Voice AI platforms demo well. Almost any of them can hold a pleasant thirty-second conversation on stage. The differences that decide whether one works for your business show up later — in production, at volume, when a call goes sideways. This is a checklist of twelve questions that surface those differences before you sign anything, grouped by what they actually protect you from.
Does it finish the job?
1. Can it act in my systems, or only talk? The single most important question. A platform that "understands" callers but cannot write a booking into your calendar or a lead into your CRM has handed you a transcription service. Ask to see it complete a real task end to end.
2. What integrations ship out of the box? Completing tasks means connecting to the booking, CRM, telephony, and messaging tools you already run. Find out which are pre-built and which require a custom plugin — and how long that plugin takes to build.
3. What happens when the goal cannot be met? A good agent captures a structured fallback (a lead, a request) rather than dropping the call. Ask what the caller experiences when the ideal outcome is not possible.
Will it tell the truth?
4. Is it grounded, or does it improvise? Push on this hard. The agent must answer from a knowledge base you control and refuse to invent facts. Ask what it does when a caller asks something not in its knowledge — the right answer is "it says it doesn't know and hands off," not a confident guess. The knowledge-base guide is a good reference for what grounding should look like.
5. How does hand-off to a human work? When the agent escalates, does the person receive the full call context, or does the caller have to start over? A clean warm transfer with context is the difference between a rescue and a frustration.
6. Can it stay in its lane on sensitive topics? For regulated work — anything touching health, money, or legal advice — the platform must let you set hard boundaries the agent cannot cross, by design rather than by prompt-wishing.
Can I see and control what it does?
7. Can I inspect and edit its behavior before it goes live? Legibility is underrated. You should be able to see how the agent will behave, test it, and adjust the scenario yourself. Platforms that hide the logic behind a black box make every change a support ticket.
8. Is every call recorded, transcribed, and traceable? When something goes wrong — and it will, occasionally — you need to replay the exact call, see each step the agent took, and find where it turned the wrong way. Ask whether you can debug a specific session down to the individual tool call. The scenarios guide shows the kind of editable, inspectable logic to look for.
9. Can I test it without a full integration? You should be able to hear the agent on your own content in minutes, not schedule a sales-led proof of concept. A platform confident in its product lets you try a live call for free.
Will it scale and travel?
10. How does it handle languages? If your customers call in more than one language, the agent should detect the language on the first turn and switch automatically. One-language-only is a hard ceiling on who you can serve.
11. How does outbound work, and at what pace? For reminders, confirmations, and lead follow-up you need outbound — ideally both single API-triggered calls and paced CSV batches — with rate control so you do not flood your own numbers. The outbound guide is a useful yardstick for what mature outbound looks like.
12. How am I billed, and does the model reward the right thing? This is the question that surprises people at scale. Per-minute billing rewards long calls; per-seat billing makes you pay for idle coverage. Outcome-based billing — a flat rate per booking, per real conversation, per spam drop — aligns the cost with the value. Ask for the full rate card and model your real call mix against it, not the demo's.
Putting the checklist to work
Notice that most of these questions are about production reality, not demo polish: acting in systems, telling the truth, being inspectable, billing sanely. A platform can ace a scripted demo and fail half of them. The way to avoid that is to run the checklist against a real call, on your real content, before you are committed.
It also helps to test against your specific job to be done, not voice AI in the abstract. If you take bookings, judge it on a booking; if you qualify leads, judge it on a qualification call. The industry pages are a useful lens here — the sales page, for instance, spells out exactly what "qualified lead" should mean and what a good hand-off looks like, which gives you concrete things to check for rather than a vague impression of "it sounded smart." Pick the industry closest to yours and use its described behavior as your acceptance test.
You can do most of it in one sitting. Paste your website on the Aitelier homepage to build an agent in about thirty seconds, place a live in-browser test call, and work down the list — ask it something outside its knowledge (question 4), request a human (question 5), switch languages mid-call (question 10). Then ask to see a session replay and the underlying scenario (questions 7 and 8), and get the full rate card (question 12). If you want to sanity-check the commercial model first, the billing guide spells out exactly what a call is charged for.
The platforms worth committing to are the ones that answer all twelve without hedging — and let you verify the answers yourself before you sign.