At the quarterly steering committee, the CIO presents the renewal for the enterprise AI coding assistant. The vendor dashboard displays a steep graph of engineering hours saved. The CFO opens the labor forecast on a second screen. Every approved headcount position is still there. The offshore contractor run-rate is completely flat. The CFO asks how the organization saved forty thousand hours of engineering time without reducing the budget by a single dollar.
The honest answer is that the savings did not go into the budget. They went into the work. AI saves minutes. Finance budgets in full-time equivalents. The units do not match.
A developer finishes a task in forty minutes instead of an hour. That twenty minutes does not show up on the P&L. When you multiply twenty minutes across a thousand developers, the vendor telemetry rolls it up into tens of thousands of hours. But you cannot bundle twenty-minute increments spread across a thousand different desks into a single cuttable role. You cannot fire one-third of a business analyst. The saved time is fragmented. It arrives in small pieces across the day, absorbed by longer coffee breaks, extra edge-case testing, or a senior engineer leaving a lengthy comment on a junior developer’s pull request instead of picking up a new ticket.
Even if you could aggregate the time, speeding up one part of the workflow does not speed up the delivery. A platform team rolls out an AI assistant to hundreds of developers. End-of-year metrics show story points completed are up slightly. But the release train is still stuck behind a manual security review that takes two weeks. The developers write the code faster, but it just piles up in the queue waiting for security review. The productivity gain is real, but it hits the next constraint. The vendor measures generation speed. Finance measures end-to-end cycle time. The constraint in the middle remains unchanged.
Then there is the problem of where the invoice lands. The AI line item sits in the IT budget. The benefit lands in the business unit. A shared services team uses an AI tool to extract data and suggest coding for invoices. Following a corporate acquisition, the volume of invoices doubles. The team processes the new volume with the exact same staff. The CFO asks why the shared services budget did not drop. The AI absorbed the growth. That is avoided cost, not cash savings. Finance does not budget counterfactuals. If you do not hire an extra analyst, there is no budget line to point to as a reduction.
The mismatch repeats across the enterprise. Legal uses an AI writing tool to summarize contracts. The review cycle shortens by three days. The sales cycle compresses, and revenue increases. The CFO sees the revenue in the sales forecast but does not attribute it to the AI subscription sitting in the IT cost center. The benefit lands where the work happens, not where the tool is bought.
When a task becomes cheaper to perform, the organization often consumes more of it. A developer does not finish the ticket faster and go home. They hold the time constant and increase the quality. They add unit tests they previously would have skipped. They refactor a legacy module. An enterprise architect generates a migration strategy document in fifteen minutes instead of three days. The stakeholders, accustomed to taking a week to read such documents, leave it in their inboxes until the following Tuesday. The project timeline does not compress. The organization absorbs the velocity by raising its baseline expectations for output.
To have a productive conversation with finance, you have to stop using the vendor’s hours saved metric. Vendor math multiplies accepted prompts by an arbitrary estimate of human typing speed. That measures the velocity of generation, not the velocity of problem-solving. Generating two hundred lines of code takes seconds. Verifying that those lines do not introduce a race condition takes the rest of the afternoon.
Measure cycle time between specific constraints. Pick a repeatable workflow with a clear start and end. Measure the time from when a ticket moves to in-progress to when the pull request is opened. Include the review, correction, and security checks in the count. Compare similar tickets, not an easy month against a hard one.
Before deploying the tool, decide how the saved capacity will be used. If you do not decide, it will be spent on more work, or it will evaporate into comfort. Present the capacity to finance as a set of choices. We freed two thousand hours this quarter. We can use that to reduce contractor spend on the next statement of work, accelerate the compliance project, or improve test coverage. Then track the decision.
If every team saves five percent, no single team can cut five percent of its headcount. The hours are not missing. They are just doing different work.