The Business Case for Voice AI: Cost Per Outcome, Not Cost Per Minute

A founder signed the cheapest per-minute voice AI he could find, and three months later the calls were cheap and the results were flat. Part 3 of the Voice AI Playbook: the Minutes Mirage, how to model cost per resolved outcome, and how to decide build versus buy.

Naveen forwarded the quote at 11pm with one question: “is six rupees a minute good?” By the time we answered, he had already signed a different vendor, the cheapest per minute he could find. Three months later the calls were cheap and the collections were flat. He had optimized the one number that did not matter and left the one that did untouched.

He never cared what a minute cost. He cared what a recovered payment cost. Nobody had shown him those were different questions.

Call the trap the Minutes Mirage: judging a voice AI investment by its per-minute rate instead of its Cost Per Resolved Outcome. A platform that is half the price per minute and resolves a third as often is not a deal. It is a more efficient way to waste money. This is part three of the Voice AI Playbook, and it is about modeling the money honestly, then deciding whether to build or buy.

Why does per-minute pricing mislead?

Per-minute pricing is seductive because it is simple and it is the number on the quote. It is also the number least connected to value. Two platforms at the same rate can differ enormously in connect rate, in how often a call reaches a useful outcome, and in how much human cleanup each call leaves behind. A cheaper minute that produces a longer, more confused call can cost more per outcome than a pricier minute that resolves on the first attempt.

The minute is an input cost. The outcome is the unit of value. Optimizing the input while ignoring the unit is the oldest mistake in operations, and per-minute is exactly the metric where the cheapest vendor wins and the buyer loses.

How do you compute Cost Per Resolved Outcome?

Cost Per Resolved Outcome is the all-in cost of a workflow divided by the outcomes it actually produced. It is arithmetic, not magic, and you can estimate it before a pilot and confirm it after. The all-in cost includes more than minutes:

• Platform and per-minute or per-call charges for every dial, connected or not.
• Telephony and number costs, which are easy to forget and add up.
• Setup and ongoing tuning time, amortized across the volume.
• Human cleanup: the minutes a person spends fixing or re-doing what the agent got wrong.

Then divide by resolved outcomes, not by calls placed. If a lending workflow dials ten thousand borrowers, connects with six thousand, and lands two thousand promise-to-pay commitments, your denominator is two thousand. The same logic applies everywhere: divide by booked visits, confirmed appointments, passed screens, or confirmed deliveries. Volume without outcomes is just cheaper noise.

A worked example

Take Naveen’s reminder workflow. Suppose the all-in cost of ten thousand dials lands near fifty thousand rupees once telephony and tuning are included. Two thousand promise-to-pay commitments puts the cost per resolved outcome around twenty-five rupees. Now compare that to a human caller who, between dialing, waiting, and notes, completes perhaps forty to sixty meaningful conversations a day. The honest comparison is not rupees per minute against rupees per minute. It is twenty-five rupees per commitment against the fully loaded cost of a human producing the same commitment. That is the comparison that survives a CFO’s questions.

Where does the value actually show up?

Cost per outcome is the floor of the business case. The ceiling is labor and revenue, and the numbers there are large. Gartner projects that conversational AI will reduce contact-center agent labor costs by 80 billion dollars by 2026, and notes that agent labor can run as high as 95 percent of what a contact center spends. The lever is not a cheaper minute. It is moving repetitive, high-volume conversations off a payroll line and onto a workflow.

The revenue side is just as workflow-specific. Calling a real estate lead in minutes instead of hours lifts the share that convert to site visits. Cutting clinic no-shows by a few points recovers booked revenue that was already sold. An on-time lending reminder turns a would-be collections case into a routine payment. A good business case names the outcome, prices it, and then shows what a higher resolution rate is worth at your volume. Cost per outcome tells you the call is affordable. The revenue lift tells you it is worth doing at all.

Build or buy: which is honestly cheaper?

Once the math works, the next question is whether to assemble it yourself from speech, language, and telephony components or to buy a platform. Building looks cheaper on a spreadsheet that counts only the model API. It stops looking cheap the moment you price the parts that are not the model: telephony reliability, retry logic, call quality across networks, CRM write-back, escalation, language coverage, monitoring, and the engineer who owns all of it at 9pm when calls start failing.

A reasonable rule: build when the conversation itself is your product and your differentiation, and buy when the conversation is plumbing that needs to work reliably and write back to systems you already run. For most real estate, healthcare, lending, recruiting, and logistics teams, voice is plumbing. The differentiation is in the workflow and the data, not in re-implementing speech infrastructure a platform has already hardened.

What changes after a quarter

Teams that adopt cost per outcome as their unit stop having vague debates about whether voice AI is “worth it.” They can say what it costs to confirm a patient, qualify a lead, screen a candidate, or secure a payment commitment, and they can watch that number move as they tune. The conversation with finance changes too. Voice AI stops being an experiment with uncertain payback and becomes a controllable line item with a known cost per outcome and a known revenue lift behind it.

The deeper bet: pricing moves to outcomes

If Naveen had asked “what will a recovered payment cost me?” instead of “is six rupees a minute good?”, he would have signed a different contract and kept three months of collections. The deeper shift is that voice AI pricing will eventually move from technical units to business outcomes, because that is the only unit buyers actually value.

Until vendors price that way, the operators who win are the ones who do the translation themselves. They refuse to be sold a minute and insist on pricing a result. Get that habit early and every later decision, which platform, which workflow, when to scale, gets easier, because you are always optimizing the number that pays the bills.