Questions to ask an AI voice agent development vendor before you sign

Ask an AI voice agent development vendor five things before you sign: how they will prove the agent works, who owns the code and prompts at the end, whose accounts carry per-minute usage, what happens when the agent cannot handle a call, and what leaving looks like. A vendor who has shipped answers all five in specifics, fast.

This is the script we would use if we were the buyer. It is organised around what each answer reveals rather than around politeness, and it includes the answers that should worry you, because the pattern of a weak answer is more informative than the content of a strong one.

The five questions, and what the answers reveal

Work through that table with the person who will run the agent after launch present, not only the person signing. The operational answers matter more to them, and their questions are usually sharper. Notice that none of the five is about the model. Model choice is the easiest thing to change later and the least useful signal of competence. Everything above is about whether the vendor has operated a voice agent after launch, which is where most of the difficulty lives.

Question one: how will you prove it works?

Voice agents cannot be judged by conversation. A demonstration call is a sample of one, chosen by the person who built it, on the intent they handled first. The only credible answer involves a scenario suite: a set of recorded or synthesised calls with known-correct outcomes, run automatically on every prompt change, model change and deployment.

Press for detail. Where do the scenarios come from, and does the vendor need your real call recordings to build them? What is the pass threshold, who agrees it, and what happens to a release when the suite goes red? Does the suite cover the long tail, meaning unclear names, callers changing their mind mid-sentence, dates said as "day after tomorrow", and numbers spoken in one language inside a sentence in another? Our position on this is not subtle: evals over demos is the engineering stance the rest of our process hangs on.

A serious vendor should also describe how it documents and reviews risk. The NIST AI Risk Management Framework sets out measure and manage functions for exactly this, and a vendor who can map its practice to a published framework is telling you the practice existed before your project.

Question two: who owns what at the end?

There are two business models in this market and they are easy to confuse. In one, the vendor builds a system and hands you the code, prompts, tool contracts, evaluation suite, infrastructure definitions and documentation, and you could keep running it if the vendor disappeared. In the other, the vendor hosts a platform and configures it for you, and the configuration is worthless without the platform.

Both can be legitimate. The mistake is buying the second while believing you bought the first. The distinction is what code and IP ownership means in a contract, and it is worth settling before commercial terms rather than after. Ask directly whether the evaluation suite and the prompts are yours, because those are the assets people most often discover they do not own. Who owns the code, prompts and models covers the contract language worth insisting on.

Question three: what will it cost to run?

Build price is the number on the proposal; running cost is the number you live with. Voice agents are metered per minute of live conversation across speech recognition, model tokens, speech synthesis and the telephony leg, so ask for a cost per resolved call at your volume with the assumptions written down, not a cost per minute in isolation.

Then ask who holds the accounts. We are blunt about this: you pay usage through your own provider accounts, and we configure budgets, routing and dashboards so it stays predictable. That keeps the billing relationship, the negotiating position and the visibility with you. For reference, AI voice agent development with us starts at $17,500 or ₹11,20,000 and runs to $56,000 or ₹38,40,000 plus that usage, with care plans from $1,000 or ₹68,000 a month; the pricing page lists the tiers.

Question four: what happens on escalation?

A large share of calls will always need a human. The quality of that hand-off decides whether your support team defends the agent or campaigns against it. Ask what the human receives at the moment of transfer, and accept nothing vaguer than a list.

The answer you want names the transcript, the customer record the agent already retrieved, the reason for the transfer and the agent's best guess at what the caller needs. Ask what triggers escalation: low confidence, a refused action, a second failed clarification, and the caller simply asking for a person, which must always work on the first attempt. Then ask whether escalation quality is measured in the scenario suite, because a hand-off nobody tests is a hand-off that degrades.

Question five: what does leaving look like?

Ask this in the first meeting, cheerfully. The reaction is the signal. A vendor confident in its work will describe the exit in concrete nouns: your repository, your infrastructure definitions, your provider accounts, a runbook, a handover session, and a clear statement that nothing needs re-implementing.

A vendor who deflects, calls the question premature or answers with "your data is exportable" is telling you the system is theirs. Data export is not portability when the prompts, evaluation suite and orchestration stay behind.

Three follow-up questions that separate the field

The five above filter out vendors who have never run a voice agent. These three separate the remainder.

What does multilingual mean in your proposal?

Multilingual is the most elastic word in this market. Ask whether it means the agent detects the language and switches, or whether it handles code-mixed speech where a caller says a Kannada sentence containing an English product name and an Arabic-numeral date. Ask which languages the evaluation suite covers, because a language with no scenarios is a language nobody has tested. Then ask what adding a fourth language costs and how long it takes.

What telephony have you actually integrated?

Ask for named platforms and named failure stories. A vendor who has genuinely shipped will talk about SIP trunk quirks, transfer semantics that differ by carrier, call recording behaviour and what happens to a call leg when the agent process restarts mid-sentence. A vendor who has only used one managed provider will answer at the level of the marketing page.

Who runs this in month six?

Voice agents are not furniture. Call mixes change, models are deprecated, policies get rewritten and somebody has to own the weekly escalation review. Ask whether that is your team, their care plan, or nobody, and make sure the answer is written down before launch rather than discovered afterwards when quality has already drifted.

Documents to ask for before you sign

• A dated delivery plan with shadow mode inside the schedule, not appended to it
• The evaluation approach in writing, including where scenarios come from and the pass threshold
• A running cost model at your call volume, with assumptions and a sensitivity range
• The ownership clause, naming code, prompts, evaluation suite, infrastructure and documentation
• A security questionnaire response, covering sub-processors, recording retention and data residency
• Two reference calls with clients running voice agents in production, not in pilot
• The change-request price list for a new intent, a new language and a policy change after launch

Our own security questionnaire for AI vendors is a reasonable starting point for the fifth item, and you are welcome to send it to us as well as to everyone else.

Where this checklist is the wrong tool

If you have not yet chosen the intent, this is premature. A vendor cannot quote sensibly against "we want a voice agent", and the answers you get will be generic because the question was. Run a discovery exercise first, pick two intents from the call log by volume, and then hold this conversation with real scope on the table.

The checklist is also overkill for a genuine experiment. If you are spending a small fixed amount to learn whether the technology fits an intent at all, ownership and exit terms matter far less than speed. Buy the experiment, learn, and apply the full script when you commit to production.

Related reading

The multilingual voice agent case study shows what shipped answers to these five questions look like in practice, and five ways AI voice agent development projects fail explains what each weak answer tends to cause. If you want to run this script against us, start a conversation and ask the exit question first.

The vendor worth signing is the one whose answers get more specific the harder you push, not the one whose slides get shinier.