All insights
Method6 min read

Finance Owns the AI Bill. Nobody Owns the Value.

Vishal Sachar

Co-Founder & CEO of CLRT

If you sit in finance, the AI bill is probably yours now. Deloitte's latest finance trends survey, advisory-firm research covering 1,434 finance leaders at companies with at least US$1 billion in revenue, found 54 percent leading cross-enterprise AI and technology capital allocation and 48 percent managing AI and technology spending and cost controls, and more than two-thirds of those holding these roles took them on within the past three years. The transfer made sense. AI stopped being an experiment run out of an innovation budget and became a recurring line large enough to need a steward. But stewardship of a cost is a different job from stewardship of a return, and the handover moved the first without the second. Finance was given the invoice. Nobody was given a measure of what the invoice buys.

12%1
Organisations that consistently weigh AI value against cost (KPMG Q3 2026 AI pulse, 2,131 leaders)
KPMG, 2026
46.9%2
Enterprises over budget on AI (Futurum 2H 2026 CIO survey)
Futurum Research, 2026
54%3
Share of finance leaders leading cross-enterprise AI and technology capital allocation, in Deloitte's survey of 1,434 finance leaders across 26 countries
Deloitte, 2026

On 24 September KPMG published the clearest picture yet of what that handover left behind. Its third-quarter global pulse of 2,131 senior leaders across 20 countries, research from a Big Four firm that also sells AI services, so read it directionally, found that 61 percent review AI costs during approval and 59 percent monitor them in operation. Then the line falls away. Only 12 percent consistently assess AI value against cost across the organisation. Set those three numbers side by side and the shape is unmistakable. The first two steps are cost controls, the kind finance has always run well. The third is the only step that answers whether the money was well spent, and almost nobody takes it. Companies have built a careful apparatus for watching what AI costs and almost none for asking what it returns.

FIG. 01KPMG's Q3 2026 global AI pulse, 2,131 senior leaders in 20 countries: cost is reviewed and monitored by a clear majority, and value is weighed against it by one in eight. Three separate measures, read directionally.
01The financed overrun

The consequence shows up when the spend runs over, which it now routinely does. Futurum's second-half CIO survey, published on 14 September by an analyst firm that tracks technology buyers, found 46.9 percent of enterprises running over budget on AI, against 5.6 percent below plan. Another 10 percent have no formal AI budget to measure against, and 5.6 percent do not know where their spending stands. What happened next is the revealing part. Among the 767 organisations that ran over, 47.6 percent sought supplemental funding, 43.3 percent absorbed the overrun into the next planning cycle, 38.7 percent reallocated money from elsewhere in IT, and 23.1 percent shifted the spend into a non-IT business-unit budget. About one in six reduced or paused the scope. The overrun was financed, deferred or moved. It was rarely questioned, because nobody held a number that could have questioned it.

FIG. 02Futurum Research's 2H 2026 CIO survey: 46.9% of enterprises ran over their AI budget, and only about one in six of the 767 that did cut scope. Responses overlap.
02Cost is not value

This is not negligence. It is what any finance function does when it holds the cost of something and not its value. Approvals, budgets and variance reports are instruments built to answer one question, whether the money spent matched the money planned, and they answer it precisely. They cannot answer the question an AI overrun actually poses. Usage might have grown because a use case is working and people want more of it, or because nobody is watching and a pilot has quietly become a standing habit. On the ledger the two look identical. A variance report can tell you that you spent more. It cannot tell you whether the extra spend bought anything. Faced with an overrun that cannot be judged, the rational move is to fund it, since cutting a working programme looks worse than paying for an idle one. So the bill rises, and each year's budget inherits last year's unexamined base.

03The missing instrument

The missing instrument is not a dashboard. It is an agreement, made for each use case before the spend, that settles four things. A baseline: what the process costs or produces today, recorded before the AI arrives, because a baseline reconstructed afterwards is a negotiation, not a measurement. An attribution: how the change will be credited to the AI rather than to the new hire, the seasonal swing or the price rise that landed in the same quarter. An owner: a named business leader who answers for the outcome, as distinct from the cost centre that pays the invoice. And a review date, fixed in advance, on which the use case either earns its next round of funding or stops. Deloitte found that even among finance leaders whose organisations approve AI on a measurement-led basis, 23 percent rely on an executive or board mandate without a precise process for measuring results. Measurement in name, mandate in practice.

FIG. 03Illustrative: the four answers a use case needs before the spend if anyone is to own its value rather than only its cost.

KPMG's own data points at the difference this makes. Among organisations reporting established returns from AI, the share that consistently weighs value against cost rises to 48 percent, four times the overall rate. That is a correlation in a survey, not proof that measuring causes returns, and part of it is circular, since a company cannot report established returns without having measured something. But the circularity is the point. The other companies may well have returns too. They simply cannot know, and a return nobody can see cannot be defended in a budget meeting, scaled with confidence, or told apart from the overrun sitting next to it. Assigning AI governance to finance without the value measure has made the CFO accountable for a number whose only visible movement is upward. Governance without an instrument is bookkeeping with a better title.

A variance report can tell you that you spent more. It cannot tell you whether the extra spend bought anything.

A deeper dive

The traps begin the moment a company tries to add value measurement after the fact. The first is the vanished baseline. Once the AI is in place, the earlier state of the process no longer exists to be observed, so the baseline is rebuilt from memory and old reports, and a rebuilt baseline always flatters whoever is doing the rebuilding. The second is the metric found afterwards. A use case measured only once it is running will always find something that moved, because some number always moves, and the measure quietly becomes whatever made the spend look wise. The third is hours that never become money. Most AI business cases are written in time saved, but saved hours turn into value only if they leave the cost base or are redeployed to work that earns, and in most organisations they simply dissolve into the week. Counting them as savings is how a programme reports a return that never reaches the profit and loss account. None of these is dishonest. Each is simply what measurement looks like when it arrives late.

The fourth trap is the one the Futurum data quietly shows. When an overrun is shifted into a business-unit budget, the cost moves but the accountability does not, and the budget now carrying the spend usually belongs to someone who never agreed what it was for. None of this is solved by better tooling, because every vendor meter measures usage, and usage is also the thing that drives the bill. Deciding what counts as value for a given process, capturing the baseline while it still exists, designing an attribution that will survive a sceptical audit committee, and choosing which use cases deserve this rigour at all are judgments about the operation itself. They are also, in practice, the same judgment as where to point AI in the first place. A use case for which nobody can write down the baseline, the owner and the stop rule is not yet a use case. It is a cost waiting to become an overrun.

Work with CLRT

CLRT works on the value side of the AI bill. We help finance and operating leaders decide which use cases deserve funding, record the baseline before the spend begins, design attribution that survives an audit committee, and put a named owner and a review date on every line, so that an overrun can be judged rather than simply financed. If you now hold the AI budget and cannot yet say which part of it is earning its keep, talk to us before the next planning cycle. And if the harder question is where AI should be pointed in the first place, that is what our diagnostic at ascent.clrtstudio.com was built to answer.

Vishal Sachar

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.

Start here

Skip the reading. See where your leverage leaks.

Ascent is our free diagnostic. Ten minutes, and you have the one workflow worth building first.