Can an AI system maintain responsibility for an outcome instead of only answering a prompt?
Context
Avorth is one startup with two strands of work. The first connected a texting assistant to Jobber and QuickBooks data. The second connected voice-model intake to a local operator desk. Both expose the same design problem: the model’s output must pass through software that knows what it is allowed to do.
What I did
- 01
Built an accounting-data adapter that rejects non-GET requests in code, even though the underlying accounting permission is broader.
- 02
Tracked whether paginated reads reached the end, so an incomplete fetch cannot silently become a claim of complete coverage.
- 03
Built a voice workflow that stores call evidence, prepares an exact action payload, and checks that payload’s hash at approval and execution.
- 04
Separated attempted delivery from a named person’s acknowledgement. In the prototype, silence can exhaust the escalation chain without being labeled completion.
Architecture / workflow
Approval is not delivery. Delivery is not acknowledgement.
What would actually count as a completed handoff?
- 01MODEL ASSISTANCE
Collect intake
Voice-model intake produces a durable call record
- 02SOFTWARE LOGIC
Bind the payload
Prepare stored content and check its hash
- 03HUMAN JUDGMENT
Approve that version
Reviewer approves the exact payload
- 04SOFTWARE LOGIC
Attempt the handoff
Local desk paging and alerts are simulated
SIMULATED / MODELED - 05SOFTWARE LOGIC
Check acknowledgement
Track acknowledgement by a named person
INSPECT THE FAILURE BOUNDARYThe payload changes, or nobody acknowledges the attempted handoff.
Stale or corrupted payloads are rejected. Silence can exhaust escalation without being labeled completion. These are inspected prototype behaviors; the suite was not rerun for this site.
This drawing covers the voice prototype. The separate accounting adapter is read-only; it does not create invoices. No live dispatch or autonomous business outcome is implied.
Where it gets difficult
An approval becomes ambiguous if the content changes afterward. The prototype binds the reviewer’s approval to the exact stored payload and rejects a stale or corrupted one.
An API can return a valid page while leaving more records unread. Coverage state is part of the result, rather than a hidden implementation detail.
A delivery attempt, a delivery receipt, and acknowledgement are different events. The state model keeps that distinction visible to the operator.
Limits & evidence
- The local voice desk simulates paging and alerts. It is not evidence of a production dispatch or customer-notification system.
- Provider demos and local implementations do not establish paying adoption or sustained business outcomes.
- “Keep completed jobs from remaining uninvoiced” is the persistent-responsibility direction. The read-only adapter does not create invoices or establish that this condition was maintained autonomously in production.
Read-only adapters and coverage state
Inspected accounting and field-service adapters reject accounting writes and expose whether paginated reads reached the end.
What this does not establish
Private code and account records are withheld. Code inspection does not establish production adoption, complete coverage of every API, or a fresh test run.
Approval, delivery, and acknowledgement
Engine code and test assertions were inspected for exact-payload approval, stale or corrupted hashes, and acknowledgement versus exhausted escalation.
What this does not establish
Paging is simulated. Assertions were read, not rerun for this site. This is a prototype account, not an independent security audit or proof of live dispatch.
What this changed
A chatbot can answer “which jobs are uninvoiced?” Maintaining that condition over time also requires durable state, coverage checks, authorization, recovery, and a way to verify what happened outside the conversation.