A voice AI project needs named owners for business policy and technical behavior. “The agency and developer will handle it” becomes unhelpful when a caller receives a false booking confirmation after hours.
Write a responsibility map before launch. Include client facts, service rules, escalation content, access authorization, implementation, testing, monitoring, usage billing, incident response, and offboarding. Shared participation can exist, but each decision needs a final owner.
Separate approval from implementation
The client or agency may approve business hours and dispatch rules. The developer implements the approved decision and verifies the action path. Qualified professionals should own domain-specific safety or clinical content where relevant.
Operational ownership matters after delivery. Decide who reviews pending requests, reconnects revoked access, updates holidays, and answers incident alerts. A developer cannot resolve an authorization question or dispatch a technician without the appropriate business authority.
Work through a fictional weekend incident
A scheduling integration loses access on Saturday. The developer can diagnose the rejected operation and suspend writes, but only the authorized client administrator can reconnect. The agency must coordinate customer follow-up under the service agreement.
A useful responsibility map identifies all three roles and the fallback while access is unavailable. Without it, everyone may believe the other party is monitoring the queue.
| Responsibility | Owner must be able to do |
|---|---|
| Business policy | Approve customer commitments |
| Access | Authorize and revoke connection |
| Technical operation | Diagnose, suspend, and recover actions |
| Staff follow-up | Resolve pending customer work |
| Client communication | Explain status and next steps |
Make the boundary testable
Rehearse one failed booking and ask each owner to perform their role. Verify contact availability, decision authority, and evidence access. A responsibility document that nobody can execute is incomplete.
Detailed responsibility maps add coordination overhead but reduce silent gaps. Overly broad developer responsibility can disguise business decisions as technical support, while overly narrow support can leave clients with no recovery route. Match the agreement to the service actually sold.
Review human call transfer for an example where telephony success and staff acceptance belong to different parts of the workflow.
Questions about delivery roles
Who approves the prompt?
Assign business-content approval and technical validation explicitly. A polished script may still encode an unsupported commitment.
Who handles third-party outages?
Define diagnosis, vendor contact, degraded operation, customer follow-up, and restoration ownership separately.
What should the client receive?
A clear responsibility map, contact routes, operating procedures, known limits, and the evidence needed to recognize when each owner must act.