If AI saves two days, who gets the time back?

Six months after Hantam Service Cooperative, a fictional shared-services cooperative in the Northern Cape, gave its cloud financial operations analyst an AI assistant and a cleaner cost dataset, her manager noticed something in the time records. The monthly cost allocation that used to take most of a week was finished by Wednesday. The member reports were drafted before the review meeting rather than after it. "She has two spare days," he said. "Let's give her the onboarding project."

Nobody in the room asked whether that was the right use of two days, or whether the two days were real. The default had already been chosen: the gain would become more work, because that is what usually happens to freed time when nobody decides otherwise.

This series is about making that decision on purpose. If AI genuinely lets a person deliver the same accepted outcomes in less time, the gain has to go somewhere. It can become more output, better service, lower cost, stronger resilience, fewer working hours or wider access to work for people who cannot or will not sell five days a week. AI does not choose between these. Leaders do, and a credible choice starts with evidence that the gain exists.

First, check that the two days exist

Faster emails are not a business case traced the five links between a minute saved on a task and a change in an organisation's results, and why most business cases assume all five hold. This article starts one step later. Suppose the saving is real at the level of a role, not just a task. Who should receive it?

Before answering, separate four things that are easy to blur:

• Task acceleration and role compression. Drafting a report faster is a task change. A role that delivers the same accepted outcomes in fewer hours, month after month, is a different and much rarer claim.
• A demonstration and sustained production evidence. A good week is not a baseline. Month-end, audit season and incident weeks have to be inside the measurement.
• Effort and elapsed time. An analyst may spend fewer hours on the allocation and still wait three days for a member to confirm a cost centre. Queueing, coordination and accountability do not compress with typing speed.
• Individual and system throughput. One person finishing early helps only if the work they feed is not stuck somewhere else.

The research does not support a general claim that AI turns five-day roles into three-day roles. A large field study of customer-support agents by Brynjolfsson, Li and Raymond found that AI assistance raised issues resolved per hour by about 15% on average in the version published in the Quarterly Journal of Economics in 2025 (14% in the 2023 working paper). The largest gains went to less experienced and lower-skilled agents, while the most experienced saw small gains in speed and small declines in quality. Gains are real in some work and uneven across people. And at the level of whole labour markets, Humlum and Vestergaard's study of Danish administrative records (NBER working paper 33777, revised March 2026) found precise null effects on earnings and recorded hours two years after ChatGPT's launch, while workers reported productivity benefits and employers reorganised tasks. Freed time is not automatically becoming shorter hours.

What Hantam actually measured

Hantam's head of shared services, Marius, did not accept "two spare days" as a figure. He asked the analyst, Naledi, to keep a simple work log for six weeks, including the month-end close, and compared it with the time records from the same period the previous year. He deliberately did not use the assistant's usage statistics, for the reason given in AI consumption is not AI value: prompts sent say nothing about work completed. All figures here are illustrative.

The role had shed about 13 hours a week, roughly a day and a half, not two days. And only part of that was AI. Three hours came from stopping a duplicate report, which needed a decision, not a model. The rest of the analyst's week, the judgement-heavy work, had barely moved.

The tasks that compress are drafting, summarising and reformatting. The tasks that do not are judging, deciding, persuading and being available when something goes wrong. A role is a bundle of both, and the second kind often sets the minimum.

Five places a verified gain can go

Once a gain has survived measurement, there are five broad destinations for it. Each has a beneficiary and a cost.

1. Customer value. Faster or more service for the same cost. At Hantam, member cost reports could be published earlier in the month, when they are still useful for decisions.
2. Quality and resilience. Time spent checking, documenting and covering for absence. This is the least visible destination and often the most valuable.
3. Organisational economics. Absorbing growth without a new hire, reducing contractor spend or, in some cases, reducing headcount.
4. Employee time. Fewer working hours, with pay treated explicitly rather than by accident.
5. Wider talent access. Roles designed for fewer days, opening work to capable people the five-day pattern excludes. Part 5 of this series takes this up.

The destinations can be combined, but every allocation is a trade-off. A gain spent on more output is not available for fewer hours. A gain spent on lower cost is not available for better service. When nobody makes the choice, it is made by default, and the default is almost always the first or third destination, through the quiet refilling of time with more work.

Four-day-week evidence is not AI evidence

The idea that productivity gains could pay for shorter working time is not new, and there is a body of trial evidence about it. It needs to be read for what it measured.

The UK pilot run by Autonomy, 4 Day Week Global and researchers at Boston College and the University of Cambridge involved 61 companies and about 2,900 workers from June to December 2022. Its full report required pay to stay at 100% with a "meaningful" reduction in working time, and each company designed its own model after two months of preparation. Of the 61 companies, 56 were continuing at the end and 18 described the change as permanent. Revenue stayed broadly the same, rising 1.4% on a size-weighted average across the 23 organisations that supplied enough data.

What it did not measure matters as much. The report explains that there was no common productivity metric, because the companies collected different data. Participation was voluntary. Much of the evidence is employee survey data. And it was a working-time trial, not an AI trial: companies prepared with workshops, coaching and peer support and designed their own working patterns. Some reformed email habits, introduced focus time or automated parts of their work, but the trial did not isolate the effect of any tool, and it ran almost entirely before generative AI assistants were in general use. A six-month pilot with mainly South African companies, which 4 Day Week Global labels its 2023 pilot, has narrower evidence still. Its results come from an organiser press release, not a peer-reviewed report: 470 employees, findings from 22 of the 28 companies that signed up, and 57% of companies with ten or fewer employees.

The honest reading is narrow. These trials show that some organisations can redesign work to reduce hours without visible loss, when they prepare deliberately and choose to. They do not show that AI produces the gain, and they do not show that every role or organisation can do it.

The allocation is a governance decision

If the destination of a gain is a choice, it needs the things any significant choice needs: an owner, evidence, a record and a review date. Hantam now treats a verified role-level gain the way it treats a benefit in an approved business case, along the lines described in The business case was approved. Who is tracking the benefits?.

The sequence has four gates:

1. Measure the role, not the task. A baseline across a full cycle, including the worst weeks.
2. Check what did not compress. Coverage duties, accountabilities and fixed deadlines that remain regardless of speed.
3. Decide the allocation with the people affected. A named decision-maker, the employee concerned and, where relevant, their representative. Changing someone's working time or pay is a change to their terms of employment, which generally needs the employee's agreement; take advice before changing terms.
4. Review on a date. A pilot with exit criteria, not a permanent change on day one.

The decision record does not need to be long. It states the verified gain, the destinations chosen and in what proportion, who gains and who bears the cost, the measures that will show whether it is working, and the date it will be reviewed. Organisations that already keep a benefits ledger can record it there, so that the same hours are not also claimed by another business case.

What Hantam decided

Marius took the measured day and a half, plus the work the team agreed to move elsewhere, to a meeting with Naledi and the finance director. The outcome, recorded in a one-page decision, split the gain deliberately.

• Members get their monthly cost reports on the fifth working day instead of the ninth.
• Resilience improves: anomaly alerts are now reviewed daily by a rota that includes a platform engineer, so cost spikes are not left until the analyst's next working day.
• Naledi, who asked to work fewer days for her own reasons, moves to a three-day role for a six-month pilot, by written agreement. Her pay is set at 75% of the full-time salary for 60% of the days. Pro rata would have been 60%; the difference is her cash share of the verified gain.
• The cooperative keeps the remaining salary difference, after the cost of the rota, and a commitment to review the arrangement against measures, not impressions.

The onboarding project went to a fixed-term contractor, explicitly funded, rather than into the gap.

Some will pay full salary for reduced hours, as the UK pilot companies did. Some will put all of the gain into service. What makes Hantam's decision credible is not the split. It is that the split was visible, agreed and measurable.

A decision table for the allocation

A note on law. In South Africa, reduced-time arrangements touch the Basic Conditions of Employment Act, the Employment Equity Act and the Labour Relations Act, including section 198C on part-time employees earning below the earnings threshold, which was inserted by the Labour Relations Amendment Act 6 of 2014. Part 2 sets out what these provisions say in outline. This series is general information, not legal advice. Any change to working time, pay or benefits should be reviewed with qualified employment-law and HR advisers.

Key points

• A productivity gain is an allocation decision. AI does not decide who receives it; leaders do.
• Measure the role across a full cycle before claiming days saved. Tasks compress; judgement, coverage and accountability often do not.
• A verified gain can go to customer value, quality and resilience, organisational economics, employee time or wider talent access. Each has a beneficiary and a cost.
• Four-day-week trials tested working-time redesign with full pay, not AI. Read them for what they measured.
• Record the allocation with an owner, the employee's agreement, measures and a review date. The silent default is more work.

Next: Part 2, A three-day role is not a five-day job compressed. All parts are on the series page.

How CloudNala can help

CloudNala helps leaders turn AI productivity claims into decisions they can defend. We measure the role-level effect of AI on a real team across a full work cycle, separate the AI gain from process changes, and facilitate a recorded allocation decision with the people affected, including the measures, review date and pilot exit criteria. Where working time, pay or benefits change, we work alongside your HR and employment-law advisers rather than in place of them.

Work with CloudNala

CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.

Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.

Talk to us about AI and work design or write to us at consult@cloudnala.co.za