AI action authority in property management: prepare, approve, or execute?

Updated July 2026

“Handle maintenance” hides many actions: receive a report, acknowledge it, suggest severity, disclose access information, contact a vendor, schedule a time, approve spend, pay an invoice, and close the work order. Classify each effect separately. A safe acknowledgement does not grant authority to approve a $2,400 replacement or disclose a lockbox code.

Write the action as a verb plus object and channel: “send one appointment confirmation by the resident’s consented email,” not “manage communications.” Atomic wording makes inputs, limits, duplicate controls, cancellation, and evidence testable.

Apply the strictest boundary first. A consequential housing or binding legal decision stays human-owned in this framework. An action that moves external money, exposes sensitive data, or cannot be intercepted pauses for approval. Only routine, reversible, source-backed steps proceed to bounded-execution design.

The ceiling is not a predicted error rate. Even a highly accurate model should not decide an applicant’s housing access simply because its confidence is high. Authority is an organizational and legal-accountability choice, not a number emitted by the model.

An action policy should identify the authoritative lease, ledger, consent, work order, bank status, vendor record, or configuration value. It should also specify which version applies and how stale, missing, or conflicting data is handled. “Use the lease” is incomplete if two signed amendments disagree in the system.

Effective time matters. A rent reminder that was valid at 8:00 a.m. can become wrong after a payment posts at 8:03. Recheck volatile facts immediately before an external effect, use idempotency or duplicate suppression, and route uncertain state to the exception queue.

List the cases that must not continue: missing consent, unmatched identity, disputed balance, conflicting lease dates, stale vendor insurance, duplicate provider event, amount above cap, unsupported jurisdiction, or absent delivery channel. For each, name the queue, required evidence, response time, and escalation owner.

An exception should preserve the attempted action and reason for stopping. A generic “AI failed” message forces a landlord to reconstruct the whole case; “appointment confirmation withheld because the resident’s email consent is absent” creates a narrow, resolvable task.

An internal log can show that software intended to act without proving the provider accepted, delivered, posted, settled, or completed it. Keep the request identifier, provider response, status transitions, timestamp, actor or policy version, linked source records, and any cancellation or reversal.

Reconcile asynchronous outcomes. A payment marked “submitted” is not settled cash; an email handed to a delivery service is not necessarily delivered; a vendor dispatch is not a completed repair. Define which receipt closes the task and when a missing or contradictory outcome becomes an exception.

The same action can be permitted in one context and blocked in another. A manager may approve a routine vendor dispatch for one property but lack authority for an owner-controlled reserve at another. A resident may consent to email appointment updates but not text messages. A vendor approval may apply only while insurance, license, geography, trade, and contract records remain current.

Encode those boundaries as inputs to the action policy, not as notes a reviewer is expected to remember. Test role changes, sold properties, ended management agreements, revoked consent, expired credentials, temporary overrides, and daylight or quiet-hour rules. Every override needs an owner, reason, scope, start, expiration, and review trail so emergency access does not become permanent ambient permission.

Use a test set containing normal, missing, stale, conflicting, duplicate, threshold-edge, provider-failure, cancellation, and human-required cases. Review both false execution and false abstention: the first can create harm, while the second can quietly recreate manual workload.

Revisit authority when the workflow, provider, model, source record, regulation, or operating team changes. Expand limits only from reviewed evidence, and preserve the narrower rollback policy. A launch decision is not permanent permission.

This is general educational information, not legal or tax advice. Rules vary by state and change over time — confirm specifics for your jurisdiction with a qualified professional.

It is a control map that assigns each atomic action to preparation, approval-required execution, bounded execution, or human-only decision-making based on consequence, data, money, reversibility, evidence, and policy.

Aptoria never decides on an applicant, and a person makes every consequential housing decision, including applicant denials. Software may organize available evidence or draft a neutral review packet, but a responsible person owns the decision and required process.

No. Confidence may help a system abstain, but authority should depend on the action’s effect, current source records, role, limits, reversibility, privacy, exceptions, and outcome evidence.

Review it whenever the workflow, model, provider, data source, law, role, or risk changes, and on a scheduled cadence informed by exceptions and outcome evidence. Keep a tested rollback policy.