Measuring ROI From AI Agents

By Katy Rumbelow | Published On: 28 August 2026 |

Every board conversation about AI eventually lands in the same place. Someone asks the ROI question. Someone answers with a use case. Everyone nods. Nobody actually knows.

The truth is that most organisations deploying agents today cannot tell you, with any confidence, what return they’re getting. They can tell you the agent is running. They can tell you people like it. They might even have a case study. What they can’t do is show you the maths.

That has to change, and not because finance teams are suddenly going to demand it. It has to change because without it, every agent decision is a guess, and guesses don’t scale.

ROI for agents is not the same as ROI for software

The instinct, when a new technology arrives, is to measure it the way we measured the last one. Software ROI is relatively simple. You buy the licence, you deploy it, you count the seats, you compare against the productivity gain or cost saved. It’s a clean calculation because the cost is fixed and the value is broadly predictable.

Agents don’t behave like that. The cost is variable, consumption-based, and often invisible until the bill arrives. The value is behavioural, sometimes indirect, and rarely captured by a single KPI. A Copilot Cowork session that compresses two weeks of work into an afternoon is worth a lot, but it doesn’t show up as a licence saving. A poorly scoped agent burning tokens on the same query twenty times a day is expensive, but it’s spread across a consumption line nobody scrutinises.

If you measure agents with software-era logic, you will get software-era answers, and they will be wrong.

The three signals that actually matter

When we talk about Agentic FinOps, we’re really talking about three signals, and the discipline of looking at them together rather than in isolation.

Cost signals. Not just the licence. The full stack. Consumption charges, per-message metering, model calls, Cowork sessions, storage, the human time spent maintaining and monitoring, and the infrastructure the agent depends on. Most organisations have visibility into one or two of these. Very few have all of them in one view. Until you do, your cost number is a fragment, not a total.
Value signals. Outcomes delivered, not activity performed. Hours saved that were reinvested in higher-value work. Cycle time reduced on a specific process. Error rates down. Customer satisfaction up. Revenue enabled, not just cost avoided. The trap here is measuring activity (the agent ran 4,000 times this month) and calling it value. Activity is not value. Activity is the receipt, not the meal.
Opportunity signals. This is the one most organisations skip, and it’s the one that matters most. What was the alternative? A human doing the same work? A different agent doing it better? Doing nothing? The opportunity cost of the agent you deployed is the value of the agent you didn’t. And the opportunity cost of the agent you didn’t retire is the value of the one you could have built instead. I wrote a piece recently about how 1 session cost me £77 – but that compressed weeks of my time, and a lawyers time, into just an afternoon, producing a number of heavy legal document drafts which would have cost thousands. It sounds much more worth it when you have the full picture.

Look at all three together and you get something close to the truth. Look at any one in isolation and you’re either overselling or underselling the estate.

The Important questions to ask

If you want to know whether you’re actually getting ROI from your agents, here are the questions to put on the table. They’re to the point and take time to answer properly, which is why they don’t come up often enough.

Which of your agents delivered measurable value in the last quarter, and how do you know?
Which of your agents cost more to run than the value they produced, and are they still running?
If you retired the bottom ten percent of your agent estate tomorrow, what would you lose?
If you doubled the investment in the top ten percent, what could you gain?
What is the fully loaded cost of your agent estate this month, including human effort, and is that number trending in the right direction?

Most leaders I ask cannot answer more than one or two of these today. That is not a criticism. It’s a symptom of an operating model that hasn’t caught up with the pace of deployment. You cannot answer ROI questions if nobody owns the agents (Blog 3), if you cannot see them (Blog 2), or if you haven’t been honest about what running them costs you (Blog 1).

Where the AOC fits

An Agent Operations Centre is, at its heart, a discipline for answering those five questions every month, for every agent, without it becoming somebody’s second job. The inventory feeds the cost view. The ownership model feeds the value view. The review cycle feeds the opportunity view. Put together, they give you a running ROI picture that stands up to a board conversation and, more importantly, informs the next decision you make about the estate.

That is the real point of Agentic FinOps. Not a dashboard. Not a report. A habit of asking the ROI question early, often, and honestly, and having the operating model to act on the answer.

Closing thought

The organisations that will win the next phase of AI are not the ones with the most agents. They are the ones who know exactly what each of their agents is doing, what it’s costing, and what it’s worth. Everyone else will be running on faith, and faith doesn’t scale.

Agents are the easy part. Measuring them properly is the work.