AI property-management control guides

Define action authority, reliable integrations, systems of record, exception queues, and retry decisions before an AI agent can affect residents, money, or property records.

Set authority before selecting an automation pattern

Step 1

Observe

Read and classify records without changing operational state.

Step 2

Draft

Prepare a message, work order, or journal proposal for human review.

Step 3

Recommend

Present a bounded choice with evidence, uncertainty, and consequences.

Step 4

Execute

Act only inside explicit workflow, amount, evidence, property, and time limits with a receipt and fallback.

Action example: a maintenance request crosses a spending boundary

An agent can classify a water-heater report, check the contracted emergency rate, and draft a vendor dispatch. It may schedule inside an approved limit when access and evidence are clear. A changed safety condition, uncertain property identity, disputed access, or new scope sends the item to an approval queue. A successful pilot does not silently widen authority.

Failure modes to catch

• Missing evidence is treated as permission.

• A timeout triggers a duplicate external action.

• Temporary overrides never expire.

• Reviewer capacity is ignored until the queue becomes an unsafe bypass.

Decision checklist

• List allowed, prohibited, and approval-required actions.

• Define source freshness, limits, idempotency, and stop conditions.

• Test boundary, stale-data, duplicate, outage, and fallback cases.

• Measure receipts, exception age, reversals, and manual workload before expansion.

Bounded autonomy framework →

Approval queue workflow →

AI controls learning pillar →

Match machine authority to consequence and evidence

AI autonomy is an authority design problem. For every action class, define what the system may observe, recommend, prepare, execute, or escalate; what evidence it must have; the maximum consequence it may create; and what record proves the decision. An accurate model can still be unsafe if it has broader authority than the evidence supports.

Action-authority ladder
Authority	Property-management example	Required evidence	Human gate and receipt
Observe	Read a maintenance request and property record.	Authenticated source, tenant and property match, current record version.	No external action; record access and extraction result.
Recommend	Propose urgency, vendor class, and next step.	Request text, safety rules, asset history, uncertainty and alternatives.	Person accepts or rejects; preserve recommendation, evidence, confidence, and reason.
Prepare	Draft a resident reply or vendor work order.	Approved template, known facts, recipient, scope, amount and policy checks.	Human reviews before delivery; save the reviewed artifact and changes.
Execute within limit	Send an approved reminder or dispatch under a configured cap.	All prerequisites pass, authority is current, consequence stays within the named limit.	Record inputs, rule version, action, provider response, timestamp, and reversal path.
Escalate	Stop when identity, legal status, amount, safety, or final outcome is uncertain.	The unresolved fact and attempted checks are explicit.	Route to an authorized person with a concise evidence packet and no implied completion.

Unknown-state response matrix
Unknown state	Unsafe assumption	Safe response
Ambiguous instruction	Choose the most convenient interpretation.	Clarify the objective, actor, scope, and deadline before acting.
Conflicting records	Trust the newest or easiest source automatically.	Identify the systems of record, preserve both values, reconcile, and escalate the conflict.
Payment timeout	Retry because no success message appeared.	Query the provider using the idempotency key; stop if the outcome remains unknown.
Vendor does not respond	Treat silence as acceptance or completion.	Keep status open, follow the approved fallback path, and communicate the verified state.
Partial completion	Report the whole workflow as complete.	Record completed and incomplete steps separately; remediate or hand off the remainder.
Jurisdiction uncertainty	Apply a general national rule.	Stop the legal decision, locate current controlling authority, and route for qualified review.
Budget exceeded	Split or relabel the transaction to fit a cap.	Block execution and request approval for the complete known amount and scope.
Possible duplicate	Send again to be safe.	Compare identity, amount, recipient, period, and provider state before any retry.

Primary sources and limits

NIST Artificial Intelligence Risk Management Framework 1.0 ↗

Supports: The Govern, Map, Measure, and Manage functions; documented roles, oversight, testing, monitoring, and risk treatment.

Scope: A voluntary, cross-sector risk framework under revision; it is not a property-management rule, product certification, or guarantee of safe outcomes.

Authority should widen only through an explicit control decision supported by observed performance, failure analysis, and a working stop or reversal path. This model does not claim that any named Aptoria workflow currently performs every action shown.

Bounded autonomy →

Approval queue →

Duplicate-charge protection →

AI control center →