BusinessGPSOwn the capacity to change.

Long read

In Capgemini’s survey, only 56% are fully confident they can forecast how AI costs will change as adoption scales. Name who owns the cost after go-live

The cost of an AI system is not settled at go-live. It moves as more people use it. In Capgemini’s survey, only 56% are fully confident they can forecast how AI costs will change as adoption scales. So name one person who owns the running cost after go-live. Then agree the evidence they will bring before the next tranche is funded.

On 12 November 2026 the Capgemini Research Institute published a survey of 500 senior executives at organisations with annual revenue above $1 billion. It was run in September 2026. Capgemini sells AI and consulting services, so it has an interest in these findings. The evidence is a survey of AI spending in general, as executives report it. It is not a measured go-live outcome, and no named organisation’s result is in it. Its direction is clear enough to act on.

The two-minute version

The scene. A programme takes an AI system live. The business case holds a cost estimate. The programme team moves on, and the estimate stays in a document nobody opens again.

In this issue. Capgemini’s survey of 500 executives, published on 12 November 2026, puts numbers on 2 things. How well organisations forecast the cost of AI, and what stands in the way of proving its value.

What changed against the belief that the cost is settled at go-live:

Bottom line: give the running cost a named owner after go-live, and agree now what evidence will fund the next tranche.

Only 56% are fully confident they can forecast AI costs as use grows

Capgemini’s release of 12 November 2026 reports these findings from the 500:

AI Yield is Capgemini’s own metric. The release defines it as “a metric that captures the verified value of an AI-assisted outcome after accounting for its full safe-delivery cost”. In plain words: what one result from the AI is worth once checked, less everything it took to deliver safely.

The release also reports 2 findings that frame the rest. Only a third of executives can confidently estimate indirect AI costs, “including human oversight, rework and governance”. And half say AI spending is approved without adequate visibility into the full cost. In the past 12 months, 59% say their organisation’s AI-related expenditure was significantly higher than expected or budgeted.

Read together, these describe one gap from several sides. Approval often happens without adequate visibility into the full cost. The cost changes as use grows, and many are not fully confident they can forecast how. And few organisations turn it into a value per outcome that would win the next round of funding.

After go-live, 3 things move the cost, and no one role carries all of it

This section is our reasoning, not Capgemini’s finding. The survey covers AI spending in general. What follows is what we think that means in the months after a go-live.

First, use grows. A tool that starts with one team spreads to 3. Volume rises. If the supplier charges per use, the invoice rises with it. So do the hours of the people who check the work.

Second, the checking changes. In the first weeks, a person usually checks everything the tool produces. Later, someone may decide a sample is enough. That decision can cut the checking hours sharply, as the example below shows. It is also a decision about risk.

Third, the people change. The programme team that built the business case often disbands soon after go-live. The business owner inherits the tool, but rarely the budget line for the hours. Finance sees the supplier’s invoice. The checkers’ manager sees the overtime.

So the running cost falls between roles. Each person sees one part. No one is asked to report the whole.

An invented example: the same tool at month 1, month 3 and month 6

Every figure here is invented, to show how the cost moves. Use your own.

A tool proposes answers to customer queries, and a person checks each one at 3 minutes.

MonthItemsCheckedChecking hoursPeople cost at an invented £40 an hour
11,000Every item50£2,000
33,000Every item150£6,000
66,0001 in 5, a sample60£2,400

From month 1 to month 3, use triples and so do the hours. By month 6, use has doubled. The hours fall, because someone decided to check a sample.

At month 6, checking every item would take 300 hours. Checking a sample takes 60. That gap of 240 hours is one decision. It needs evidence that the tool is accurate enough to stop checking everything. Whoever makes that call owns the running cost, whether or not anyone named them.

Name one person who owns the running cost after go-live

Choose the person who will keep the tool after the programme team leaves. Usually that is the business owner of the process it supports. Pair them with a finance partner who will read the figures each month.

Write down what they own: the supplier’s charge, the checking hours and any review time, as one monthly figure. Ask them to accept it in writing. Check that the budget line that pays for the checking hours moves to them, or to a named budget holder. If it does not, tell the sponsor before the owner accepts.

If no one will take it, say so to the sponsor before go-live. An unowned cost still gets paid. It is paid from budgets that never agreed to it.

If you are already live, the move is the same. Name the owner now, and start their first report from this month’s figures.

Done when: one named person has accepted the running cost in writing, with a finance partner beside them.

Set the 5 figures the owner reports each month

Agree a short monthly report. Keep it to 5 figures. They are volume, the supplier’s charge, the checking and correcting hours, the items corrected, and the total cost with the hours priced.

Set the first report for the end of the first full month after go-live. Put the business case’s estimate beside each figure, so the gap is visible from the start.

Where the business case holds no estimate for a figure, write “none” beside it. That gap is a finding in itself. If you built the one-tool sheet from this issue’s page, start the first report from it.

Done when: the 5 figures are agreed, the first report date is set, and the business case’s estimate sits beside each one.

Agree what evidence funds the next tranche, before the first one ends

Before the next tranche goes for approval, agree what evidence it needs. Agree it with the sponsor and the owner now, not at the approval meeting.

At a minimum, the list holds 4 items:

Capgemini’s AI Yield is one way to frame the value. Your own measure will do, if it counts the people as well as the invoice.

Write the evidence list down, with the date of the tranche decision. Then the owner knows from month 1 what they will be asked for.

This is where the 57% bites. A tranche that asks for proof at the approval meeting leaves little time to gather it. A tranche that agreed the proof months earlier gets the figures it asked for.

Done when: a written evidence list is agreed by the sponsor and the owner, with the date of the next funding decision.

Decide who may reduce the checking, and on what proof

Write one rule. It says who may move the tool from checking every item to checking a sample. It also says what proof they need, such as an error rate below an agreed level over a set number of weeks.

This is the decision that moves the cost most. It should be made on purpose, by a named person, with the evidence on the record.

It also works in reverse. If corrections rise after the change, the same rule says who moves the tool back to full checking, and how fast.

Done when: the rule names who decides, the proof required and where the decision will be recorded.

What comes next

BusinessGPS’s Baseline report grades how an organisation runs change, step by step, in red, amber or green. To see how the grading reads, open the BusinessGPS Baseline report, which BusinessGPS ran on itself.

Sources

  1. Capgemini Research Institute, press release on the economics of AI, Paris, 12 November 2026, read on 15 November 2026. capgemini.com

Every edition