A voice-agent demo can sound convincing while doing very little. If you are hiring someone to deliver an agent for your agency, ask them to show what happened in the connected system after the call.
That request changes the conversation. You can inspect a customer record, a service request or a transfer outcome, rather than judging the developer by a pleasant voice and a polished dashboard.
What does a voice AI developer actually own?
The job can include conversation design, speech providers, phone routing, backend integrations, deployment and ongoing support. Those are separate responsibilities. A developer who handles one well does not automatically cover the others.
Start by describing one end-to-end task. For an illustrative home-service workflow, a caller explains the problem, the agent checks the service area, saves a request, and the office follows up. Nobody has promised a technician or a confirmed appointment at that point.
The brief should say which steps the developer owns and which remain with your team. Without that boundary, a proposal for an agent can quietly become a proposal for an entire operations product.
Ask for one normal call and one failed integration
I would use the same short assessment for every candidate. Ask them to demonstrate a normal request, then make an external dependency unavailable. They can use fictional data and a test system; private customer recordings are unnecessary.
| Ask the developer to show | What you are checking |
|---|---|
| The caller changes a detail | The latest value reaches the integration |
| A service rejects a request | The agent explains the failure without claiming success |
| The call ends unexpectedly | Staff can find unfinished work |
| The same operation is retried | The backend has an explicit duplicate-handling policy |
| The developer is unavailable tomorrow | Your team has usable setup and support notes |
A candidate may need time to prepare this. Agree on a small paid assessment when the work goes beyond showing an existing example. A substantial unpaid build is a poor substitute for a properly scoped engagement.
Read the proposal for missing responsibilities
Look for named integration accounts, test access, supported call paths and acceptance criteria. Ask whether the price includes phone usage, model usage, hosting, test calls and maintenance. If these are separate, the proposal should say so.
Ownership matters after delivery. Your agency needs to know where the code lives, who controls the phone number, how credentials are rotated, and who can roll back a release. Agree on those terms before providing access.
A launch date also needs assumptions. An engineer cannot finish a calendar integration against an API nobody has permission to use. List dependencies and put a date against the decisions your team must make.
Check project evidence without confusing a demo with client work
My Voho case study describes my contribution to an existing platform: LiveKit agents, tool calling, knowledge retrieval, SIP and work across the product and backend. It does not claim sole ownership of the platform.
TalkPatch is different. It is an independent demonstration of agent configuration and a receptionist workflow. Both can help you assess engineering work, provided the author explains what each example actually proves.
That distinction is useful when reading any portfolio. Look for a clear personal contribution, an inspectable artifact and stated limits. Treat unexplained business metrics as questions to investigate, rather than automatic proof of engineering quality.
Questions agencies ask before hiring
Do we need a developer or a managed platform?
If your workflow fits a platform's existing features and your team can maintain it, configuration help may be sufficient. Custom integrations, application features or unusual failure handling can justify a broader engineering role.
Should we hire hourly or agree on a fixed scope?
A fixed scope works when the task and acceptance checks are understood. An investigation with unknown causes may need a time-limited diagnosis before anyone can price the repair responsibly.
Does a successful test mean the agent is ready for every caller?
No. A test supports the scenarios and conditions it covered. Ask what remains untested, including language coverage, concurrency, transfers and recovery after service failures.
Before your next hiring call, write down one successful outcome, one failure outcome and the records that would prove each. Use those as the starting point for the engineering brief.