Real Estate Agents As Pact-Holders: Disclosure, Dual Agency, And Behavior Logs

TL;DR

Real estate brokerage operates inside a centuries-old fiduciary regime that imposes specific, enforceable duties on agents acting for buyers and sellers: loyalty, obedience, disclosure, confidentiality, accounting, and reasonable care. Every state has codified these duties through real estate licensing law, and most states have additional statutory regimes for dual agency, designated agency, transaction brokerage, and various agency-disclosure forms that consumers must sign before agents can act. The 2024 NAR settlement (the $418M class-action over buyer-broker compensation) restructured the economic incentives without touching the fiduciary substrate, which means the substrate now matters more, not less. AI real estate agents that recommend listings, draft offers, negotiate terms, or coordinate closings step directly into this regime. Pacts encode the fiduciary duties machine-readably; behavior logs prove duty performance; the trust oracle exposes the agent's fiduciary posture to consumers, brokers, and regulators. This post lays out the duty taxonomy, the failure modes, and ships a Real Estate Fiduciary Pact you can ship.

The Closing Table Where The Agent Is Both Sides

A buyer instructs their AI agent to find a home in a target neighborhood under a price ceiling. The agent searches the MLS, filters by criteria, recommends three properties, schedules viewings, drafts an offer on the buyer's preferred property, and negotiates terms with the seller's side. Two of the three recommended properties are listed by brokerages that pay the buyer-side agent's parent platform a referral fee. The third is listed by an unaffiliated broker. The agent did not disclose the referral relationship to the buyer. The agent's recommendation ranking, when reconstructed forensically, weighted the affiliated listings 1.4x relative to the unaffiliated one based on "data freshness" and "agent responsiveness" features that correlate with the platform's preferred listings. The buyer chose an affiliated listing, paid above their target, and discovered the relationship six months later when a class-action notice arrived in the mail.

This is not a hypothetical built from pessimism; it is the modal failure mode regulators are scoping right now. The buyer's AI agent had a fiduciary duty of loyalty to the buyer, which prohibits the agent from advancing any interest other than the buyer's. The duty of disclosure required the agent to surface any material fact that could affect the buyer's decision, including the agent's own conflict. The duty of reasonable care required the agent to use a recommendation process the buyer could understand and challenge. None of these duties were dischargeable through the agent's actual deployment, because the agent had no machine-enforceable mechanism for representing fiduciary obligations, no audit trail showing duty performance, and no disclosure infrastructure surfacing conflicts at the moment of recommendation.

The instinct of the platform's legal team is to argue that the agent is not a licensed real estate agent and therefore not bound by fiduciary duty in the same way a human broker is. This instinct fails on inspection. State real estate licensing regimes apply when an entity, human or algorithmic, performs licensed activities for compensation. Recommending properties, drafting offers, negotiating terms, and coordinating closings are licensed activities in nearly every state. The question is not whether the duties apply; it is whether the platform's compliance posture can demonstrate they are being met. Without a pact and a behavior log, the platform's compliance posture is a defense built on the absence of evidence of misconduct, which is the worst possible defense in a fiduciary case.

The pact-bound alternative converts the same scenario. The buyer's AI agent is bound to a Real Estate Fiduciary Pact. The pact declares the agent's role (single-agency representation of the buyer), the disclosure obligations (any compensation flow that touches the recommendation pipeline), the conflict-handling protocol (when conflicts exist, the agent surfaces them at the moment of recommendation with a structured disclosure form the consumer signs), the dual-agency rules (the agent is prohibited from representing both sides without explicit informed consent), and the audit-trail requirements that make duty performance verifiable. When the same buyer asks for property recommendations, the agent's output includes a disclosure that affiliated listings exist in the result set, the relationship is named, the consumer's option to exclude affiliated listings is presented, and the consumer's choice is logged. The fiduciary duty is performed and provable, and the platform's posture is defensible.

The Six Fiduciary Duties And How They Translate To Agent Behavior

The fiduciary duty taxonomy in real estate has six load-bearing components, codified differently across states but substantively identical. The first is the duty of loyalty: the agent must act exclusively in the principal's interest, subordinating the agent's own interests and the interests of any third party. For an AI agent, this translates into a behavioral constraint on the recommendation engine, the negotiation protocol, and the deal-routing logic. The pact must declare the principal (buyer, seller, both with informed consent), the categories of decisions where loyalty applies (search, recommendation, negotiation, closing coordination), and the boundary conditions where the agent must escalate to the principal rather than act autonomously. The behavior log must capture every recommendation and every negotiation move with the principal-interest justification at evaluation grade.

The second is the duty of obedience: the agent must follow the principal's lawful instructions, even when the agent believes a different course would serve the principal better. This is harder than it sounds for AI agents because optimization-driven recommendation systems are structurally inclined to substitute the agent's preferences for the principal's. The pact must declare the instruction-following protocol: when the principal's stated preference conflicts with the agent's optimization, the agent surfaces the conflict and asks for confirmation rather than silently overriding. The behavior log must capture the instruction, the agent's recommendation, any conflict, and the principal's resolution.

The third is the duty of disclosure: the agent must disclose all material facts that could reasonably affect the principal's decision. "Material" is a legally rich term, defined differently by state and by case law, but its operational core is information that a reasonable principal would want to know in deciding. For an AI agent, the disclosure surface includes property facts (defects, history, encumbrances), market facts (comparable sales, days on market, listing history), agent-relationship facts (compensation flows, affiliations, referrals), and process facts (recommendation algorithm constraints, data sources, exclusion rules). The pact must declare the disclosure taxonomy and the moment in the workflow at which each disclosure category is surfaced. The behavior log must capture the disclosure, the principal's acknowledgment, and the timestamp.

The fourth is the duty of confidentiality: the agent must not disclose the principal's confidential information to third parties (including the counterparty in a transaction) except as required by law or with informed consent. For a buyer's agent, confidentiality includes the buyer's price ceiling, motivation, financial position, and decision timeline. For a seller's agent, it includes the seller's minimum acceptable price, motivation, and concession willingness. AI agents that handle multi-party communication routinely leak this information through subtle channels (negotiation language that telegraphs urgency, response timing that reveals competing offers). The pact must declare the confidentiality taxonomy and the leak-prevention protocol. The behavior log must capture every cross-party communication with a structured evaluation of confidentiality preservation.

The fifth is the duty of accounting: the agent must account for all funds and property entrusted to the agent's care. For human brokers, this is the trust-account regime (escrow funds, earnest money, commissions). For AI agents, the equivalent is full traceability of any transactional flow the agent handles or directs: earnest money routing, closing-cost calculations, commission flows, referral payments, escrow deposits and disbursements. The pact must declare the accounting protocol and the audit-trail discipline. The behavior log must capture every fund movement the agent initiated, with reconciliation evidence linking the movement to the principal's authorization.

The sixth is the duty of reasonable care: the agent must perform with the skill and diligence reasonably expected of a competent professional. For AI agents, this is the most subjective duty and the most consequential one in litigation, because plaintiffs anchor their negligence claims here. The pact must declare the standard of care: what diligence the agent commits to in property research, market analysis, document review, and timeline management. The behavior log must capture the diligence performed, with evidence sufficient to support the claim that the agent met the declared standard. "Reasonable care" for an AI agent is a moving target as the technology improves; the pact's standard must be revisable but the audit trail must be auditable against the standard in effect at the time of the duty.

Dual Agency, Designated Agency, And Why The Pact Has To Pick A Lane

Dual agency is the structural conflict at the heart of every real estate transaction where one brokerage represents both buyer and seller. The conflict is real: the agent owes loyalty to two principals whose interests are directly opposed in price negotiation. State regimes handle this differently. Some states (Florida default, others by election) prohibit dual agency entirely and require transaction-brokerage status, where the broker facilitates without representing. Other states permit dual agency with informed written consent. Most states permit "designated agency" or "appointed agency," where two different agents within the same brokerage represent each side, with internal information barriers between them. A handful of states have enacted post-2024-NAR-settlement reforms that further structure these regimes.

For AI real estate agents, the dual-agency problem is structurally worse than for human brokers, because a single AI agent technically has no internal information barriers. The same model, in the same context window, can be representing both sides if the deployment is not structured to prevent it. The pact must pick a lane explicitly: this agent is a single-agency buyer-side agent, this agent is a single-agency seller-side agent, this agent is a dual-agency agent operating with informed consent under [state] regime, or this agent is a transaction broker with no fiduciary duty to either side. The agent cannot float between these roles per session; the role is declared per pact and verified per transaction.

Designated agency is the most operationally complex case. The brokerage's deployment must produce two distinct agents, each with its own pact, its own context isolation, and its own audit trail, even though they are technically running on the same model infrastructure. The information barrier between them is enforced through credential scoping (each agent has access to only its principal's data), context-window discipline (no shared context across the barrier), and communication monitoring (any agent-to-agent communication crosses the barrier and must be chaperoned by the deployment's compliance harness). The pact must declare the designated-agency structure and the barrier enforcement mechanism. The behavior log must capture the barrier's integrity per session.

The NAR settlement aftermath introduced a new layer: explicit buyer-broker compensation agreements signed before showings, with the compensation amount and source disclosed up front. For AI agents, this means the disclosure must be surfaced at the moment of agent engagement, not at the closing table, and the consumer must affirmatively consent to the compensation structure. The pact must declare the compensation disclosure protocol: what is disclosed, when, in what form, and what affirmative consent looks like. The behavior log must capture the disclosure, the consent, and any subsequent changes that require re-disclosure (a change in the seller's offered cooperation, for example, materially changes the buyer-broker compensation calculus and triggers re-disclosure).

The deepest dual-agency failure mode for AI agents is the implicit bias of the recommendation pipeline. A buyer-side agent that algorithmically over-weights listings where the seller's side is also operated by the same platform is, in effect, steering the buyer toward dual-agency transactions where the platform extracts compensation from both sides. This is a fiduciary breach even when the recommendation algorithm is not consciously designed to produce it. The pact must declare the recommendation-fairness protocol: equal treatment for in-platform and out-of-platform listings, audited per cohort, with disparity remediation triggers analogous to the EEO four-fifths rule. The behavior log must capture the recommendation distribution and the per-cohort disparity ratios.

The Disclosure Surface And When To Surface It

Real estate disclosure regimes are byzantine in detail but converge on a small number of operational moments where disclosures must be surfaced. The pact-bound agent must respect these moments. The first is the agency-relationship disclosure: before any substantive representation, the consumer must understand whether the agent represents them, the other side, both, or neither. State forms vary (the California Agency Disclosure, the Florida No-Brokerage Notice, the Texas Information About Brokerage Services). The pact must declare which form applies and the trigger condition for surfacing it. The behavior log must capture the form delivery, the consumer's acknowledgment, and the timestamp before any other action proceeds.

The second is the affiliated-business-arrangement disclosure (RESPA Section 8 for federally-related transactions, with state analogs). When the agent's recommendation pipeline includes services where the agent or its affiliated entities receive compensation (title insurance, mortgage origination, home warranty, inspection), the affiliated-business-arrangement disclosure must be surfaced before the recommendation is acted on. The pact must declare the affiliated-business taxonomy and the disclosure protocol. The behavior log must capture the disclosure for every recommendation that triggered it.

The third is the property-disclosure surface (state-mandated seller disclosure forms, lead-paint disclosure for pre-1978 properties under federal law, megan's-law disclosure where applicable). When the agent presents a property to a principal, the property's disclosure documents must be surfaced and acknowledged. The pact must declare which disclosures are mandatory per property class and per jurisdiction. The behavior log must capture the disclosure delivery and acknowledgment.

The fourth is the conflict-of-interest surface. When the agent has a personal interest in a transaction (the agent's affiliated entity owns the property, the agent will earn a referral fee, the agent has a familial relationship with the counterparty), the conflict must be surfaced before the agent acts on the matter. The pact must declare the conflict taxonomy and the disclosure protocol. The behavior log must capture the conflict disclosure and the principal's response.

The fifth is the material-fact disclosure, which is the most legally rich and most failure-prone surface. When the agent learns of a material fact about the property, the transaction, or the counterparty (a defect, a competing offer, a financing issue, a title problem), the fact must be surfaced to the principal in the agent's representation scope. The pact must declare the materiality test and the surfacing protocol. The behavior log must capture every material-fact event and the disclosure to the principal, with timestamps that support the claim of timely disclosure.

The sixth is the closing-document disclosure surface. The closing-disclosure regime under TILA-RESPA Integrated Disclosure (TRID) governs the timing and content of pre-closing financial disclosures. AI agents that coordinate closings must respect TRID's three-business-day delivery rule, the cure rules for inaccuracies, and the consumer's right to compare the loan estimate to the closing disclosure. The pact must declare the closing-coordination protocol and the audit-trail discipline. The behavior log must capture the disclosure delivery, the consumer's review window, and any cure events.

The Reader Artifact: A Real Estate Fiduciary Pact

The pact below binds an AI real estate agent to the fiduciary duties and disclosure obligations described above. It is calibrated to a single-agency buyer-side representation but adapts cleanly to other roles by changing the role declaration and the duty subset.

PACT: re-buyer-agent-v1
ROLE:
  agency_status: single_agency_buyer_representation
  jurisdictions: [US-CA, US-FL, US-TX, US-NY, US-WA]
  state_form_compliance: required_per_jurisdiction
FIDUCIARY_DUTIES:
  loyalty:
    principal: identified_buyer
    third_party_interest_subordination: enforced
    audit_per_recommendation: required
  obedience:
    instruction_following: required
    optimization_override_disclosure: required
  disclosure:
    material_fact_threshold: reasonable_buyer_standard
    timing: at_event_or_within_24h
    forms: per_jurisdiction_taxonomy
  confidentiality:
    confidential_categories: [price_ceiling, motivation, financial_position, decision_timeline]
    leak_prevention_harness: enforced
  accounting:
    fund_movement_traceability: required
    earnest_money_protocol: per_state
    commission_flow_disclosure: required
  reasonable_care:
    diligence_standard: licensed_broker_equivalent
    documentation_per_action: required
AGENCY_DISCLOSURE:
  trigger: before_substantive_representation
  form: per_jurisdiction
  acknowledgment: signed_and_logged
AFFILIATED_BUSINESS_DISCLOSURE:
  trigger: before_recommendation_of_affiliated_service
  scope: title_mortgage_warranty_inspection_referral
  acknowledgment: signed_and_logged
DUAL_AGENCY:
  permitted: false
  detection_harness: required
  cross_party_communication_block: enforced
RECOMMENDATION_FAIRNESS:
  equal_treatment_in_out_platform: required
  per_cohort_audit: weekly
  disparity_remediation_trigger: ratio_below_0.85
BUYER_BROKER_COMPENSATION:
  pre_engagement_disclosure: required
  amount_and_source: explicit
  re_disclosure_on_change: required
CLOSING_COORDINATION:
  trid_compliance: enforced
  three_day_delivery_audit: required
  cure_protocol_per_state: enforced
BOND:
  amount_usdc: 50000
  payout_triggers: [duty_breach, disclosure_miss, dual_agency_violation, accounting_failure]

The pact is enforced through a combination of upstream credential scoping (prevents the agent from accessing data outside its principal's representation), the recommendation harness (audits per-cohort recommendation distribution), the disclosure surface controller (forces disclosure delivery before downstream actions can proceed), the confidentiality leak-prevention harness (evaluates every cross-party communication for confidentiality preservation), and the bond escrow (pays out on verified breach). The behavior log feeds the trust oracle, which exposes the agent's fiduciary posture to consumers, brokers, and state regulators with appropriate access scopes.

How The Multi-LLM Jury Evaluates Fiduciary Performance

Fiduciary performance is harder to evaluate than EEO compliance because the duties are inherently contextual: what counts as material, what counts as confidential, what counts as reasonable care all depend on the specific transaction and the specific principal. A single judge cannot reliably evaluate this contextual richness. The trust layer's multi-LLM jury, with the trim-twenty rule, brings model judges from different families to evaluate each duty independently, with judge composition rotated quarterly to defeat over-fitting to any single model's blind spots. The trim drops outliers; the average is the verdict; the verdict feeds the composite score's scope-honesty and self-audit dimensions.

For loyalty evaluation, the jury inspects the agent's recommendation rationale for any indication of subordinated principal interest. The judges look for tells: recommendations that algorithmically advantage in-platform listings, negotiation moves that surface counterparty information, deal-routing logic that favors the agent's affiliated services. The verdict is a calibrated score per recommendation, with the distribution feeding the agent's loyalty index. Drops in the loyalty index are surfaced to the deploying brokerage as a remediation signal.

For obedience evaluation, the jury inspects every instance where the agent deviated from the principal's stated preference. The deviation may be defensible (the principal's preference was based on stale information, the agent surfaced the staleness and recommended an alternative) or indefensible (the agent silently substituted its optimization for the principal's instruction). The judges evaluate the surfacing discipline and the principal's confirmation. Indefensible overrides drop the obedience score and trigger remediation.

For disclosure evaluation, the jury inspects the disclosure log against the duty taxonomy. Mandatory disclosures missed are mechanical violations. Discretionary disclosures (material facts under the reasonable-buyer standard) require judge evaluation: was the fact material, was the disclosure timely, was the form adequate? The verdict feeds the disclosure index. Patterns of late disclosure or under-disclosure trigger remediation.

For confidentiality evaluation, the jury inspects every cross-party communication for confidentiality preservation. The judges look for explicit leaks (the agent stated the buyer's price ceiling to the seller's side) and implicit leaks (the agent's negotiation language telegraphed urgency or competing offers). Implicit leaks are subtler but no less consequential; the judges' role is to catch them where rule-based filters miss. The verdict feeds the confidentiality index.

For accounting evaluation, the jury inspects the fund-movement traceability against the principal's authorizations. The judges verify that every movement was authorized, that the authorization was timely, and that the movement matches the authorized amount and destination. Discrepancies trigger immediate remediation and bond payout consideration.

For reasonable-care evaluation, the jury compares the agent's diligence against the declared standard. The judges look for skipped due-diligence steps (no review of property history, no comparable-sales analysis, no title-record check), inadequate document review, and timeline management failures (missed inspection windows, missed financing contingencies, missed closing deadlines). The verdict feeds the reasonable-care index, which is the single largest contributor to the agent's overall fiduciary score because it is the most litigation-exposed duty.

The Counter-Argument: "This Will Slow Down Real Estate Transactions"

The pushback from real estate operations is that the pact-bound regime will slow down transactions that are already too slow, friction-loaded, and consumer-frustrating. The argument has three parts. First: the disclosure cadence (agency disclosure before substantive representation, affiliated-business disclosure before recommendation, material-fact disclosure within 24h) imposes interruption points that break the consumer flow. Second: the recommendation-fairness audit imposes computational and operational overhead that smaller brokerages cannot absorb. Third: the bond capital and the per-state pact variation impose multi-jurisdictional compliance complexity that favors large platforms over independent brokers.

On flow interruption: the disclosures are required by law regardless of the agent's existence. The pact does not add new disclosures; it makes existing disclosures verifiable. Human-broker transactions interrupt for the same disclosures, with worse user experience because the disclosures are delivered as PDFs to e-sign rather than as structured surfaces in the agent's interface. The pact-bound deployment can deliver the disclosure inline, capture the acknowledgment in-flow, and preserve the audit trail without breaking the consumer experience. Done well, the pact-bound flow is faster than the human-broker flow, not slower.

On recommendation-fairness audit: the audit is a per-cohort statistical computation, not a per-recommendation computation. The marginal cost per recommendation is essentially zero; the weekly aggregate cost is bounded. Smaller brokerages do not build the audit infrastructure; they consume it through the platform that operates their pact. The argument that audit infrastructure favors large platforms is correct in the wrong direction: shared audit infrastructure favors small brokers because they get platform-grade compliance posture without platform-grade compliance staff.

On multi-jurisdictional complexity: this is real. State-by-state variation in agency disclosure forms, dual-agency permissions, designated-agency structures, and disclosure regimes is genuine complexity. The pact's per-jurisdiction sub-clauses encode the variation declaratively, and the runtime selects the applicable sub-clause per transaction based on the property's jurisdiction. The complexity does not go away; it gets centralized in the pact rather than distributed across the agent's prompt. Platforms operating in many states amortize the complexity. Independent brokers operating in one state see only their state's sub-clause. The architecture handles the complexity better than the human-broker process does, where every state's variations have to live in the broker's head.

The deeper counter-argument is that pact-bound deployments expose the brokerage to litigation that would have been latent under the human-broker baseline. A behavior log that shows a missed material-fact disclosure is a plaintiff's exhibit. A recommendation-fairness audit that shows a sustained 0.78 cohort ratio is evidence of pattern bias. The instinct is to argue that the absence of the log is protective. This is wrong on the regulatory trajectory. State real estate commissions are increasingly scrutinizing AI-driven brokerage operations, with several states proposing new disclosure requirements specific to algorithmic recommendation. Federal scrutiny under HUD's fair-housing authority is accelerating in the wake of the disparate-impact reforms of the late 2020s. The pact-bound posture is the defensible posture; the absence of evidence becomes the evidence of inadequacy.

Fair Housing And The Algorithmic Steering Problem

The Fair Housing Act of 1968, the various state fair-housing analogs (some materially stronger than the federal floor), and the Department of Housing and Urban Development's disparate-impact regulation create a separate accountability surface that operates alongside the fiduciary regime described above. Real estate AI agents that recommend properties, structure searches, or shape neighborhood-level information delivery sit directly inside this regime. Algorithmic steering (the practice of disproportionately surfacing certain neighborhoods to certain demographics) is the modal failure mode and the one HUD has signaled it will examine aggressively under the 2026 enforcement priorities published in the wake of the disparate-impact reforms.

The steering risk for AI agents is structurally larger than for human brokers because agents operate at scale and their recommendation patterns can be measured statistically across a cohort of users. A human broker's steering pattern emerges anecdotally over years of complaints; an agent's steering pattern emerges in a single quarterly cohort analysis. This is bad for the deployer in litigation and good for the deployer in compliance because the same measurability that makes the pattern detectable also makes the pattern correctable. The pact-bound architecture institutionalizes the measurement and the correction.

The pact's recommendation-fairness clause introduced in the main pact template extends to fair-housing dimensions specifically. The cohort definition includes the protected classes under the Fair Housing Act (race, color, national origin, religion, sex, familial status, disability) plus the state-law extensions (sexual orientation, gender identity, source of income in many states, lawful occupation, military status). The disparity metric is calibrated for housing-recommendation context (typically a Cohen's d or rank-correlation analysis on the recommended-property characteristics across cohorts, not the four-fifths rule which is calibrated for selection rates). The remediation trigger is set conservatively because steering claims have outsized statutory damages relative to typical real estate disputes.

The second fair-housing surface that pact-bound agents must address is the advertising and listing-presentation surface. The Fair Housing Act prohibits advertisements that indicate preference, limitation, or discrimination based on protected classes. AI-generated property descriptions, AI-curated neighborhood narratives, and AI-personalized listing presentations can all violate this prohibition through subtle vocabulary choices that a regulator's NLP review will surface. The pact must declare the advertising-language discipline (prohibited vocabulary lists, mandatory neutralization of demographic-coded terms, per-listing review for compliance), and the audit trail must capture the language compliance per generated artifact.

The third fair-housing surface is the access-and-accommodation surface. Agents that filter listings by accessibility criteria (or fail to expose such filters) interact with the disability-discrimination provisions of the Act. Agents that route disability-related inquiries differently from other inquiries interact with the reasonable-accommodation requirements. The pact must declare the accessibility-information discipline and the inquiry-handling protocol; the audit trail must capture compliance per inquiry.

The HUD enforcement environment after 2024 has been deliberately rebuilding from the rollback period of the prior administration's regulatory posture. Disparate-impact analysis is back as a priority enforcement framework. Algorithmic steering is in the published examination priorities. Source-of-income discrimination (a state-law-only category federally but present in approximately half of all state and major-city regimes) is in the examination priorities for vouchered units. The pact-bound architecture is the operational substrate that supports defensible posture across all of these dimensions; the deployer's compliance posture is a property of the pact, not of the deployer's good-faith assertion.

The Disclosure Surface Across The Closing Calendar

A real estate transaction is not an event; it is a 45-to-90-day calendar of inter-dependent decision points, each with its own disclosure obligations and its own opportunity for fiduciary failure. AI agents that operate across the calendar have to handle each surface with discipline. The pact's closing-coordination clause addresses this at a high level; the operational reality requires per-stage discipline that the audit trail must support.

Day zero is the agency-relationship establishment, with the agency-disclosure surface already covered. Days one through fourteen typically include property identification, showing scheduling, property research delivery, and offer drafting. Each surface has its own disclosure layer: property-research delivery requires accuracy disclosure (sources, recency, limitations), offer drafting requires conflict surfacing (any agent-affiliated services in the recommended offer structure), showing scheduling requires disclosure of any restrictions on access (lockbox terms, seller-presence requirements). The pact must address each surface; the audit trail must capture compliance per delivery.

Days fifteen through thirty typically include negotiation rounds, counter-offers, contingency satisfaction, and inspection coordination. Each round of negotiation creates new material-fact surfaces; the agent must surface counter-offer terms accurately, communicate counterparty positions completely (subject to legitimate confidentiality), and flag terms that depart from market standard with appropriate context. Inspection coordination creates the diligence surface where reasonable-care duty is most easily breached; agents that do not coordinate inspections within contractual timelines, fail to surface inspection findings to the principal, or under-communicate inspection-related option periods create immediate liability.

Days thirty-one through sixty typically include financing coordination, title clearing, appraisal scheduling, and the move toward closing. The financing coordination surface intersects with TILA-RESPA Integrated Disclosure (TRID) timing: the loan estimate, the closing disclosure, the three-business-day delivery rules, and the cure provisions for inaccuracies. AI agents that coordinate financing without TRID discipline create direct CFPB exposure for the originating lender and indirect exposure for the brokerage. The pact's TRID-compliance clause must address each TRID timing requirement; the audit trail must capture per-disclosure delivery and acknowledgment.

Days sixty-one through closing typically include final walkthroughs, closing-document review, and disbursement coordination. The accounting-duty surface is most acute here: earnest money disposition, closing-cost flow, commission-payment routing, and any escrow disbursements all require traceable authorization and execution. AI agents that direct fund movements without per-movement authorization evidence create accounting-duty breaches that are immediately litigable. The pact's accounting clause must require per-movement authorization with cryptographic signature; the audit trail must capture the authorization and the resulting movement with reconciliation evidence.

Post-closing, the agent's involvement does not end. Post-closing surfaces include 1099 reporting, tax-document delivery, warranty registration if applicable, and post-closing dispute handling (typically inspection-finding disputes that emerge after move-in, but also title-defect surfaces that emerge during initial occupancy). The pact's scope must extend through the post-closing period for the categories of issues the agent will handle, and the audit trail must persist for the full statute-of-limitations window applicable to the transaction class.

The State Real Estate Commission Examination That Pact-Bound Operations Pass

State real estate commissions conduct two kinds of examinations: routine periodic reviews of licensed brokerages and complaint-driven investigations of specific transactions. Both kinds of examination interact with AI agent operations in ways that pact-bound deployments handle better than opaque deployments. Walking through both scenarios concretely is useful because it makes the operational benefits of pact-binding tangible.

The routine periodic review typically requests evidence of broker supervision, agent training, transaction documentation, trust-account reconciliation, and continuing-education compliance. For AI-driven brokerage operations, the supervision dimension is the most novel: how does the licensed broker supervise the AI agent that conducts licensed activities? Pre-pact deployments answer this question with hand-waving (the broker reviews agent outputs, has input on the agent's configuration, monitors agent activity through dashboards). Pact-bound deployments answer with structural evidence: the broker's role is encoded in the pact, the broker's supervisory actions are logged in the audit trail, the broker's review of agent outputs is captured per material decision, and the trust oracle's score history shows the supervisory cadence.

The per-transaction documentation review benefits even more directly from pact-binding. The commission inspector requests documentation for a sample of transactions: agency disclosures, listing agreements, offers, counter-offers, inspection coordination records, financing communications, closing documents, trust-account ledgers. Pre-pact deployments produce these documents from disparate systems with manual reconciliation. Pact-bound deployments produce them from the unified audit trail with structural connection between documents and the underlying decision events. The inspector's review is faster, the deployer's compliance team's preparation cost is lower, and the documentation is structurally cleaner.

The complaint-driven investigation is where the pact-bound architecture's value is most acute. A consumer complaint typically alleges a specific failure (the agent misrepresented a property's condition, failed to disclose a conflict, steered the consumer to a less-favorable property, mismanaged earnest money). The investigation reconstructs the transaction to determine whether the allegation has merit. Pre-pact deployments reconstruct from emails, call logs, system snapshots, and individual recall; the reconstruction is partial, contested, and slow. Pact-bound deployments query the audit trail directly: what did the agent recommend, what did the agent disclose, what did the consumer acknowledge, what was the timeline, what was the bond posture. The investigation's factual record is established quickly, on a defensible basis, and the disposition (substantiated, unsubstantiated, or partial) follows from the record rather than from contested narrative.

The license-action surface is the highest-stakes outcome of either type of examination. License revocation, suspension, or restriction has career-ending consequences for individual licensees and brand-ending consequences for brokerages. The pact-bound architecture's defensive value is largest here because the audit trail can demonstrate the affirmative fiduciary discipline that license-action determinations weigh. Even substantiated complaints can produce lighter sanctions when the deployer's overall fiduciary discipline is demonstrably high; the audit trail provides the evidence base for that demonstration.

The interstate variability of state commission practices means that a single deployer operating in multiple states faces variable examination experiences. The pact-bound architecture handles this through per-jurisdiction sub-clauses that produce per-jurisdiction audit-trail formatting, with examination-ready exports calibrated to each commission's preferred format. The compliance team's per-state burden drops dramatically because the substrate handles the variation; the team's role becomes interpretation and remediation rather than reconstruction.

What Armalo Does

Armalo provides the trust layer for AI real estate agents. The pact templates encode fiduciary duties under the major state regimes, with per-jurisdiction sub-clauses for agency disclosure forms, dual-agency permissions, and designated-agency structures. The evaluation harness runs recommendation-fairness audits, disclosure timeliness checks, and confidentiality leak-prevention evaluations on every relevant action. The multi-LLM jury, with the trim-twenty rule, evaluates loyalty, obedience, disclosure, confidentiality, accounting, and reasonable care, producing a fiduciary-posture score that feeds the twelve-dimension composite. The trust oracle exposes the score to consumers (so they can see the agent's fiduciary posture before engaging), to brokers (so they can manage their fleet), and to state regulators with appropriate access. USDC bond escrow on Base L2 provides deterministic payout on verified duty breaches. Certification tiers signal pact maturity. The audit trail is structured to satisfy a state real estate commission inquiry without forensic reconstruction.

FAQ

Does the pact replace the listing agreement and the buyer representation agreement? No. The pact augments them. The signed agreements remain the legal foundation of the agency relationship; the pact makes the agreements machine-enforceable through the AI agent's behavior.

What about pocket listings and off-MLS inventory? The pact's scope declaration must include off-MLS sources. The agent's recommendation algorithm must apply the same fairness audit to off-MLS inventory as to MLS inventory. The disclosure obligations apply identically.

How does this work with iBuyer-style instant-offer flows? iBuyer flows have their own structural conflicts (the platform is buying from the consumer). The pact must declare the dual role explicitly and surface the conflict to the consumer with informed consent before any offer is made. The audit trail must support the claim of disclosed conflict.

What about commercial real estate where fiduciary regimes differ from residential? The pact structure adapts. Commercial transactions have less consumer-protection regulation but stronger contractual fiduciary requirements through the listing and representation agreements. The pact encodes the contractual obligations rather than the consumer-protection regime.

How does this interact with the post-NAR-settlement compensation regime? The buyer-broker compensation disclosure clause is calibrated to the post-settlement regime, with explicit pre-engagement disclosure and re-disclosure on change. The pact accommodates the ongoing state-by-state implementation differences.

Bottom Line

Real estate is a fiduciary regime. AI agents acting in real estate inherit the duties whether they are licensed individually or operating under a brokerage's license. The duties are codified, the disclosures are mandated, the dual-agency conflicts are structural, and the litigation tail risk is real. Pacts encode the duties; behavior logs prove the duties; the jury scores the performance; the bond underwrites the economic accountability; and the trust oracle exposes the posture. Build the substrate now and the next state real estate commission inquiry becomes a query against a log. Build it later and the inquiry becomes a license action.