The measured size of it
- 24%
- Average AI budget overrun
- 53%
- Executive concern about AI cost
- 52%
- CFO involvement in AI approvals
Agentic AI is a software system that takes a stated objective, breaks it into steps, calls the tools and data sources those steps require, checks what comes back and repeats the cycle until the task is finished or a person takes over. It spends money on the way to an outcome, so the unit of account is a completed task.
A claims team at a UK insurer built one to triage motor claims. The demonstration persuaded the room: the agent read the claim, checked the policy, requested a contractor quote, booked an assessor and drafted the customer letter. Then the pilot invoice arrived. Token spend had risen with every retry, a reviewer was sampling outcomes, the integration had taken longer than the model work, and the savings case still rested on a spreadsheet of assumed handling times. The question became whether the completed job cost less than the value it released.
That is the test this page makes routine. ET Edge Insights reported on a Capgemini Research Institute study, The Cost of Intelligence, published in November 2026, which found that nearly three in five organisations had seen AI spending exceed expectations or budget over the previous year. The study's answer is a ratio it calls AI Yield: verified business value divided by the full cost of safely delivering an AI-assisted outcome. The phrase ET Edge Insights uses is "controlling the cost of intelligence". That is the right frame for a UK board, because the exposure sits in the gap between a pilot that works and a business case that holds. Agents that fail get shut down early. Agents that work well enough to spread before anyone has proven the economics are where the capital goes.
A completed task is the unit of account
Finance should fund a completed, checked piece of work, priced across every attempt it took. An agent that closes a ticket incorrectly has created rework, a customer complaint and a second review. The full cost of that outcome includes the first attempt, the correction, the oversight and the delay. AI Yield puts all of those items in the denominator. Adoption metrics do not. A seat count or a token total can rise while the business gets no better at the work.
The discipline is to define the outcome before the pilot. For the claims agent, the outcome is a claim triaged to the right handling route within the service standard, with the customer informed and no rework. That definition gives a finance team something it can audit. It also gives the agent a stopping condition. If the agent cannot reach that state, a person must take over, and the handover cost belongs in the business case.
Elementum, writing about return on investment, reports that 95% of enterprise generative AI pilots deliver no measurable profit-and-loss impact. That figure is a warning about measurement design. A pilot that cannot show its baseline cannot show its value, and a pilot that counts only the successful runs has hidden the cost of the failures.
The full cost sits in the invisible layers
The costs that decide an agent's economics are oversight, integration, retries, failures and rework. The figures in this table come from the Capgemini Research Institute study, The Cost of Intelligence, as reported by ET Edge Insights.
| Measure from the Capgemini study | Finding |
|---|---|
| AI spending above expectation or budget | Nearly three in five organisations |
| Average overrun among those organisations | 24% above budget |
| Organisations more than 30% over budget | One in five |
| Overrun organisations reporting negative financial consequences | Two-thirds |
| Organisations without a complete enterprise-wide view of AI expenditure | Six in ten |
| Organisations measuring cloud, compute and infrastructure costs | 81% |
| Organisations measuring human oversight, skills and workflow change | 51% |
| Organisations tracking operations, retries, failures and rework | 48% |
| Organisations calculating AI Yield | One in five |
| Expenditure connected to verified business value among spend-trackers | 45% |
| Executives fully confident in forecasting AI costs as adoption scales | 56% |
| Executive concern about the financial cost of AI | 53%, up from 45% a year earlier |
| Agentic token consumption growth at OpenRouter in six months | 14-fold, against 2.8-fold for human-driven usage |
| Token consumption of agentic coding tasks compared with code-chat | About 1,000 times |
| Executives expecting multi-model strategies | 61% |
| Executives expecting more open-weight models | 58% |
| Executives expecting more smaller or specialised models | 50% |
| Performance gap between leading closed and open-weight models | About 3.3% |
| Open-weight token price advantage in the comparison cited | About 87% lower |
| Cost reduction from intelligent model routing | 27% to 85% |
| CFO involvement in approving major AI investments | 52%, up from 42% a year earlier |
The pattern is clear: organisations can see infrastructure spend and miss the work that makes an outcome reliable. AI Yield forces the denominator to include every layer that safely delivers the outcome. The study found that few organisations calculate it. Practical control starts with instrumentation. Log every model call, tool call, retry, human review and exception. Assign a cost per completed task. Review the number monthly with finance and operations in the same room. The Capgemini study shows agentic usage growing much faster than human-driven usage, and coding agents consuming orders of magnitude more tokens than code-chat. A cheaper token price can sit beside a larger bill.
Route work to the cheapest capable model
Model routing turns cost control into an architectural decision. Routine classification, summarisation and structured decisions can run on smaller or open-weight models. High-stakes exceptions can go to a frontier model. Set the routing rule against a quality threshold and a cost per completed task.
| Work type | Model class | Quality gate | Cost check |
|---|---|---|---|
| Routine classification | Small or open-weight | Accuracy above an agreed threshold on a held-out set | Cost per completed task below the baseline |
| Structured decision | Mid-tier or open-weight | First-pass success rate and a review sample | Cost per completed task with rework included |
| High-stakes exception | Frontier | Human sign-off and a complete audit trail | Value at risk and oversight cost |
Katherine Kellogg at MIT Sloan warned that reclaiming 20% of someone's time does not translate into a 20% labour-cost saving. Report reclaimed capacity as capacity until a paid cost leaves the budget. Model routing only works when the quality gate is measured on the work the agent actually does.
Map the process before choosing the tool
The first agent should be chosen because the work needs contextual judgement across tools. Many processes that look agentic on a slide are better served by rules or workflow automation. The test is whether the work requires the system to interpret context, choose between tools and act within a boundary. If the task follows fixed rules, an agent adds cost and variability for no gain.
| Option | What it does | Best suited to | Typical failure |
|---|---|---|---|
| Agentic AI | Plans steps, calls tools, checks results and acts towards a goal | Work with variable inputs, multiple systems and a need for contextual judgement | Over-privileged access, unpredictable behaviour and rework that is not counted |
| Robotic process automation | Follows predefined rules and moves data between systems | High-volume, stable, rule-based tasks | Brittle when the process or interface changes |
| Workflow automation | Coordinates a defined sequence of steps and approvals | Processes with clear stages, owners and handoffs | Automates a process that was never mapped correctly |
TechRadar published an account from Pete Evans at Atos of a procurement process that looked like four steps on paper and closer to ten in practice. People moved between ERP, Outlook and spreadsheets, copying information, waiting for replies and re-entering data. A baseline on four steps would have measured the wrong process. An agent given four steps could not have completed a job that takes ten.
MIT Sloan reported that in one agentic project 80% of the work was data engineering, stakeholder alignment, governance and workflow integration. Prompt engineering and model fine-tuning made up the rest. That is the budget line most business cases miss. Map the process as it is done, then decide whether agentic AI, RPA or workflow automation fits.
Set gates that can stop the rollout
A rollout needs gates that can stop it before the bill compounds. Stage the work in three gates: bounded pilot, controlled rollout and scale. Each gate needs evidence that finance and operations can test.
| Gate | Evidence required | Stop condition |
|---|---|---|
| Bounded pilot | Baseline, cost per completed task, first-pass success rate and human override rate | Cost per completed task above the baseline or override rate above the agreed ceiling |
| Controlled rollout | Verified value against AI Yield, rework rate, audit trail and a named owner | Value unverified for two review cycles or rework above the threshold |
| Scale | CFO sign-off, model routing performance, incident review and a capacity plan | Forecast cost per completed task exceeds the approved case |
Governance is part of the cost model. The National Cyber Security Centre advises teams to start with tightly bounded pilots, grant least privilege and name a human owner for every agent. Brookings and Carnegie Mellon University convened evaluators who concluded that static benchmarks cannot predict deployed behaviour; field testing and continuous monitoring are needed.
The Capgemini study, as reported by ET Edge Insights, shows that budget tightening was the common response to overruns, and few organisations changed their underlying architectures. Finance is already in the room. The board question changes from whether an agent can do the task to whether the completed task pays for itself. AI Yield gives the board a number it can test. Start with a bounded workflow, capture the baseline, count the full cost, route each task to the cheapest capable model, and set a gate that can stop the rollout. Then scale the agents that clear the gate.
Sources
- Why enterprises must measure business value before scaling agentic AI - ET Edge Insights
- Stop choosing the wrong Agentic AI use cases | TechRadar
- Agentic AI, explained | MIT Sloan
- Agentic AI ROI: How to Prove Value to Finance in 2026 | Elementum
- Thinking carefully before adopting agentic AI | National Cyber Security Centre
- How can we best evaluate agentic AI? | Brookings
Source: Why enterprises must measure business value before scaling agentic AI, ET Edge Insights





