Gaurav Vashisht has over 22 years of experience building and supporting finance and enterprise systems.
Enterprises can count every token their AI agents consume. Almost none can confidently say what they got back and showcase real, tangible and quantifiable ROI.
That is the state of AI measurement today: adoption charts climbing in every direction and no defensible answer to the only question that matters to a board, which is what the investment returned.
This is the central unsolved problem of AI ROI, and it is about to matter enormously because next year’s AI budgets are being set right now on the basis of adoption charts rather than return.
Why The Number Is So Hard To Get
Return needs a denominator. To know what an AI agent saved, you need to know what the work cost before.
This is where the trouble starts because nobody in the building agrees. Ask finance how long an invoice reconciliation takes, and they may say five minutes because they only see the approval click. Ask operations, and they may say five hours because they do the matching. Ask engineering, and they may say five days because they count the exceptions. Build an ROI figure on any one of those numbers, and the other two departments will dispute it. Disputed numbers do not get acted on.
The problem gets worse when the work is knowledge work. A developer fixes a bug with the help of a coding agent. The ticket says it took six days, but the developer handled 20 other tickets and attended a conference in between. The agent log says it took 30 minutes, but before opening the agent, the developer listened to podcasts, sat through team calls and consulted senior colleagues. The real effort happened outside of anything the tools can see. Undercount it, and you overstate what AI did. Developers know this, which is why they do not trust the numbers, which is why the numbers never get used.
And there is a third trap. Once you can attribute work to AI, the instinct is to use it to rank people. Do that naively, and the top of the leaderboard fills with the heaviest token consumers and the most prolific pull request submitters. Then talk to their managers. Often, these are the people whose code gets rewritten every week and whose bugs keep coming back. Volume is not impact.
Three Ideas That Close The Gap
The approach I have been developing to use in my own work takes each of these problems in turn.
1. Stop Asking What The Work Costs And Observe It
Connect to the systems where work actually happens (the ERP, the ticketing tool, email, chat) and time each step of a workflow inside its own system. Take the median across people and a rolling average across months.
On the first day of measurement, a manager can click a button and see that a workflow takes 3.9 hours manually, based on hundreds of observed runs. That is a measurement, not an opinion, and it ends the five-minutes-versus-five-days argument for good.
2. Reconstruct The Effort That Happens Outside The Loop
When a pull request is submitted, walk the graph of everything connected to it, including the ticket, the agent sessions, the calls, the chat threads, the documents and the research. Each carries its own observed time. Add them up, then apply what I call a “channel tax,” a multiplier that grows with the number of distinct channels the work passed through, because coordination has a cost no single system records.
Developers look at the result and recognize their week reflected in it. That recognition is what makes the number usable for incentives.
3. Penalize Rework
If the same bug comes back to the same person, or a later revision rewrites most of their code, the earlier credit was not earned, and it is reduced. Small edits are ignored; code is always touched. But a weekly pattern of wholesale rewrites moves a contributor down the ranking, not up. It is attrition analysis applied to AI: We ask not just who produced output but also whose output lasted.
What Happens When Value Becomes Visible
Put the three together on top of the tools a company already runs, through connectors to its agents, skills and workflows, and the AI conversation changes. Hours saved per week and per month become reported figures, traceable to observed baselines. Business value sits next to token cost, so return per dollar is a number rather than a slide. Every figure carries an audit trail back to the runs, people and rates that produced it.
Something else happens, too, and it surprises people. Total AI spend goes down. Not because anyone mandates it, but because teams can finally see which skills and agents create value and which do not, and they stop paying for the second kind. Measurement, it turns out, is the most effective cost control there is.
Outcome And Takeaway
Usage is a cost line. Every company running AI, regardless of which agents, skills or ecosystems they run, needs this. Not eventually. Now, while next year’s budgets are being set.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?