Nearly three in five large enterprises spent significantly more on AI over the past year than they expected or budgeted, and the ones that overshot ran 24% over on average. One in five organizations blew through their budget by more than 30%. Those are the headline numbers from The cost of intelligence, a research snapshot the Capgemini Research Institute released this week, based on a September 2026 survey of 500 senior executives at companies with more than $1 billion in revenue across 10 countries and 13 industries.
The overruns were not free. Two-thirds of the organizations that overshot their AI budget reported negative financial consequences, such as lower free cash flow or cuts to discretionary spending. Concern is rising with the bills: 53% of executives now say they are worried about the overall cost of AI, up from 45% a year ago, and only 56% are fully confident they can forecast how AI costs will evolve as adoption scales. Capgemini quotes Uber CTO Praveen Neppalli Naga, who put the mood in one line: “I’m back to the drawing board, because the budget I thought I would need is blown away already.”
The most important finding, though, is not the size of the overrun. It is that most enterprises cannot say what the money bought. That turns AI cost optimization from a procurement exercise into a measurement problem, and it is the problem we keep returning to on this blog, from the Uber AI budget blowout to the unit cost gap in Deloitte data.
What Capgemini found
Start with visibility. Six in ten organizations lack a consistent, enterprise-wide view of their AI costs. The gaps are not random, either. Among executives responsible for AI cost management, 81% say their organization actively measures cloud, compute and infrastructure costs, and 79% measure model, application and consumption costs. Only 51% measure the cost of human oversight, skills and workflow change, and just 48% measure ongoing operations, retries, failures and rework. In Capgemini’s words, the least visible AI costs sit inside the operating model, not the monthly model bill.
Then look at value. Among organizations with established AI spend tracking that could estimate their coverage, 61% of AI spending can be systematically identified and tracked, but only 45% can be linked to verified business value. Put differently, even the enterprises with the most mature AI spend tracking cannot connect more than half of it to an outcome anyone has checked. Only one in five organizations calculates what Capgemini calls AI Yield, the verified value of an AI-assisted outcome after accounting for the full cost of delivering it safely, including model usage, infrastructure, integration, human oversight, governance and rework.
Agents are about to make all of this harder. Capgemini cites OpenRouter data showing agentic token use grew 14 times in six months while human-driven use grew 2.8 times, and Stanford Digital Economy Lab research finding that agentic coding tasks use roughly 1,000 times more tokens than code chat, with the same agent on the same task varying up to 30 times in cost. A bill that moves that much from one run to the next cannot be managed with an annual budget line.
The cheaper-model reflex in AI cost optimization
How are enterprises responding? Mostly by squeezing. Among organizations that overran, 53% tightened budgets or introduced spending controls, 22% kept going because the expected value justified the cost, 20% rethought their tools or processes, and 5% paused or stopped the affected initiative. Looking ahead, 61% expect to adopt a multi-model strategy that selects models by cost, capability, latency and risk, and 58% plan to use self-hosted or externally hosted open-weight models. As Reuters reported on Thursday, that shift echoes SAP, whose finance chief said last month it was replacing some of its costliest models with cheaper ones where performance held up.
The economics behind the shift are real. Capgemini says open-weight tokens cost 87% less on average than closed-model tokens, and on OpenRouter open-weight models handled 72% of tokens in September while accounting for only 27% of spend. Research on task-based routing cited in the snapshot shows cost reductions of 27% to 85% against always using the strongest single model, depending on the quality level a team accepts. We have covered the mechanics in how model routing works and the trade-offs of self-hosting your own model.
But a cheaper token is not the same as a cheaper outcome. If a smaller model gets fewer answers right on the first pass, the savings on tokens can disappear into retries, human review and rework, which are exactly the costs fewer than half of enterprises measure. We made this point when prices fell and bills kept rising in your AI got cheaper, your bill did not. Capgemini’s own advice is telling: route work to the lowest-cost model capable of completing the task, which presumes you know, task by task, which model completed it and at what quality. Capgemini also predicts first-pass success rate will become the top-priority AI metric, and that cost per successful outcome will gain ground on token counts and total spend.
Agents break the budget-line approach
Banks are learning the same lesson in public. According to Reuters coverage of McKinsey’s Global Banking Annual Review, published November 10, spending on the tokens that power AI agents could easily reach 20% of total labor costs at banks, and responsibility for controlling it has shifted to new teams some banks call AI FinOps. Early efforts to route employee queries to cheaper models by complexity stopped working once agents were involved. “Those models are ineffective when AI agents are working, because they communicate with each other and tokens quickly add up,” McKinsey senior partner Vik Sohoni told Reuters. The response, he said, is to be more selective about where agents are deployed and to focus on the areas with the strongest expected returns.
That is the right instinct, and it requires a dataset most enterprises do not have: cost and outcome per agent, per workflow, per team, across every vendor. A FinOps for AI practice that only sees invoices can cap spending. It cannot tell you which spending to protect. This is the gap Olakai was built for. Olakai is a vendor-neutral Enterprise AI Intelligence Platform that acts as a system of record for the entire AI stack, connecting usage and spend from coding tools, assistants and agents to the business outcomes they produce. Instead of a monthly bill by provider, finance and AI teams see cost per completed task, cost per successful outcome and AI ROI by use case, which is what Capgemini’s AI Yield asks for.
What CFOs and AI leaders should do now
Capgemini found that CFOs now help approve major AI investments at 52% of organizations, up from 42% a year ago, so finance is already in the room. The question is what it brings to the table. Four moves turn AI cost optimization from cutting into choosing:
- Measure the full cost of an outcome, not just the model bill. Include oversight, retries and rework, the cost lines fewer than half of enterprises track today.
- Make first-pass success rate and cost per successful outcome standard metrics for every material AI workflow, and compare them with a non-AI baseline.
- Before swapping models to save money, benchmark the cheaper option on real tasks and keep measuring after the switch, so savings on tokens are not lost to rework.
- Review the portfolio on a fixed cadence and decide to scale, fix or retire each initiative based on realized value, not on how much it consumed.
Each leader needs a different cut of the same record. The CFO and FinOps team need AI spend attributed to owners and outcomes so a budget overrun can be explained, not just stopped. The head of AI needs to see which models and agents deliver at the required quality, so routing decisions are made on evidence. With Olakai, these come from one dataset, and the analytics and custom KPIs let each team track the outcome metrics that matter to its own workflows.
Governance considerations
Cost control and governance are converging. A multi-model portfolio with open-weight models in the mix means more providers, more hosting arrangements and more places where data flows, and every new route is a new surface to govern. Capgemini recommends defining permitted agent actions and autonomy levels, with risk-based guardrails and human escalation. That oversight has a cost, and it should be measured like any other. In regulated sectors such as financial services, a cheaper model that needs more human review may not be cheaper at all. AI governance that records which model handled which task, under which policy, gives both the auditor and the CFO what they need.
Where to start
Pick the five AI workflows with the largest spend and answer three questions for each: what did it cost last month including human review and rework, what share of its outputs were right the first time, and what verified outcome did it produce. If those answers take a week of spreadsheets, a model swap will save money you cannot see and may cost value you cannot measure. The choice in front of most enterprises is to put a measurement layer like Olakai in place now, or to join the three in five that overshot their AI budget and still cannot say which half of the spend was worth it.
Talk to an expert about measuring cost per successful outcome across every model, agent and AI tool in one record, and see how the platform turns AI spend into decisions.