Support AI ROI narrows at both ends

The quarterly budget meeting is half over when the finance lead turns the vendor deck back to the first page. The number on it says seventy-six percent of tickets resolve without a human, at ninety-nine cents each, and the monthly bill would be half of what the support organisation spends today. After the meeting, someone re-runs the arithmetic on last month real tickets and lands on a net saving of a little over twenty percent.

Both numbers are honest. They differ in what they count. The vendor counts what its agent can absorb; the buyer has to count how much of the whole ticket inventory can be handed over at all. What makes this year unusual is that both ends of that calculation are moving at once: the unit price is climbing, and the share of automatable work is a ceiling fixed by the business. Read only one end and the budget comes out optimistic.

The unit price is climbing

Gartner put a direction on this in January: by 2030 the cost per resolution for generative AI will exceed three dollars, higher than many consumer-facing offshore human agents. The reasons listed are rising data centre costs, a pivot from subsidised growth to profitability among AI vendors, and use cases that keep getting more complex. Complex tickets consume more tokens and need more expensive people behind them.

The same forecast carries a line that gets quoted less: by 2028, regulatory changes related to AI will increase assisted service volume by thirty percent. That matters more than it sounds for a budget model. It says the human workload pushed out by AI does not disappear. It moves: part of it becomes reviewing the model output, sampling its answers, and handling appeals.

Put those two together and the ninety-nine cent price has a character. It is a subsidy price, set while a vendor buys market share with losses. A contract that locks three years of volume at today unit price hands the buyer a saving on paper and keeps the upside risk at home.

What procurement can actually do is concrete. Write the unit price into the contract with a cap on the annual increase. When pricing is per outcome, write down who defines the outcome. Settle who keeps the evaluation set and the call records on exit. None of the three solves a technical problem, and all three decide the shape of the bill three years out.

Pricing models put the risk on different people

The same support agent can be sold three ways, and each puts the risk somewhere else. Per seat, the vendor wants more seats, so its incentive is coverage rather than resolution, and the bill stops tracking ticket volume and starts tracking headcount. Per conversation, the buyer pays for the conversations that were not resolved, so the risk sits with the buyer. Per outcome, the buyer pays only when the agent resolved, which is the friendliest shape, and it pushes the risk into a different place: the definition.

What counts as one outcome, who decides, and whether the decision can be checked. Widen the definition by one notch and the bill grows by a slice. This is already being treated as a method problem in the field: counting the share of conversations where the customer did not ask for a human comes out twenty to forty points higher than counting the ones where the problem was actually solved. That gap is not a technology gap. It is a definition gap.

So the thing to negotiate in an outcome contract is not the price. It is the decision rule, and the more it reads like a test case the less it argues at settlement time. Who writes it, who reviews it, what share gets sampled by a human, and how a wrong decision is traced back. Those four belong in the contract, not in the appendix.

The other end is ticket structure

How much of the ticket inventory can go to an agent is decided by the structure of the tickets, not by how strong the model is. The same vendor reports very different numbers across industries. In ecommerce, seventy to eighty-four percent of tickets fall into predictable categories and cost drops by forty-seven percent. In SaaS and technology, the average is thirty-nine percent, with technical escalations holding resolution between fifty-five and sixty-five percent. Financial services resolve less but save more per ticket, because a human takes longer on each one. These figures come from the Fin industry benchmark page, which is a first-party source, so read it knowing it is both the data and the vendor.

Together the numbers say something simple: net cost reduction across a whole support organisation lands at twenty to thirty-five percent in year one, not the seventy percent in the deck. The difference is infrastructure, configuration time, and the tail that nobody avoids, where complex tickets end up back with people.

The human reference price is thirteen dollars and fifty cents per contact in the Gartner figure, with other benchmarks between six and twelve. That range matters because it is both the comparable for AI and the buyer own coordinate. A company that costs its human contact at six dollars has a third of the saving room that a thirteen-dollar company has, and that company is usually the one most eager to buy, because its margin per ticket is thin.

What the arithmetic looks like with both ends moving

Take one set of explicit assumptions, which are assumptions rather than measurements, and swap in your own numbers: fifty thousand conversations a month, eight dollars per human contact, ninety-nine cents per AI resolution.

At forty percent automatable, twenty thousand conversations move, human cost is one hundred and sixty thousand dollars, AI cost is nineteen thousand eight hundred, and the month saves one hundred and forty thousand two hundred. Put the 2030 price of three dollars in place, and the AI cost rises to sixty thousand: the saving becomes one hundred thousand, down by forty thousand two hundred, or twenty-nine percent.

At sixty percent automatable, thirty thousand conversations, the saving is two hundred and ten thousand three hundred at ninety-nine cents; at three dollars it is one hundred and fifty thousand, a drop of roughly thirty percent. At twenty percent, the saving falls from seventy thousand one hundred to fifty thousand.

Set the three cases side by side and one thing is visible. A three-fold increase in unit price costs about thirty percent of the saving. Moving the automatable share from twenty to sixty percent multiplies the saving roughly three times. Of the two variables, only the second one is yours to move. The unit price follows the vendor cost structure and its funding cycle. The share follows your own ticket taxonomy and your own process work.

There is also a denominator that gets left out. The thirty percent increase in assisted service volume means the human side of the work grows because of AI: reviewing what it produced, handling what it got wrong, and meeting the sampling and appeal requirements that come with regulation. A model that counts only the conversations handed over puts this at zero, and it returns later as training and quality review.

Consequences you can check

If this holds, several things should be observable. First, swapping in a stronger model leaves the automatable share where it was. Second, companies that did process work move the share, and the movement shows up in specific categories: approvals moved earlier, data sources unified, decidable conditions written down, and a slice of the tickets that used to sit in the judgment box moves into the decidable box. Third, price increases show up first in the most complex categories, where token consumption and human cost both rise.

The reverse also holds. When a company reports a large saving increase right after a model upgrade, the likely explanation is that its ticket taxonomy was too coarse and a sampling pass redrew the boxes, not that the model got stronger. This test helps a buyer judge whether a vendor case study is real, because model upgrades have a ceiling and process work does not.

Our judgment

We do not accept the default assumption that return on support AI is a curve that improves with time. What 2026 shows is two lines crossing: the unit price climbing, and the automatable share fixed by the shape of the business. The part that actually moves is in the middle, in how you sort tickets, where you draw the human boundary, and how much process expense you are willing to pay up front.

That is why the first action we take with a client is not choosing a model. It is counting tickets. Sort last month into three boxes: answers that live in the knowledge base are decidable, answers that require checking another system are judgment, and answers that require authorisation or payout are authority. Only the first box is natural ground for an agent. The second needs tooling and process. The third mostly stays with people. Measure the three shares and the ceiling appears, and it barely moves when the model changes.

The second judgment is about the contract. Per-outcome pricing sounds friendly to a buyer, but once the unit price is in an upward channel it moves the risk into the definition. A wider definition means a larger bill, and no technical work fixes a definition. Only the contract does.

The third is about the organisation. The more tickets the agent absorbs, the harder the remaining ones are, and the job shifts from throughput to judgment. That layer does not get cheaper when AI goes live. It comes back as training, review and quality checks. Counting it before the fact is cheaper than finding it at year end.

Where this arithmetic does not apply

The boundaries belong in the piece. A company handling a few thousand tickets a month cannot estimate shares this way, because the sample is too small and a two-hundred-ticket check will skew the ratio. That company should run a single-category pilot and decide afterwards whether to widen.

A company with a stable outsourcing contract priced per head has a comparable of three or four dollars, not thirteen fifty. In that setting AI saves far less than the model predicts, and the real return shows up in response time and language coverage rather than cost. Judging it with a cost model misreads the case.

Regulated work such as insurance claims and medical advice is constrained by something else as well. The automatable share there is set not only by ticket structure but by who is allowed to make the decision. Sort by authority level, not by subject matter.

Three actions you can take today

First, compute the three-box shares from last month own tickets instead of the vendor industry average. It is one day of work: use the labels you already have and hand-check two hundred tickets to correct the bias.

Second, build a sensitivity table at ninety-nine cents, one fifty, and three dollars. If the three-dollar case does not cover the up-front spend, payback runs past three years and the contract term should follow.

Third, write three lines into the contract: unit price with a cap on annual increase, who defines an outcome and what share is sampled, and who keeps the evaluation set and call records on exit.

A question for whoever signs: when was your ticket taxonomy last updated. If the categories are three years old, the ceiling is a three-year-old ceiling, and no model upgrade will lift it.

References

External figures come from the two pages below, with vendor claims marked as such. The unit prices and volumes in the worked example are assumptions set here, not measurements from either source.

Gartner, January 26 2026 press release: GenAI cost per resolution above three dollars by 2030, higher than many offshore human agents, with a thirty percent rise in assisted service volume by 2028

Fin (Intercom) industry benchmark page: six to thirteen fifty per human contact, fifty cents to two dollars per AI ticket, per-industry resolution rates, twenty to thirty-five percent year-one net saving (first-party vendor source)