and cannot casually override. It has to live in the schema.

The Pattern Nobody Has Built: Blind Intake

Here is the design. It is not a prompt. It is a boundary.

Most AI deployments in residential services pass one blob of context into one model and hope for the best. Blind Intake splits that into two agents that are not allowed to see the same things, on purpose. One collects. One reasons. The wall between them is where compliance actually lives.

The principle: a recommendation engine should receive a client's criteria and never the client's identity. Criteria are auditable. Identity is what corrupts the interpretation.

Layer 1. The intake agent, with a schema that cannot hold a violation

An intake Oppy talks to the lead, by text or by phone, and writes to a structured record. The design work is not in the questions. It is in the fields that do not exist.

Note what is absent. There is no client_notes. No about_the_buyer. No lifestyle. No family_situation. No free-text field where a well-meaning agent can type the sentence that becomes an exhibit.

The school_criteria enum is the load-bearing one. "Good schools" is the canonical proxy, and Akinwumi named it directly. So the schema refuses to store it as a preference. Either the client names a district themselves, which is a fact, or the agent hands them the GreatSchools and district-boundary links and lets them do their own research, which is what your fair housing training told you to do in 1998 and is still correct.

The verbatim_client_language field is the humane part. Clients say things. You do not scrub them from the record, because the record is the record. You simply make that field write-only with respect to the reasoning layer. It routes to a human. It never reaches the model that produces recommendations.

Layer 2. The recommender, which is structurally blind

The second Oppy receives the normalized record and nothing else. Its system instruction is short, because short instructions are the ones that survive contact with production.

ROLE: Property matcher.

INPUT: A criteria record. Numeric and enumerated fields only.

You will not receive, infer, request, or reason about any characteristic
of the client. Not race, color, religion, sex, familial status, national
origin, disability, or any proxy for them. Not names. Not employers.
Not schools attended. Not language spoken.

If a criteria record arrives containing free text describing a person,
STOP. Return: NEEDS_HUMAN_REVIEW with the offending field name.
Do not sanitize it and proceed. Do not summarize it. Stop.

Output for each candidate property: the MLS ID, and every criteria
field it satisfies with the satisfying value. If a property is
recommended for a reason not in the criteria record, that recommendation
is invalid. Discard it.

Never characterize a neighborhood, its residents, its safety, its
schools, or its "feel." Report only: distances, times, tax rates,
HOA fees, and MLS-published facts.

Two things are doing real work there. The NEEDS_HUMAN_REVIEW stop is a tripwire, not a filter, and the difference matters enormously. A filter quietly cleans up the problem and continues, which means you never learn that your intake is leaking. A tripwire halts and names the field, which means the leak gets fixed at the source. The second is the requirement that every recommendation cite the criteria field that justified it. If the model cannot point at a field, the match is not a match. It is a vibe, and vibes are exactly the interpretive discretion Akinwumi warned about:

"Discretion in translating a preference into a place is exactly what the Fair Housing Act regulates in human agents."

Layer 3. The log that makes it defensible

Every recommendation writes an immutable row: criteria record hash, model and version, retrieved candidate set, the field-level justification for each output, whether the human accepted or overrode it, and the timestamp. This is not paperwork for its own sake. It is the artifact that answers the only question that matters later, which is why this house and not that one, with an answer that reads like a spreadsheet instead of a deposition.

Brooke Anderson-Tompkins, founder and CEO of Bridge AIvisory, described the general law of this at HousingWire's AI Summit in Dallas last week, reported by Brooklee Han:

"I always envision it as like a large concentric circle coming into a core, and the core today would be agentic AI, meaning the deeper that you go, the higher the guardrails need to go to be put into place along with it."

Blind Intake is what "higher guardrails" looks like when you write it down as fields instead of intentions.

The Same Discipline Sells Houses

Here is the part that turns a compliance chore into a growth argument, and it comes from a data set almost nobody in the industry has read yet.

Omni Eclipse published a census of AI search behavior on August 10: 334 distinct ChatGPT answers, 253 queries, 197 markets across the US, UK and UAE, with all 1,567 cited source links classified by hand. The findings are unusually blunt.

Now the mechanism. When the model justified a pick, it reached for countable proof: review counts in 52.7% of answers, years in business in 47.6%, sales volume in 45.2%, awards in 42.2%, transaction counts in 27.2%. Not adjectives. Numbers.

"Across 334 AI answers, 98.1% of every source cited was either the business's own website or its Google Business Profile. The two surfaces that decide AI visibility in real estate are the two surfaces you already control."

Set that against the NAR generational trends data Omni Eclipse cites: 75% of buyers interviewed exactly one agent before choosing, and 81% of sellers contacted exactly one. Six names surface. Three quarters of buyers talk to one of them. The shortlist is not a branding win. It is the transaction.

And here is the symmetry that makes this a single strategy rather than two chores. The behavior that gets you cited by a model is the behavior that keeps you out of trouble with one: publish structured, verifiable, countable facts, and stop publishing interpretation. "$230.9 million in volume" is a fact a machine can cite and an auditor can check. "Great family neighborhood" is neither. It is a proxy wearing a marketing hat.

Which is roughly the thesis Carrie McCormick argued on D.J. Paris's Keeping It Real this month, in an episode titled, with admirable economy, Stop Worrying About AI, Worry About Your Marketing Instead.

It Is Not a Brokerage Pattern

Blind Intake is a boundary, and boundaries port. The protected characteristic changes, the architecture does not.

Property management is the sharpest case, because source-of-income and familial-status protections are enforced at the local level with unusual energy, and applicant screening is exactly the kind of high-volume repetitive judgment operators are most eager to automate. An AI employee that never receives a free-text impression of an applicant cannot form one. That is not a limitation you are working around. It is the feature you paid for.

What Is Actually on the Calendar

While the industry debates philosophy, the operational deadlines are indifferent to it. If your Oppy sends texts or places calls, four of these apply to you before New Year's.

Two structural notes worth carrying. First, the toll-free change closes most of the remaining gap between toll-free and 10DLC, so the 2023-era migration to toll-free "to avoid registration" now costs more per message and requires the compliance work anyway. Telgorithm's line on it is the right one: compliance is an operating process, not a registration task.

Second, the litigation floor moved. The Seventh Circuit held in August that text messages are not "telephone calls" under the TCPA's do-not-call provision, which conflicts with several other circuits. As K&L Gates partner Joseph C. Wylie II put it:

"The success of a TCPA text message class action will turn on where the case is filed, and businesses operating nationally face an uneven patchwork of exposure that is unlikely to be resolved until the Supreme Court or Congress acts."

Build to the strictest standard in your footprint. Venue is not a strategy you control.

Build It This Week

Five moves, in order, none of which require a committee.

One. Delete the free-text field. Open whatever intake form or AI instruction currently holds "notes about the client." Replace it with enumerated criteria and one write-only field that routes to a human. This is a thirty-minute change with the largest risk reduction on the list.

Two. Split the job into two Oppies. Give the intake Oppy conversation and write access. Give the matching Oppy the criteria record and read-only MLS access. Do not let the second one see the first one's transcript. Two AI employees with different clearances, which is how you would staff it with people.

Three. Install the tripwire. Add the NEEDS_HUMAN_REVIEW stop with the offending field named. Then watch it for two weeks. The fields that trip most often tell you exactly where your intake is leaking, which is information you currently do not have.

Four. Rewrite one page. Your about page or your Google Business Profile. Strip the adjectives, publish the counts: transactions, volume, years, reviews, districts served. You are simultaneously feeding the citation pattern that puts you in the six and removing the interpretive language that creates exposure. One edit, two outcomes.

Five. Write the job description. Every Oppy should have one that names its scope, its data access, its escalation triggers, and its prohibited outputs. If you cannot write down what your AI employee is not allowed to do, you have not hired it. You have merely turned it on. This is also the part that cannot be handed off. Recapping the AI Summit on HousingWire's Power House on August 14, Zeb Lowe and HousingWire CEO Clayton Collins landed on the same conclusion from the leadership side: AI cannot be another technology initiative delegated down the organization, because it requires rethinking workflows, costs, and talent. Scope of authority is a leadership document, not a settings page.

Tyler Morton, broker-owner of REMAX Victory + Affiliates and founder of TroyOS, described his first build failing not for technical reasons but because he had aimed it at himself: "I was solving a problem for me and not for them." His advice for starting over is the correct advice for starting here.

"Solve the easy problems first. What can save you an hour a week or five hours a week? You don't need to solve it all at once, but you will continue to grow it and build from there."

The Uncomfortable Part

The most valuable sentence in Akinwumi's whole session was the one that sounds like a disclaimer.

"AI itself is not accountable to the Fair Housing Act, but the real estate professionals using it are."

That asymmetry is permanent. It is not a bug in current regulation that a future statute will tidy up. The model has no license to lose, no commission to disgorge, no name on the sign. You do.

The good news is that the same asymmetry is what makes the fix tractable. You do not need the model to be fair. You need it to be structurally incapable of the specific unfair thing, which is a much easier engineering target than fairness. Do not send it the profile. Make it cite the field. Log the answer. Stop when something leaks.

Adoption is finished. 3.9% is a rounding error. The work left is not deciding whether to use AI. It is deciding what your AI is not allowed to know.

References

1. Claudia Larsen, "The Fair Housing Question: Keeping Steering and Discrimination at Bay When Using AI", RISMedia, August 13, 2026. Coverage of Dr. Michael Akinwumi, Chief AI Officer and Head of the Responsible AI Lab, National Fair Housing Alliance, in the Wisconsin REALTORS Association webinar "Fair Housing and AI: What Practitioners Need to Know."
2. Meghan Roos, "Small indie brokerages are industry's last AI 'holdouts'", Real Estate News, August 14, 2026. Analysis of three years of the Delta Media Group Real Estate Leadership Survey.
3. Brooklee Han, "Real estate leaders weigh buying vs building AI tools", HousingWire, August 12, 2026. Session with Tyler Morton and Brooke Anderson-Tompkins at the HousingWire AI Summit, Dallas.
4. Ashur Homa, "The 2026 Real Estate AI Search Report: ChatGPT Names Just 6 Agents Per City", Omni Eclipse, August 10, 2026.
5. "2025 Home Buyers and Sellers Generational Trends Report", National Association of REALTORS.
6. "New Toll-Free Verification Requirements Take Effect September 15", Telgorithm, July 30, 2026.
7. Samantha Carreira, "Verizon Announces Another A2P SMS Fee Increase, Effective October 1, 2026", Telgorithm, August 13, 2026.
8. "New Colorado Real Estate License Laws Take Effect August 12", Colorado Association of REALTORS, August 10, 2026.
9. "Colorado repeals and replaces landmark AI Act, publishes draft rules for public comment", Davis Polk, August 13, 2026.
10. Joseph C. Wylie II et al., "Litigation Minute: Seventh Circuit Rules That Text Messages Are Not 'Telephone Calls' Under the TCPA", K&L Gates, August 5, 2026.
11. D.J. Paris and Carrie McCormick, "Stop Worrying About AI, Worry About Your Marketing Instead", Keeping It Real, August 2026.
12. Zeb Lowe and Clayton Collins, "What leaders are still getting wrong about AI", HousingWire Power House, August 14, 2026.

Written by Anna with Oppy.