Autonomy

An AI agent can decide which tool to call and interpret the response, but it does not become the payment processor, accounting platform, messaging carrier, or maintenance vendor. Those systems own external states the model cannot truthfully infer. A timeout may occur before acceptance, after acceptance, or while the response is returning.

A provider receipt captures what the destination reported: provider and account, external object ID, internal action and idempotency identity, request and response times, supported status, and later changes. It excludes secrets and unnecessary resident data. The receipt narrows uncertainty; it does not prove that money finally settled, a tenant read a message, or a repair was satisfactory.

A landlord approves a $425 invoice. The agent sends a payment instruction and receives no response before its deadline. If it labels the bill unpaid and tries again, the vendor may receive $850. If it labels the bill paid, the first request may have failed before reaching the provider. Both errors begin with treating missing confirmation as a known outcome.

The safe artifact chain is explicit: approval record, action identity, outbound attempt, provider lookup or receipt, payment state, bank or processor settlement, ledger entry, and reconciliation result. An unknown step pauses consequential follow-up. The landlord sees one exception instead of a confident but invented conclusion.

A good explanation should cite the records that support it. “I sent the reminder” is stronger when paired with the carrier message ID and accepted status; “the rent is settled” needs the payment event and ledger reconciliation. When the receipt is absent or contradictory, the explanation should say what remains unknown and which lookup or reviewer owns the next step.

Data provenance extends the chain backward. It shows which lease, ledger balance, resident record, policy version, and approval the agent used before acting. The provider receipt extends it outward. Together they let an owner inspect authority, evidence, external effect, and correction without asking the model to grade itself.

Routine, reversible work can often proceed inside a configured boundary when receipts are reliable and exceptions are visible. Higher-consequence actions need narrower authority: money movement, screening, lease changes, and legal or safety communication should pause whenever source facts or external state are incomplete.

During a software evaluation, ask for a failed-call demonstration. Does the action remain pending? Can the system look up the original provider object? Is another attempt fenced by the same identity? Does the final ledger link the receipt and any reversal? The answers reveal more about operational autonomy than a staged conversation with a chatbot.

Dumping every payload into a log is neither safe nor useful. The product should retain the minimum fields that correlate authority, request, provider object, state, and correction while excluding credentials, full payment details, and unrelated resident information. The owner needs a readable view; technical support may need a protected deeper record.

Retention and export matter too. A receipt that disappears before a chargeback, lease dispute, or annual close cannot support the work it described. Ask which records survive cancellation, which timestamps use provider time versus local receipt time, and whether the exported chain remains understandable without access to the original dashboard.