Automated customer service is cheap to start and not free to run. Gartner's forecast that the cost of a generative-AI resolution will exceed $3 by 2030 makes that explicit, and it moves the real question from the price of a unit to the price of a result, and to who carries the risk when the result does not come.
We have argued for a decade that the cost of a single success matters more than the cost of a minute. The forecast does not prove us right about anything, but it does push the whole market towards the same arithmetic. This article explains what the forecast says, what it does not, and how to structure a contract so that risk sits with whoever can control it.
What Gartner predicted#
In January 2026 Gartner published a prediction that by 2030 the cost per resolution of generative AI in customer service will exceed $3, which the press release frames as higher than the cost of many B2C offshore human agents. It also predicted that by 2028 AI-related regulatory changes will increase assisted-service volume, where a person is involved, by 30%, as customers exercise a right to talk to a human.
The useful part is the unit. Gartner does not talk about cost per token, per minute or per conversation. It talks about cost per resolution. That is the right unit, and it is the one most buyers still do not have in their contracts.
Why running costs rise#
There are several reasons why automation that looked almost free in a pilot becomes a material line in the budget at scale. None of them is exotic.
- Long conversations cost more. A generative system that talks at length, repeats context or loops through clarifications consumes compute and telephony for every extra turn.
- Failures cost twice. A conversation that does not resolve the case still costs money, and then the case is handled again by a person or by another attempt.
- Oversight costs money. Listening, testing, versioning content, audit trails and incident handling are part of running a process in a regulated sector. They do not appear in a demo.
- Integration and change cost money. Every new portfolio, product or country brings data exchange, rules and dialogue work.
- Regulation adds human steps. Where a person must be offered or involved, the cost of that route belongs in the cost per resolution.
Other research points the same way. Gartner's August 2026 survey of 199 service leaders found AI spending up 38% while overall service budgets rose 2%. Forrester predicts that rising AI costs will force B2C leaders to scale back about a third of AI initiatives in marketing, CX and digital in 2027; that prediction covers B2C marketing and digital, not contact centres alone. The common thread is that the bill arrives after the pilot.
Unit price and price per success#
The mistake we see most often in procurement is comparing suppliers on unit price. A minute, a call or a conversation is an input. What the business buys is an output: a payment, an attended visit, a qualified lead, a completed survey.
The arithmetic is simple. If one solution costs half as much per call but succeeds five times less often, it costs two and a half times more per success. In 2022 our founder made this point in an interview with NowyMarketing: a cheap solution with a low success rate can cost more than twice as much per success as a dearer one that works. The numbers change by process; the logic does not.
If we are paid by the minute, we are not playing towards the same goal as our client. – Dawid Wójcicki, Founder & CEO of Vee
Per-minute billing makes the problem worse, not better. It rewards long calls, while in many processes the result comes from short ones. A promise to pay in early collections, for example, is often obtained in a short, specific conversation. A supplier paid by the minute has no reason to make that conversation shorter.
Who carries the risk in each model#
Every pricing model allocates two risks: the risk that running costs rise, and the risk that the result does not come. It helps to name them explicitly.
| Model | Risk of rising running cost | Risk of no result |
|---|---|---|
| Licence or platform, client runs the process | Client | Client |
| Per minute or per token | Client | Client |
| Per conversation or per call | Shared, mostly client | Client |
| Per case or account, independent of outcome | Supplier | Client |
| Base fee plus a performance layer tied to verified results | Supplier | Shared |
Simplified; actual allocation depends on contract terms, volumes and how results are measured.
The direction of travel is clear. Our own billing has moved away from minutes: new agreements are no longer billed by the minute, though some existing contracts still are. The model we use has three layers. A one-off setup. A base operation fee per case or per account, independent of the outcome, which moves the running-cost risk away from the client on terms set out in the agreement. And an optional performance layer, a KPI incentive or a fee for a verified result, where the result can be measured in the client's own systems.
We do not claim that outcome-linked pricing fits every process, and we do not offer to work for results alone. Some processes have no clean, attributable outcome; some clients prefer predictability. The point is to choose consciously who carries which risk, rather than letting the unit of billing decide by default.
Defining a resolution before you pay for one#
Outcome-linked pricing only works when the outcome is defined precisely enough that neither side can argue about it later. Every success definition we agree with a client answers six questions.
- What counts as success: a payment, a visit attended, a contract signed, a survey completed, not a declaration.
- When it must happen: the measurement window after contact.
- Where it is measured: the source of truth, normally the client's system, not the supplier's dashboard.
- Attribution: how the effect of contact is separated from what would have happened anyway, usually against a comparator on the same buckets and portfolios.
- Exclusions: which cases do not count, such as those transferred to the client's staff or stopped for legal reasons.
- Disputes: how disagreements about data are settled, and by whom.
The fourth question is the one that separates honest outcome pricing from marketing. A supplier that claims the whole recovery rate is claiming payments that would have arrived without a call. We claim the increment. In early collections we have measured an increase of 7–13% in on-time payment frequency, historically, versus a human-agent comparator. That is the kind of figure that belongs in a performance layer: measured, scoped and incremental.
What a resolution looks like in practice#
“Resolution” sounds universal, but it means something different in every process, and the risk profile changes with it.
Early collections#
The resolution is a payment within a window, not a promise. For a bank the expensive part of delinquency is rarely the cost of the call; it is the delay that pushes an account into the next bucket and forces it to build reserves. The value of fast, complete contact is therefore measured in money that arrives earlier, and the comparator matters more than anywhere else.
Appointments in healthcare#
The resolution is often learning early that a visit will not happen, so that the slot can be offered to someone else. A cancellation obtained two days ahead is a success; a confirmation that turns into a no-show is not. Paying for confirmed calls would reward the wrong thing.
Lead qualification#
The resolution is an interested customer handed to a salesperson who has time to take the call. Whether that customer is eventually granted credit depends on the lender's own criteria, which the contact supplier does not control. That is a clear case for sharing risk only on the part of the funnel the supplier can influence.
How architecture affects the cost of a result#
Cost per resolution is also a design choice. In regulated processes we run deterministic dialogue where it matters: replies come from a closed, versioned library, and models are used to recognise what the customer means rather than to write what the system says. That keeps conversations short, behaviour predictable and the run cost stable. Generative components are used where they help and where the risk is low.
We also own the telephony layer, the dialler logic and the retry strategy. Reach is the first rung of any result, and a large share of the cost of a resolution is spent before anyone answers. How and when customers are reached changes the cost per result more than the choice of model.
What to ask a supplier#
- What is the unit of your price, and which risks does it leave with us?
- If conversations get longer or need more attempts, who pays?
- How do you define a resolution in our process, and where is it measured?
- Against which comparator will you measure the effect, on which buckets or segments?
- Which parts of the process use generative replies, and which use approved, versioned content?
- What happens to the price if regulation requires more human involvement?
The $3 forecast will be argued about, and it may turn out high or low. The more durable lesson is that automation has a running cost and a failure cost, and both need an owner. Buy results where they can be measured, pay for operation where they cannot, and make sure the supplier carries the risks it controls.
Follow The First Second
Occasional issues on the mechanics and economics of AI-powered customer contact. No noise, no vendor news.
Follow via RSS