AI voice agent

The four parts you are actually buying

• Hearing. Speech turned into text, per minute. Accuracy is quoted on clean recordings and measured on your callers, who are in vans and kitchens. This is where a quote and reality separate first.
• Deciding. The model, per call and per token. The cheapest part of the bill and the part every demo is about. If a vendor spends the whole meeting here, they are showing you the easy quarter.
• Speaking. Text turned back into voice, per character or per second, and the premium voices cost several times the plain ones. Worth hearing yours read your own product names before you sign anything.
• The line. The number, the minutes, the carrier. Boring, cheap, and the part that decides whether you can move to another supplier next year or start again.

Where the money goes, and it is not the build

Every layer above bills by use, so a voice agent costs by the minute rather than by the month. That is the whole economic argument: the bill grows with how much people talk to you, not with headcount, which is good news at four hundred calls a day and bad news at forty. The platform in the middle takes a margin on all four parts, and its price per minute usually hides which of them moved.

The number to bring to any conversation is your own: minutes of calls last month, split by the ones a machine could plausibly take. Multiply it out before the pilot rather than after. I would rather tell you in the free hour that the arithmetic does not work than build you something that quietly costs more than the person it replaced.

Buy the platform, or build on the parts

• Buy, almost always, to start. The voice platforms have solved the hard, unglamorous half — telephony, interruption handling, keeping the pause short enough to sound human. That is real engineering you should not pay me to repeat.
• Build the part that is yours. What it may never say, which system it reads, what happens at the handover. No platform writes those, because they are decisions about your business rather than software.
• Keep the line and the records portable. The number and the transcripts should be somewhere you own. That single condition is what stops a per-minute price rising because they know moving is expensive.
• Rebuild only when the boundary is odd. A rule no platform can express, or a system only you can reach. That is the case for building, and it is rarer than the people who sell builds will tell you.

Five tests you can run in ten minutes, before you sign

Call the vendor's demo number yourself and do these. Every one of them is a thing real callers do on the first day, and every one of them is cheap to check now and expensive to discover later.

• Interrupt it mid-sentence with something unrelated. It should stop, listen, and carry on from where you took it — not finish its paragraph.
• Spell a surname with two double letters, then ask it to read it back. Anything it writes into your systems has to survive being said out loud.
• Ask it whether it is a person. The answer should be immediate and plain, and if the vendor treats that as a weakness, that tells you what they optimise for.
• Ask for a human. Time how long it takes and check whether the person who picks up already knows what you said.
• Call from a speakerphone in a noisy room. The accuracy you were quoted was not measured here, and this is where most of your callers actually are.

When a voice is the wrong shape for the job

• The call is really a form. If the caller always says the same six things in the same order, a link by text costs nothing per minute and is faster for both of you.
• The answer lives somewhere that cannot be read live. A voice that has to say "let me check and call you back" is a slower version of the person you already have.
• Your bill would be smaller than the person it replaces. Per minute has no economy of scale to offer at low volume — work out both numbers before anyone builds anything.
• Nobody has written down what it must never say. That list is the deliverable, and until it exists there is nothing to build and nothing to test against.