A new executive survey highlights how difficult it can be to connect AI spending with verified business value. The practical lesson is to follow a workflow from first request to accepted result, accounting for the people, corrections and infrastructure needed along the way
An AI subscription or usage invoice is easy to recognise as a cost. The time a colleague spends checking its output may be less visible, especially when that effort is scattered across a working day. Yet both belong to the economics of the same task. A 12 November report from Capgemini Research Institute gives this problem a timely business context. In a September survey of 500 senior executives, 59 per cent said their organisation’s AI expenditure had been significantly higher than expected or budgeted. Only a third were confident estimating indirect costs such as oversight, rework and governance.
The survey covers organisations with annual revenues above US$1 billion, across ten countries and thirteen industries. It reflects executives’ reported experience, rather than independently audited accounts, and it cannot establish how common the same problem is among small firms. Its relevance extends beyond that population through a management question: what does a useful AI-assisted result actually cost? This article develops that question as an operating discipline. The objective is not simply to make the technology cheaper. It is to understand whether the combination of software, data, people and process improves the work the business needs to complete.
Choose a business outcome before choosing a cost metric
A company can count requests sent to a model without knowing whether those requests accomplished anything. An illustrative support operation might generate many draft replies while still needing employees to resolve the underlying inquiries. In that situation, a falling cost per draft would not necessarily mean a falling cost per resolved case. The distinction is familiar in other areas of management: an activity is not always a result. Before comparing AI options, the business should define what completion means, including the quality threshold and any required approval. That definition determines which costs and benefits are relevant to the decision.
The FinOps Foundation’s guidance on unit economics connects technology spending with the value of products, services or activities, distinguishing technical resource measures from business measures. Applied to an AI workflow, that suggests looking beyond consumption towards a unit the organisation recognises, such as an accepted document or a resolved inquiry. This is an application of the principle, not a claim that every outcome can be reduced to a single number. A cost measure should be accompanied by evidence of quality and timeliness, so that a cheaper process is not rewarded for leaving difficult work unfinished.
Follow the task through its corrections
Consider a hypothetical small consultancy using AI to prepare the first version of a client report. The model’s output might be quick and inexpensive, yet the final report could require source checking, factual corrections and substantial rewriting. Those activities do not automatically invalidate the tool. They establish the boundary of its contribution. A useful comparison would examine the complete process with and without assistance, using tasks of comparable difficulty. It would include the effort needed to make the work acceptable rather than stop the clock when the first draft appears. Otherwise, the apparent saving could be created by measuring different stages of completion.
This does not require treating every minute of human judgement as waste. Review may be a necessary part of delivering a dependable service, and an AI-assisted process could improve the quality of what staff review. The relevant question is whether the overall arrangement produces better results for the resources used. An illustrative evaluation might distinguish routine edits from substantive corrections and record when a task must be restarted. That evidence could help identify where the tool is useful and where it creates disproportionate work. It would be more informative than asking staff only whether they felt faster while using it.
Released capacity is not automatically a cash saving
Time saved can have real value without immediately reducing expenditure. A professional who finishes preparation earlier might spend more time with customers, respond sooner or absorb additional demand. Those are possible benefits that should be described in their own terms. If the wage bill remains unchanged, the business should be careful about calling the same benefit a realised cost reduction. The distinction matters for investment decisions because a proposal can look more profitable when potential capacity, cash savings and revenue are combined without checking whether they overlap. A credible business case explains what changes in practice and how that change will be observed.
The same care applies to revenue claims. More quickly generated sales material does not by itself demonstrate additional profitable sales. An organisation could investigate whether assistance improves response times, the suitability of proposals or conversion, while recognising that market conditions and other changes also affect results. These are suggested evaluation questions, not outcomes reported by the survey. The goal is a defensible chain from the intervention to the benefit. A firm may decide that a service improvement justifies the expense even when direct financial attribution remains difficult, but it should make that judgement explicitly rather than manufacture a precise return.
A cheaper model can make a workflow more expensive
Choosing a less expensive model is an obvious response to a growing bill. It can be sensible when the task’s requirements are met, but the comparison should preserve the same definition of acceptable output. An illustrative model that costs less per request could require more retries or more specialist review. Another might be more expensive to call but complete the relevant task with fewer corrections. The appropriate comparison is therefore conditional on the workload, the prompts and the surrounding process. There is no universal inference that a larger model creates more value or that a smaller model is always the economical choice.
Capgemini reports that more than three in five respondents expect their organisations to adopt a multi-model strategy over the next year. That is an intention, not evidence that routing work among models has already delivered a return. For a business considering the approach, routing also introduces work: deciding which tasks go where, maintaining integrations and checking that quality remains acceptable after updates. The potential saving should be assessed alongside that complexity. A small firm with a modest workload may find a simple, bounded arrangement easier to sustain than a sophisticated selection layer whose maintenance exceeds the value it creates.
Finance, operations and technology need the same task boundary
AI costs can fall into different budgets while benefits appear elsewhere. An IT team might pay for access, an operations team might provide reviewers and a business unit might claim the resulting capacity. An illustrative reporting arrangement that captures only the first budget would miss part of the delivery cost. The organisational response is to agree which workflow is being evaluated and who owns its result. That owner need not personally control every component, but should be able to bring the relevant evidence together. Without that shared boundary, separate departments can each report success while the company cannot judge the whole process.
A practical pilot could therefore involve the people who understand service quality, the people who maintain the system and the people who assess expenditure. They could agree the starting point, completion criteria and period over which the comparison will be made. This is a proposed operating approach, not a proven intervention from the survey. It should also include a way to record exceptions, because unusual cases may consume the expertise that an average hides. The purpose is not to build an elaborate reporting exercise. It is to prevent a local reduction in visible effort from being mistaken for an organisation-wide improvement.
Small businesses can begin with a narrow record
The survey’s large-company population should not be used to prescribe enterprise-scale machinery for a sole trader. A small business could begin with one recurring task and a straightforward record of direct charges, preparation, review and final acceptance. An illustrative independent designer might assess assistance with internal proposal preparation before expanding it to other activities. The comparison would need to include the time spent correcting descriptions, checking commitments and adapting the result for the client. Such a record would not produce a scientifically generalisable estimate, but it could reveal enough about the firm’s own work to inform a limited decision.
The choice of task matters as much as the record. A process with a clear output and repeatable quality criteria is easier to assess than an open-ended promise to make the whole business more intelligent. Starting narrowly also limits the burden of preparing data and training colleagues. If the assessment shows that reorganising existing information solves much of the problem, that is useful learning rather than a failed AI project. The business objective remains better work. The technology earns its place when its contribution can be identified within that work, with requirements proportionate to the organisation’s capacity to maintain it.
Manage the economics as the process changes
An evaluation is a baseline rather than a permanent verdict. A workflow can change when staff become more experienced, a supplier changes pricing or a model update alters the output. Periodic reassessment could ask whether the original benefit persists and whether the cost of exceptions has moved. It should also consider the consequences of expanding volume: a successful small pilot may encounter different cases when used more widely. These are reasons to retain a clear link between usage and accepted outcomes. They do not imply that every variation requires a new procurement exercise, only that a decision should remain open to evidence from operation.
The important business message in the new survey is that confidence in AI’s potential does not remove the need to understand delivery. A visible bill is one part of that delivery, and a polished output is one stage of it. The useful unit lies where the organisation can say that the work is complete and good enough for its purpose. By following the resources needed to reach that point, leaders can make more credible choices about scope, tools and investment. The result may be expansion, redesign or a decision to stop. Each can be sensible when it follows a clear account of what the business receives for what it spends.
Image credit: Desktop with laptop and calculator, anonymous/unknown photographer as recorded by Wikimedia Commons, originally published on Unsplash in 2016 under CC0 1.0. Illustrative photograph unrelated to Capgemini or the surveyed organisations. Resized for this site; WordPress may crop display versions.
