Start with the failure modes
A voice agent demo shows the happy path. Your risk lives elsewhere: the caller who is angry, the system that is down, the question the agent should not answer, the number that should not be dialed. Evaluate each platform by asking how it behaves in those moments.
1. Control over what the agent can do
- Can you allowlist the tools the agent may call, and validate every argument?
- Are permissions checked by the server, independent of what the model requests?
- Can high-impact actions require human approval?
- Is there any path for the model to run queries or commands directly? (There should not be.)
2. Truthfulness
- Where do balances, prices, coverage or eligibility come from during a call?
- What does the agent say when the source system is unavailable?
- Can answers cite their source and version?
- How are outdated documents prevented from answering?
3. Testing and release
- Is there a test console with scripted scenarios and expected outcomes?
- Are published versions immutable, so you can trace a call to the exact configuration?
- Can you roll out gradually and roll back instantly?
- Are prompt-injection and escalation behaviors part of the tests?
4. Human handoff
- How does the agent decide to transfer, and can you configure the conditions?
- Does the person who receives the call see a summary and the reason?
- What happens at 2 a.m., or when no one is available?
5. Consent, recording and disclosure
- Are consent and suppression checked immediately before each outbound call?
- Can recording and transcription be disabled by region or workflow?
- Can you configure disclosure prompts?
- Who can access recordings and transcripts, and is that access logged?
6. Telephony and quality
- Which components handle media, and how are provider failures handled?
- Are latency, interruptions, time to first audio and disconnect causes measured per call?
- Can you set concurrency limits, destination restrictions and spend caps?
- Has the platform been tested on real phone networks, not only in a browser?
7. Evidence and operations
- Is there an append-only audit trail with actor, reason and approvals?
- Is there a kill switch for an agent, a campaign and all outbound calling, and how fast does it act?
- How are costs attributed per agent and campaign?
- What is the failure behavior if the policy service is unreachable? (Outbound should be blocked.)
Ask vendors to demonstrate, not describe. If a control cannot be shown in a trial environment, treat it as unproven.