Hours saved aren't dollars saved: turning AI productivity into financial value

In brief

AI time savings become money only when someone acts on them. Four routes that reach the P&L, how to measure each, and a worked example of the gap.

Steve Elliott

Steve Elliott · · 6 min read


On this page
  1. Why don't AI time savings show up in the P&L?
  2. How big are measured AI productivity gains?
  3. Which routes turn time savings into money?
  4. A worked example: what is 40 minutes a day worth?
  5. What should finance measure before banking the benefit?
  6. Sources

AI time savings become money only when someone acts on them: a planned hire is not made, a contractor is released, revenue work moves faster, or more output is sold. Left alone, saved minutes are absorbed into other work, which is what a large study linking chatbot use to Danish payroll records found.

In a survey of about 25,000 Danish workers, 85% of people using chatbots said they moved the time saved to other job tasks, and records showed no measurable effect on earnings or hours. In our view, a business case that multiplies minutes saved by a wage rate is counting capacity, not cash.

Why don't AI time savings show up in the P&L?

Because time is not a line in the ledger. Payroll, contractor invoices, revenue and working capital are. Saved time reaches one of those only through a decision, and the evidence so far suggests few organizations have.

  • Slack is absorbed. In the same study, by Humlum and Vestergaard for NBER and revised March 2026, fewer than 10% of chatbot adopters reported taking more breaks or leisure, and about 30% reported spending more time on the tasks they had sped up. The estimates rule out effects on earnings or hours larger than 2%, two years after ChatGPT's launch.
  • Little capacity has been released. The Census Bureau's 2026 AI supplement to its Business Trends and Outlook Survey, covering November 2025 to January 2026, found that 66% of AI-using firms use it only to augment tasks. AI-related employment decreases occurred in only 2% of firms.
  • The aggregate number is small. In the Real-Time Population Survey, reported by the St. Louis Fed, self-reported time savings among employed US adults rose from 1.6% to 2.2% of work hours over roughly Q3 2024 to Q2 2026. Bick, Blandin and Deming, who built the survey, note that workers who finish faster without their employer knowing may take the time as "on-the-job leisure", which would raise welfare but not productivity.
  • Self-reports can run high. In one small study, METR's early-2025 trial, 16 experienced developers expected AI to speed them up by 24%, and afterward believed it had sped them up by about 20%. Measured, they took 19% longer.

None of this says AI doesn't make people faster at tasks. It says the speed has to be converted, deliberately, before finance can see it.

How big are measured AI productivity gains?

Real, uneven, and smaller at the level of a job than at the level of a task. The studies most often cited in business cases measure different things, and their limits matter as much as their headlines.

StudyDesignSampleResultMain limit
Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025Staggered rollout of a conversational assistant5,172 customer support agentsIssues resolved per hour up 15% on average; the 2023 working paper put novice and low-skilled gains at 34%Most experienced agents saw small speed gains and small declines in quality
Dell'Acqua et al., Harvard Business School working paper, 2023Pre-registered randomized field experiment758 BCG consultantsInside AI's capabilities: 12.2% more tasks, 25.1% faster, over 40% higher qualityOutside them: 19 percentage points less likely to be correct
METR, July 2025Randomized by task16 experienced open-source developers, 246 tasks19% slower with AI (confidence interval 2% to 39% slower, per METR's 2026 update)Small sample on large, familiar codebases; early-2025 models
METR, February 2026 updateRandomized by task57 developers, 800+ tasksReturning developers 18% faster, new developers 4% faster; both intervals include zeroAuthors call it "very weak evidence"; many developers declined to work without AI
Anthropic, November 2025Model-estimated time per task from 100,000 conversationsClaude.ai conversationsEstimated time per task cut by about 80% (median)Excludes time spent checking or refining output outside the conversation; estimates not validated against real-world outcomes

Two readings follow, in our view. First, task-time cuts of about 80% and economy-wide savings of 2.2% of hours are not in conflict: most of a job is not the task the assistant speeds up. Second, quality moves too, in both directions. The BCG experiment found large quality gains on some tasks and more wrong answers on others. A benefit model that counts only minutes misses both.

Which routes turn time savings into money?

We see four, and each lands in a different place on the financial statements. The labels are ours; the measures are standard finance.

RouteWhat has to happenWhere it showsHow to measure it
Avoided hires or unfilled leaversVolume grows and the hiring plan shrinks to match, or leavers are not replacedCost avoidance against plan for avoided hires (payroll flat); lower payroll for unfilled leaversApproved headcount plan before the AI decision vs actual, adjusted for volume
Contractor or outsourcing releaseVariable labor is cut by the capacity freedLower invoices within a quarterContractor or BPO spend vs a pre-launch baseline, per unit of volume
Cycle time on cash or revenue workA process that gates cash or launches gets shorterWorking capital (one-time cash) or earlier revenueDays sales outstanding, close days, time to quote, time to market
Throughput soldMore output per person, and demand exists to buy itRevenue per headOutput per FTE and conversion against a control group with the same mix

Not backfilling leavers turns capacity into payroll savings without layoffs, but only if someone owns the decision to hold the role open. Each route needs a named owner and a baseline set before launch. Without both, in our view, the benefit is unauditable.

A worked example: what is 40 minutes a day worth?

The numbers below are illustrative. They don't come from any company.

A claims operation has 100 employees at a fully burdened labor rate of $90,000 a year, plus overflow contractors equal to 10 FTE at $120,000 each. An AI assistant, rolled out to employees only, reports in its usage dashboard that they save 40 minutes a day. The all-in run cost, covering licenses, integration, evaluation and training, is $120,000 a year.

The headline. 40 minutes over 220 working days is about 147 hours per person, or 14,667 hours across the team. At $90,000 divided by 1,760 hours, about $51 an hour, that is $750,000. This is the business-case headline.

The measured gain. A controlled comparison, matched on claim mix, shows claims closed per paid hour up 4%, not the 8.3% that 40 minutes of an 8-hour day implies. The rest goes to checking output, other tasks and rework. Capacity freed: about 4 FTE-equivalents, rounded (the same output now needs 100 divided by 1.04, or about 96.2 people), worth about $360,000 at employee cost.

What reaches the P&L depends on the decision taken:

DecisionGross valueNet of $120,000 run costWhere finance sees it
Headline in the business case$750,000$630,000Nowhere: it is a capacity estimate
Release about 4 FTE of overflow contractors$480,000$360,000Contractor invoices, next quarter
Employee-handled volume grows 4%; the 4 planned hires are not made$360,000$240,000Against plan; payroll flat year on year
No decision; capacity absorbed$0-$120,000Run cost only, unless quality or backlog has a priced value

The contractor route is worth more because contractors cost more per FTE. It is also the fastest to bank, because contracts can end within a quarter. The avoided-hire route is real money, but it is cost avoidance: a CFO comparing this year's payroll with last year's will see no change, so the comparison has to be against the approved plan.

A cycle-time variant. Suppose the same kind of assistant cuts invoice dispute resolution so that days sales outstanding fall by 2 days on $400 million of annual revenue. That releases about $2.2 million of cash once ($400 million divided by 365, times 2). At a 9% cost of capital, it is worth about $197,000 a year. The value sits on the balance sheet, and it is a one-time release, not an annual saving.

What should finance measure before banking the benefit?

We suggest five steps, in this order:

  1. Set the baseline before launch. Unit volume, output per paid hour, quality rate, contractor spend and the approved headcount plan, frozen on the day of the decision.
  2. Measure output, not reported time. Use output per paid hour against a control group or a staggered rollout, as the support and consulting studies did. METR's results show how far self-reports can drift from measured time.
  3. Pair every speed measure with a quality measure. Error rate, rework, escalations or customer satisfaction. The research shows quality can rise or fall with speed.
  4. Name the route and its owner. Which decision converts the capacity: the hiring plan, the contractor contract, the process change. Write it into the case with a date.
  5. Report net of the full run cost. Tokens are a small share of AI cost; we covered the split in AI initiative costs: tokens vs labor.

We think a CFO should fund AI productivity cases on the measured gain and the named route, and treat the headline hours as an upper bound. For the rest of a fundable case, see what a CFO-ready AI business case contains. AI value realization then becomes a tracking question: did the contractor invoices fall, did the hiring plan shrink, did days sales outstanding move?

In our view, capacity becomes money reliably when the AI investment, the team it serves, its plan and its budget sit in one operating model, where the decision to release capacity has an owner.

Sources

Frequently asked questions

Why don't AI time savings reduce costs on their own?

Because saved time is spent somewhere. In the Danish study by Humlum and Vestergaard, 85% of chatbot adopters moved saved time to other job tasks and fewer than 10% took more breaks, with no measurable effect on earnings or hours. Cost falls only when someone changes the hiring plan, contractor spend or a process that gates cash or revenue.

Why is the measured gain smaller than reported time saved?

Because reported time counts minutes on the task the assistant touches, not output across the whole job. Some of those minutes go to checking AI output, rework or other tasks. Anthropic's own estimate of tasks cut by about 80% excludes time spent checking output outside the conversation. Output per paid hour, measured against a control group, shows what survives those leaks.

Is avoided hiring a real saving?

Yes, but it is cost avoidance, not cost reduction. Payroll stays flat while volume grows, so a year-on-year comparison shows nothing. The saving is visible only against the headcount plan approved before the AI decision. We suggest freezing that plan as the baseline and having the business owner, not the AI team, confirm each role that was not filled.

Are self-reported AI time savings reliable?

Not on their own. In METR's early-2025 randomized trial of 16 experienced developers, participants believed AI had sped them up by about 20% when they had in fact taken 19% longer. The St. Louis Fed's FRED Blog also calls survey-based time savings approximate. Use self-reports to find where to look, then measure output per paid hour and quality before banking any benefit.

Free playbook

Run the enterprise from any AI.

Read The Headless Enterprise
Steve Elliott

Steve Elliott

Published November 22, 2026

More on AI Economics

Tokens, inference cost, cost per task.