If your AI voice agent quotes a caller a price you would never have quoted, you pay it. Not the platform. I read the live terms of service of three AI voice companies on 21 September 2026, and two of them put that in writing. The third does not use the word output anywhere in its contract, which is worse for you, not better.
Every buyer I have worked with compares these platforms the same way: cost per minute, concurrency, how natural the voice sounds on a demo call. Nobody opens the contract. The contract is the only document that describes what happens on the day the agent books a Saturday emergency callout at a weekday rate.
Your vendor has already decided who owns what the agent says. It is not the vendor.
The mistake: you shopped this on the per-minute rate
The question I get is always about price per minute. Five cents, nine cents, twelve. Then somebody asks whether the voice sounds robotic, and the decision is effectively made. I have built phone automation for home services companies, law firms and lenders, and I can count on one hand the clients who opened the terms of service before signing up.
That is the wrong order. The per-minute rate is the smallest number in this purchase. An agent talking to every inbound caller for eight hours a day, unsupervised, naming timelines and prices, is the largest exposure in it. The contract is where that exposure gets assigned, and it is assigned before you place your first call.
So I read three of them on the same morning: Vapi, Retell and Bland. Same product category, three completely different levels of care about the same question.
Three contracts, three levels of care about the same risk
Vapi's terms, last updated 14 September 2026, define Output as the material the service generates on your behalf, naming transcripts, recordings and audio. Then they go one step further than I expected. Where the service performs or implements decisions, recommendations or actions on your behalf, the contract calls those Service Actions, says they are deemed to be performed on your behalf, and leaves all rights and obligations for those actions and their outcomes with you. Booking the job is a Service Action. So is transferring the call, or writing the wrong date into your calendar.
Retell's terms run a section on AI-Generated Content that covers voice outputs, transcripts, summaries and call analytics, and state plainly that this content is produced by automated systems and may contain errors, inaccuracies or unintended outputs. Its warranty disclaimer says the entire risk associated with use of the service is borne solely by the customer. Retell also carries a dedicated telemarketing compliance clause naming the TCPA, the Telemarketing Sales Rule, the Do Not Call registry, the FTC Act, state telemarketing laws and federal and state anti-wiretapping laws. Vapi names the TCPA too, inside its indemnity, alongside the Telemarketing and Consumer Fraud and Abuse Prevention Act.
Bland's terms, last updated 28 August 2026, run about a quarter the length of the other two. Searched the same morning, the words output, voice, generated, recording, TCPA and telemarketing appear zero times on that page. Its user content clause covers what you contribute, post or transmit, which is your side of the conversation, not the agent's.
An AI employee that answers, qualifies and books for you.
Nothing in that sentence tells you who owns the booking when it is wrong.
Actions deemed to be performed on your behalf, with the outcomes yours.
Same feature, named honestly, in the only document that decides anything.
That is a measurement of three published pages on one day, not a prediction of how any of these companies would argue a dispute. Enterprise agreements and acceptable use policies can say more than the self-serve terms do. But the self-serve terms are what most operators actually click through, and on those pages silence is not protection. All three disclaim every warranty, and two of them say in those words that the risk of use is the customer's. What separates them is whether the contract has a word for the thing your customer actually hears.
A tribunal already answered this in 2024
British Columbia's Civil Resolution Tribunal settled the underlying question in February 2024, in Moffatt v. Air Canada. A passenger relied on the airline's website chatbot for its bereavement fare rule, the chatbot was wrong, and, in words CBC News quoted from the decision, the airline argued the bot was "a separate legal entity that is responsible for its own actions."
The tribunal member, Christopher Rivers, called that submission remarkable and refused it. A chatbot is part of your website. A business is responsible for the information on its website. It makes no difference whether that information came off a static page or out of the bot. He also found the airline had not taken reasonable care to make sure the chatbot was accurate.
Now hold that next to your vendor's contract. The vendor tells you the output may contain errors and that the risk of using it is yours. The tribunal tells you that you were required to take reasonable care it was accurate. Two documents written by people with opposite incentives, pointing at the same person. The decision is Canadian and binds nobody in the United States, but no American operator should want to be the case that tests it.
Three moves before you sign anything
Cap what the agent is allowed to assert
Prices, arrival windows, warranty terms, eligibility, anything a caller can act on. My standing rule is that the agent quotes nothing it cannot read from a live system. A line like "our dispatcher confirms the price when she assigns the truck" costs you no bookings and closes the gap the contract leaves open.
Search the vendor's terms for three strings
Output, indemnif, and the name of the telephone law in your market. It takes thirty seconds and it is the only pre-purchase test that tells you something the pricing page never will. If none of the three appear, you have learned that the contract has no language for the event you are most likely to have.
Keep the audio and the transcript, and check they are yours in writing
You will want the recording on the day a caller says your agent promised something. Vapi's Output definition names transcripts and recordings as belonging to you. Confirm your own vendor's does, and confirm your retention window outlasts the complaint window in your trade.
There is one vertical where this compounds. If your calls come from a home services business, where the caller is describing a burst pipe and wants a price and a time in the first thirty seconds, the agent's opening lines are the whole transaction. That is the build I do most often, and it is why I write the script against the contract rather than against the demo. If that is your world, the shape of it is here: AI calling for home services.
The bottom line
Read the contract before the pricing page. Cap what the agent is allowed to assert, keep the recordings, and run the whole build on the assumption that every word the agent says is a word you said. Two of the three vendors wrote that assumption into the contract. A tribunal already ruled it. The only party still treating the AI as a separate third party is the operator who has not read either document.