AI ROI Hidden Costs: The Hours Nobody Puts in the Model

The largest hidden cost in an AI ROI model is not tokens, licenses, or compute. It is labor. Glean's Work AI Index, a survey of 6,000 full-time digital workers across the United States, United Kingdom and Australia conducted between December 2025 and January 2026, found that workers spend an average of 6.4 hours a week feeding AI systems context, supervising their output, debugging their mistakes, and cleaning up after them. The report names this botsitting and defines it as the labor of making AI usable. The same survey found workers reporting roughly 11 hours a week saved through automation. Both numbers came from the same respondents in the same instrument. Almost every AI business case circulating in enterprises right now carries the second number and not the first.

Two things are true at once. AI is producing real time savings that employees can feel and describe. And the gross figure they describe is being presented to boards as if it were a net figure, because the offsetting cost lands in a place no financial system observes.

What is the hidden cost in an AI ROI model?

An AI ROI model usually has a clean denominator and a generous numerator. The denominator holds things that generate an invoice: seats, tokens, compute, integration work. The numerator holds time saved, usually collected by asking people how much time they saved.

Supervision labor belongs in the denominator and is almost never there. It generates no invoice, because the person doing it was already on payroll. It generates no vendor telemetry, because correcting a bad output happens after the session ends and often in a different application. It is a reallocation of hours the company had already bought, which is precisely why it is invisible: nothing changed on any budget line, so nothing triggered a review.

Where the time actually goes

The Work AI Index breaks the 6.4 hours into components. The pattern matters more than any single figure, because it shows the cost is not one behavior but four.

Component shares are rounded and sum to 39%. The report states the total at 37% of all AI interaction time, against 36% spent producing work and 27% spent learning tools and building agents. The report also found that 36% of AI sessions fail outright, requiring a full restart or substantial rework, and that 77% of workers had corrected or redone AI-assisted work in the previous month.

How to read these numbers honestly

Two caveats belong on the table, and stating them is not a reason to discard it. First, every figure here is self-reported, which makes it subject to recall bias in both directions: the 11 hours saved and the 6.4 hours spent are equally soft. Second, the report is published by a vendor whose product addresses the problem it describes. That is a structural conflict, not an accusation, and it applies to most published AI research including studies whose conclusions we like. Treat the direction as credible and the magnitude as an estimate your own organization should replace with a measurement.

Why does the cost stay invisible?

Three mechanisms keep supervision labor off the ledger, and they compound.

The first is that it is labor, not spend. Finance can see token consumption climb on the invoice. It cannot see an engineer spending an extra half hour reviewing generated code, because that engineer's cost did not change.

The second is that the spend side is already incomplete, so nobody is looking further. The 2026 State of AI Cost Governance Report from Mavvrik and Benchmarkit, based on 396 enterprise organizations surveyed in April and May 2026, found that only 11% of organizations can forecast AI costs within plus or minus 10%, down from 15% the year before, and that while 98% of engineering organizations use AI coding assistants, only 42% of organizations include developer AI tools in AI cost reporting. The same study reports that 62% of organizations had an unexpected AI cost materially alter a business decision in the past year, and that a quarter delayed or cancelled an AI initiative outright, the finding CFO Dive led with when it covered the report. Mavvrik sells AI cost governance software, which is worth stating plainly. Even discounted for that, the study describes organizations that cannot yet close the books on the invoiced portion of AI. The uninvoiced portion is not a priority when the invoiced portion is still moving.

The third is that most measurement points at the tool rather than the work. Seats, sessions, and prompts describe what happened inside the application. Supervision happens outside it. This is the same structural blind spot we described when shadow AI moves work outside the instrumented perimeter: the instrument is pointed at the software, and the cost is in the workflow.

This pattern is old, and it predates AI. In our 2024 workforce research conversations, a CFO at a private equity backed IT services holding company described spending three days on manual reconciliation before every cash flow forecast, because each acquired agency kept its own tools. Three days of senior finance time, every cycle, appeared in no system as a cost. The work was real. The record of it was not.

What Deloitte's readiness data says about the mechanism

Deloitte's agentic AI readiness survey, fielded to 501 senior manager to C-suite respondents in the United States across five industries between April and June 2026, all at organizations at least piloting agentic AI, found readiness inverted against where the work happens.

Strategy is the most prepared area. Business processes and workforce are the two least prepared. Deloitte reports that most leaders it interviewed said their organizations are layering AI agents on top of existing processes rather than redesigning those processes, and that fewer than a third expect the majority of their processes to be redesigned around agentic AI within two years.

Layering is what produces supervision labor. When an agent is placed on top of a process that was not changed to accommodate it, a person has to stand in the gap between the two: supplying the context the process never captured, checking output the process was not designed to validate, and reconciling the result back into a system the agent cannot write to. That person is the integration layer. Their time is the cost of not redesigning the process, and it is charged to the workforce rather than to the project.

What an honest measurement looks like

The fix is not another survey. It is measuring the work, because supervision labor leaves a trace even when nobody records it as a task.

1. Put a supervision line in the model. Even a rough estimate stated as a range is better than an implicit zero. An implicit zero is a claim, and it is the least defensible one in the document.
2. Measure rework, not just adoption. Adoption tells you the tool is open. Rework, revision counts, reopened tickets, and passes to close tell you what the output cost to make usable.
3. Compare cohorts rather than periods. Teams that adopted a capability against comparable teams that did not, controlling for tenure, role mix, work type, and tool stack, with a confidence score attached and the limits disclosed. That is controlled cohort analysis with confidence scoring, and it is described in full on the Levos measurement methodology page.
4. Redesign one process instead of layering five agents. The Deloitte data suggests the supervision tax is the price of layering. One redesigned process will produce a cleaner return than five bolted-on agents, and it will be measurable.

The organizations that will be able to defend an AI number in twelve months are not the ones with the highest reported savings. They are the ones whose reported savings are net of something. A gross number that has never had a cost subtracted from it is not a return. It is a hope with a decimal point.

The cost side connects to two other pieces: AI cost governance on spend controls, and AI license utilization on seats versus use.

Frequently asked questions

What are the hidden costs of AI adoption?

The largest is labor rather than spend: the 6.4 hours a week workers report spending on feeding AI context, supervising output, and debugging. It never appears on a cost line because it is a reallocation of hours already inside payroll.

Why do AI ROI models miss the cost of supervising AI?

Because the model is assembled from things that generate a record. Supervision labor generates no invoice and no vendor telemetry, so the numerator is complete and the denominator is not.

How much of employee AI time goes to supervision rather than output?

By Glean's accounting, 37% goes to supervision, 36% to producing work, and 27% to learning tools and building agents. The report also found 36% of AI sessions fail outright.

Is AI spend reporting capturing the full cost of AI?

Not yet. Only 11% of organizations can forecast AI costs within plus or minus 10%, and only 42% include developer AI tools in AI cost reporting, according to the Mavvrik and Benchmarkit study of 396 organizations.

How do you measure the real net effect of an AI tool?

By observing throughput, cycle time and rework in the systems where work lands, and comparing adopting teams to comparable non-adopting teams with a confidence score attached, rather than asking people to estimate their own savings.

See what your AI investment is actually returning, net of what it costs to supervise. Levos is a Human Capital Operating System, the intelligence layer above your existing stack, measuring what changes in the work rather than what shows up in a vendor dashboard. Request a Demo, read how the AI Impact signal family separates adoption from outcome, or start with what a Human Capital Operating System is.

Levos is accepting design partner applications from US organizations of 150 employees or more with an active AI rollout.