A board can approve a software subscription in an afternoon. It can take a year to discover whose Saturday paid for it. NatWest Group reported that AI cut the handling time for complaints by about nineteen minutes per case, equivalent to more than 22,000 hours during 2025. A bank put its name to that figure in its annual report. An investor trying to price the opportunity for an AI supplier still needs the next page of the calculation. How many hours came off paid shifts? How many went into resolving other complaints? Who received the money the change made available? 72
Those are different questions for the bank and its employees. A customer whose disputed payment is settled on Thursday instead of the following Tuesday has gained something tangible. An employee who depended on overtime to cover the rent might lose income even while keeping the same job title and base pay. A junior employee released from gathering documents might finally get to hear the difficult calls. Another could spend the year checking generated summaries and never hear what a customer sounds like when the summary is wrong. Our accounts have a line for the software fee. They have little to say about either career.
What the clock measured
The cleanest productivity figure in this debate comes from a place where there was already a good reason to count the work. A software company wanted to know whether customers' problems had been resolved. It could observe how many conversations an agent handled, whether the issue was settled and how much time the agent spent at the desk. A worker could conduct several chats at once. A shorter conversation and a larger number of successful resolutions in an hour therefore measured different achievements.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered arrival of a conversational assistant across 5,172 support agents. The main rollout occurred in autumn 2020 and winter 2021. The system drew on past conversations, giving extra weight to chats involving strong performers. Agents could use or discard a suggestion. In the authors' preferred specification the tool increased output by 0.301 resolved issues per hour, roughly fifteen per cent of the pretreatment mean. Newer and previously lower-performing agents benefited most. Among highly skilled agents the small improvement in speed came with some deterioration in quality. 20
Two studies that cannot be averaged
One study observed customer problems resolved in an hour. The other timed developers working on issues they had chosen. Read each estimate against its own scale.
Issues resolved per hour
+0.3015,172 agents · approximately 15% more output
The preferred model adjusts for agent, tenure, location and month. The line is an approximate 95% interval of +0.237 to +0.365 issues per hour.
Sample and calculation
Rollout mainly in autumn 2020 and winter 2021. Table II, column 3 uses 12,295 agent-month observations and reports a coefficient of 0.301 with standard error 0.0329. The interval shown uses 1.96 standard errors either side of the estimate. The study describes a pretreatment mean of 1.97 resolutions per hour.
Time to finish an issue
19% more time16 developers · 246 tasks
Reported interval +2% to +39%. Experienced contributors specified work on familiar repositories before random assignment. Both ends of the reported interval imply a longer task with the early-2025 tools.
Original experimentThat last detail would matter to the people negotiating a bank contract. The senior employee whose calls taught the assistant how to respond might have little use for its script. A supervisor who rewards adherence could make her worse at precisely the difficult conversations through which her experience earns its return. The support study also documents what happened during outages, when suggestions temporarily stopped. The patterns in chat duration suggest some learning persisted. Outages were rare, and the measure available at that level was duration. For a bank interested in training, the relevant follow-up would include whether a colleague could reach the right decision when the assistant went dark.
The firm’s own agents received hourly pay and performance bonuses. Managers told the researchers that performance bonuses could account for twenty to forty per cent of typical take-home pay. The researchers lacked individual compensation records. A higher resolution rate could move an agent across a bonus threshold. Redrawing that threshold could return the gain to the employer. Two workers could gain equally on the operational measure and receive different amounts at payday. The study identifies where to look for that division. Resolving it requires the bonus schedules and the payments. 20
Look at the other clock. METR recruited sixteen experienced open-source developers to work on their own projects, where they had about five years of experience on average. They specified issues before each was randomly assigned to a condition permitting or excluding AI tools. Across 246 tasks completed with the early 2025 toolset, access increased completion time by nineteen per cent. Before starting, developers expected a twenty-four per cent reduction. Afterwards, they believed they had been about twenty per cent faster. A buyer who interviewed the people doing the work would have come away with the opposite sign to the estimate from the experiment. 21
By the follow-up, the experiment itself had become harder to run. Ten original participants returned and 47 new developers joined. The raw time estimates pointed towards faster completion, with intervals spanning either side of zero. Some developers declined to enter tasks they especially wanted to do with AI. Others were reluctant to take part under terms that could deny them a tool they used daily. The hourly study payment had also fallen from $150 to $50. Concurrent agents complicated timekeeping because a person waiting for one job could already be working on another. The researchers called for a change in design. 22
Put the two clocks beside each other and the tempting national forecast dissolves. One counts accepted customer outcomes per hour after access arrived at different times. The other estimates time required for a prespecified programming task under random assignment. Even the arithmetic runs in opposite directions. At constant demand and quality, a fifteen per cent rise in output per hour would allow 100 hours of identical work to be done in about 87 hours. It would take a further decision to cut thirteen paid hours from the rota. Nothing in either study establishes who could sign for that reduction at our bank.
The entries on the payslip
Anders Humlum and Emilie Vestergaard could follow workers across the boundary that frustrates an enterprise pilot. Working with Statistics Denmark, they linked surveys of chatbot adoption to administrative records of earnings, hours and jobs. Their late 2024 survey included about 25,000 people across 7,000 workplaces in eleven exposed occupations. In the March 2026 revision of their working paper, differences between adopters and comparison workers before and after ChatGPT's arrival yielded estimates near zero for earnings and recorded hours. The authors could rule out average effects larger than two per cent in their observation window, which ended in December 2024. 73
The same investigation found employees writing material, checking machine output and helping integrate the tools into existing processes. Some of these duties even reached people who never used the systems. Most chatbot users reported assigning time saved to other tasks. The source for that last finding was a survey. The source for earnings was an administrative record. Taken together they explain how a firm can change the contents of a job quite substantially before it changes what it pays for the job. A study with an early closing date also leaves room for a later round of hiring decisions. 73
An American payroll series raises a sharper question about entry. In their August 2026 working paper, Brynjolfsson, Bharat Chandar and Ruyu Chen compared young workers in occupations with different degrees of AI exposure using ADP data through June 2026. Employment among 22-to-25-year-olds in exposed work stood nineteen per cent below the level implied by keeping pace with their less-exposed peers. Reduced recruitment accounted for most of the divergence. An employer could contribute to that gap year after year without announcing a single redundancy. A graduating class would experience the change one unanswered application at a time. 74
The authors describe their result as a pattern requiring explanation. The estimates weaken with controls for education, some differences predate widespread generative AI use and the divergence is more pronounced in their payroll sample than in national surveys. Occupational exposure also leaves open whether a specific employer adopted any tool. A separate New York Fed analysis of US vacancies found relative weakness in postings for exposed occupations already visible before ChatGPT, with little evidence of a new divergence between junior and senior postings. A vacancy, a hire and an employee on payroll are three successive events. The studies observe different ones. 74 75
Three minutes in the bank
There is a test at the boundary between departments. A call centre can remove a minute from its record by sending a doubtful payment directly to investigations. If the investigator spends four minutes finding material omitted from the referral, the apparent improvement at the first desk has increased total labour. Match referrals to downstream records using the same case identifier, including files that cross a month end. Draw the comparison around the customer’s completed case. Departmental averages will otherwise reward the team that moves work across the reporting boundary.
Where recovered hours go
Follow three minutes recovered from each of 24,000 comparable cases. A funded temporary assignment and a cancellable overtime shift have different consequences at payday.
The destination of 1,200 hours
One bank · one budget yearEvery column accounts for the same cohort. Select a destination to see its share and its cash treatment.
160 hours of new checks, 50% of discretionary time to backlog, funded temporary cover and cancellable overtime.
| Against existing paid shifts | $1,000 higher |
|---|---|
| Against the funded plan | $7,000 lower |
Trace every hour and reproduce the cash comparison
| Destination | Ordinary repair | AI service | Total |
|---|---|---|---|
| Clear backlog | 240 | 250 | 490 |
| Other duties | 0 | 250 | 250 |
| Added checks | 0 | 160 | 160 |
| Avoid new hire | 0 | 200 | 200 |
| Cut overtime | 0 | 100 | 100 |
| Total | 240 | 960 | 1,200 |
Ordinary repair recovers 120 hours from each prior task. The remaining 600 assembly hours and 360 matching hours are attributable to the service in this scenario. The bank finances new checks with time from the same cohort. It cancels up to 100 previously paid overtime hours at $60, then fills up to 200 hours of approved temporary cover at $40 with regular staff. It divides remaining whole hours between older files and other duties. Unchanged salaries cancel in both comparisons. The $7,000 lab fee includes setup and support. Both routes assume the same case mix and acceptance after 30 days, staff with the necessary skills at the required times, and equivalent review of errors. The cash comparison values customer wait reductions and salaried redeployment at zero.
Give the bank a cohort small enough to audit. For a full budget year assume 24,000 payment files of comparable difficulty. The old process takes an extra three minutes per file to assemble evidence and match records. On an accepted file the evidence must still resolve to the original entries, with the same controls and the same thirty days in which a correction can reopen. Count work against a mature cohort, including returns. The three minutes yield 1,200 annual hours. They measure two steps in these files. The bank has many other reasons to employ the people who perform them.
Divide the old work into 720 hours spent assembling the evidence and 480 spent matching it. Put a competent manager in charge of the old system before offering the bank anything new. Assume a common case identifier can save 120 hours in each step. That releases 240 hours for the existing backlog with the incumbent process. The proposed service can take over the remaining work, freeing a further 960 hours before checking. Budget 160 of them for examining exceptions and maintaining the comparison files. The bank now has 800 hours whose destination can reasonably be attributed to its AI purchase.
Put the shifts that affect the cash budget first. Suppose one hundred hours of overtime would otherwise have been paid this year and the new rota cancels them. At an assumed marginal $60 per hour, the bank's wage bill falls by $6,000. The same transaction can take $6,000 of gross pay from people who previously worked those hours, if the whole amount was wages. Where the employer's rate includes payroll contributions, the workers' own reduction will be smaller. An annual salary divided by a standard number of hours gives an accounting rate. It rarely tells you what the next cancelled shift actually saves.
The next 200 hours belong to a temporary assignment the bank had budgeted at $40 an hour. Existing employees can cover it in their ordinary shifts. The bank can withdraw an approved requisition worth $8,000 provided the work and the budget really existed. No employee on last year's payroll has had $8,000 taken from their pay for that reason. Somebody who might have been hired loses a possible piece of work. The evidence a reporter would want here is a dated requisition for the right skills and season, followed by its withdrawal. A manager's recollection of intending to hire carries less weight.
Assign half the remaining 500 hours to other duties and half to the waiting queue. The queue now receives 490 hours in all, 240 from the simple repair and another 250 from AI. Customers might value those hours more than the entire contract price. To turn them into cash the bank would need a specific consequence, such as a service credit it would otherwise have paid. Value them at zero in the cash calculation while recording how long the oldest complaint has been open. A service team deserves a measure of what it delivers even when the customer never receives a separate bill for patience.
The supplier wants $7,000 for the year, including setup and support. Compared with the current paid schedule after ordinary repair, the bank spends $1,000 more once the fee is included. Against a funded hiring plan it spends $7,000 less. Lose the requisition and the second case for buying disappears. Increase checking time until it consumes all 960 incremental hours and the bank owes the fee while its ordinary backlog improvement survives. This is a plausible route from a wages budget to AI revenue. The transfer depends on an employment decision that a technology director acting alone may have no authority to make.
The person learning to sign
Removing preparation also removes a place where a junior employee used to encounter errors at relatively low cost. Some mistakes belong in training. A new payment officer can learn the difference between a genuine reversal and a second debit by searching for the initiating instruction, comparing the dates and explaining the discrepancy to a colleague. If every clean case is prepared automatically and every hard case goes straight to a senior colleague, where does the next senior colleague acquire that experience? The buyer drawing up the service specification gets to decide whether any of the recovered hours answer that question.
Anthropic ran an experiment with 52 mostly junior software engineers learning an unfamiliar Python library. Participants could use it to complete two features, then took a quiz on what they had just encountered. Those assigned AI assistance scored fifty per cent on average, compared with sixty-seven per cent for participants who coded without it. The difference in completion time was statistically insignificant. An immediate quiz on a single library has a short reach. It does, however, separate today’s submission from the participant’s ability to diagnose the next failure. A bank could design the same distinction into an assessment with anonymised historical files. 76
Training has an expense on the other side of the desk. Natalia Emanuel, Emma Harrington and Amanda Pallais examined engineers at one large firm around office closures and later returns to the office. Sitting near teammates increased coding feedback by 18.3 per cent. Less experienced colleagues gained most, while senior engineers wrote less code when near teammates. This was a study of proximity. Its relevance to an AI rollout is the accounting problem it exposes. A manager measuring the senior employee only by personal output could classify time spent developing somebody else's judgment as a loss. 77
Who gets the recovered afternoon
An organisation can also fill a newly available hour several times in expectation. Berkeley researcher Xingqi Maggie Ye described eight months of observation and interviews at a US technology company with about 200 employees. Workers used generative tools to attempt a broader range of tasks, and work could extend into more of the day. The study traces a mechanism within one company. It offers no population estimate of unpaid hours. For our bank it suggests a practical check. Ask the person who signed the file when they reviewed it, especially if the system records the draft as complete at the instant the assistant produced it. 78
Consider the worker on the closing shift. A queue can fall during the day while complicated returns accumulate as branches shut. Averages over all employees will look better even if one team loses control of its evenings. Compare hours by shift, seniority and case severity. Check reopening and customer complaints after the same interval. Record cancelled overtime against payslips and withdrawn hiring against approved budgets. Ask workers taking on the new checking whether the preparation guide gives them enough authority to send a doubtful file back. Somebody whose only option is to approve can generate an astonishing acceptance rate.
At renewal I would put the person who owns the rota next to the person who owns the software contract. Give them the gross minutes, the matched queue, the payment records and a list of responsibilities that changed hands. Ask which part of the supplier’s next invoice can be financed from a smaller cash bill, which part buys a better service and which part buys the ability to train people for tomorrow. The employee who gains an afternoon, the graduate who never receives an offer and the duty manager who takes the call on Saturday each know something the procurement system cannot tell us. The economics of AI computing depends on whether anyone asks them.