Policy WindowAI, evidence & public choice
Menu
Human by choice
Policy Window advocacyWiki · neutral reference ↗

When AI saves time, who gets it?

If a service uses AI to finish administration sooner, people might get longer appointments, workers might get a lighter week, or the organisation might buy less paid labour. None follows automatically. Employers and funders make choices, within constraints set by schedules, contracts and money.

A money value attached to staff time can describe capacity that is already being paid for. It does not establish a reduction in spending, or show that the hours are available when people need them. Ask who controls the time, who pays for the technology and whose interests count.

HM Treasury's Government Efficiency Framework separates direct spending reductions from benefits delivered with unchanged staffing and budgets. Its rules require efficiency savings to be net of costs, without worse performance or outcomes. These are central-government reporting categories, not proof that another organisation can release money. They help expose a distinction a service's own plan must resolve. [1]

A service that exists only on paper

Imagine a community service providing 1,000 appointments each year. Every number here is an invented assumption, not a market price, pilot result, estimate of AI performance or advice for a real service.

A manual appointment takes 30 paid minutes: 20 in direct conversation and ten in administration. The AI-assisted version takes 25: the same 20 minutes of conversation and five for checking and administration. The five-minute reduction must be net of required checking and corrections. Keeping conversation time constant does not establish equal quality or access.

Assign paid time a value of £40 an hour. Software costs £2 per appointment; setup costs £2,000 once, in year one. Hold volume and prices constant for three years, with no discounting or residual value. Other costs would need identifying before a real decision; this illustration does not claim they are zero.

The manual workload is 500 paid hours a year, assigned a £20,000 cost. The assisted tasks take 416⅔ hours, potentially releasing 83⅓ hours. These are annual totals, not a promise that convenient blocks of time appear in a rota.

Two ways to use the same five minutes

A: keep the staff budget fixed. The service retains 500 paid hours worth £20,000 each year. Salary cash savings are zero. Technology adds £4,000 in the first year and £2,000 in each later year: £8,000 extra spending across three years.

If all the released time could genuinely be redeployed, each appointment could contain 25 minutes of conversation and five of checking and administration within the original 30-minute block. That is an allocation of time, not evidence of better quality, more appointments or shorter waits. Scattered minutes, coverage needs or extra work could prevent it. More appointments would also change software costs and are outside this model.

Workers could gain lighter workloads or shorter hours on unchanged pay; services could offer follow-up. These choices need a plan. Retained pay remains expenditure.

B: actually reduce paid costs. Now assume the full £40 for every released hour is genuinely avoidable spending. The service buys only 416⅔ paid hours, costing £16,666⅔ annually. It saves £3,333⅓ in staff cash each year before technology costs.

That assumption is demanding. A calculated hourly cost may include premises, management and other overheads that remain. Existing salaries may stay unchanged. Avoiding overtime or not filling a vacancy might reduce spending; reducing paid hours might reduce someone's income. None is assumed feasible, lawful or acceptable here. Contract terms and the consequences for workers need examination.

The two cases have different cash results:

Period A: extra cash needed, staff budget fixed B: staff cash saved, if avoidable B: net cash result after technology
Year 1 £4,000 £3,333 £667 extra cost
Year 2 £2,000 £3,333 £1,333 saving
Year 3 £2,000 £3,333 £1,333 saving
Three-year total £8,000 £10,000 £2,000 saving

Amounts are rounded to the nearest pound; totals use the unrounded calculations. Both cases include the same software and setup costs.

In B, the exact first-year result is £666⅔ of extra cost, followed by £1,333⅓ saved in each later year. The three-year saving is £2,000. Dividing that total into an annual average would conceal the first-year funding gap. Spreading setup over accounting periods would not change when this model pays it.

Under B's avoidable-cost assumptions, the first year needs six net minutes saved per appointment to break even. A recurring year needs three; the three-year average needs four. The assumed five minutes sits only one minute above that last threshold. At two minutes, the three-year result becomes £4,000 extra cost; at eight, £8,000 saved. Those are recalculations, not performance predictions.

Measure what happened to the time

The government's Microsoft 365 Copilot trial illustrates the measurement gap. Its time estimates were self-reported and calculated from response bands. The report could not identify how saved time was spent. Those findings should not be treated as measured payroll savings, or as a productivity rate for our example. [2]

A local test should examine the whole task, including corrections, checking and work shifted to colleagues or service users. Record paid hours and actual spending separately. Check errors, follow-up work, conversation time, waiting and access for different users, alongside workload and workers' experience. A faster administrative step can coexist with a worse service.

Extra quality could also justify paying more. A tool might be worth buying without any time saving. Conversely, preserving 20 minutes of conversation cannot compensate automatically for inaccurate records or barriers to access. The decision needs evidence about the intended benefit, not only a favourable conversion from minutes into pounds.

Make the allocation a decision

Human by choice proposes a practical rule: before adopting an AI change, the responsible budget holder or commissioner should state how any released time and money are intended to be used. A commissioner is the body buying or funding the service. This statement should sit in the existing business case or decision record.

Name who can alter rotas, staffing budgets and service standards. Say who funds software, setup and evaluation, and whether costs fall on a different budget from the expected benefit. Identify the intended beneficiaries: workers through pay or hours, service users through access or attention, or the organisation through lower expenditure. Do not promise the same minutes to everyone.

An employer could adopt this through its management or governing body. A commissioner could propose proportionate reporting through its commissioning process, subject to its actual contractual powers and lawful authority. This is a proposal for organisational practice, not a claim that a universal duty already exists. Workers and service users should help judge the options; a manager's statement does not create their consent.

Compare the strongest feasible alternatives, including improvements to the manual process, staffing or training without buying the tool. Ask what the same money could achieve that way. A pilot should specify a review point and what would trigger redesign or withdrawal: for example, net time gains below the budget case, errors rising, or workload simply moving elsewhere. Failure of the cash case still leaves room to defend a worthwhile service benefit openly.

There is a cost to this rule: recording and checking gains takes staff time. Small organisations should not need a large new reporting system. Use existing records where they answer the question, budget for necessary evaluation, and publish only appropriate aggregate information. If credible records already reveal the allocation and consequences, another form adds little. Evidence of that would weaken the case for an additional statement.

The aim is to make economic power visible. Staff time, money and authority are different things. A useful AI change may create room to improve work or services; deciding who receives that room requires a choice people can see and challenge.

Human by choice is Policy Window's advocacy programme on AI, work and who decides. This article is analysis and advocacy, distinct from the neutral Policy Window wiki.

Sources

[1] HM Treasury: Government Efficiency Framework. Updated 24 November 2025; sections 4.1-4.5.

[2] Government Digital Service: Copilot experiment findings. Published 2 June 2025; methods, time estimates and conclusions.

Where will the time go?

Proposed allocation statement

Human by choice · Policy Window advocacy

Use this with workers, service users and the budget holder to record who receives released time and who pays. Put it in the existing decision record; link evidence already available. This proposes organisational practice, not a legal duty, compliance certificate or existing reporting requirement.

Keep answers brief. Mark each estimate, observation and unknown clearly.

1. Decision and responsibility

Service and exact workflow: ___. Proposed change: ___. Decision owner and date: ___. Who controls the budget, rota and service standards? Record their actual powers and any further permission needed.

2. Starting point

For which period and workload? Record paid hours, staffing, spending and how work is currently done. What do existing records show about quality, errors, waiting, access and workload? Identify gaps in this baseline.

3. Time genuinely available

Compare the whole task before and after, including checking, corrections, follow-up and work moved to colleagues or users. What time remains available, to whom and when? Separate observed changes from assumptions; identify scattered minutes that cannot be reassigned.

4. Cash, costs and payer

Which spending could actually fall, when and through what action? Keep this separate from capacity retained on the same payroll. Name who pays for setup, software, training, checking and evaluation. Separate one-off and recurring cash payments from already-funded staff time; identify missing costs and overheads that remain.

5. Allocation and feasibility

Who receives the available time through attention, access, follow-up, lighter workload or shorter hours? What happens to pay? Who loses income, service or capacity, or takes extra work? Allocate each saving once. State the rota, contract, coverage and operational assumptions, and who can implement them.

6. Worker and user input

Who was asked, how and about which options? Record concerns, disagreement, missing voices and the response. Distinguish participation from agreement; give people a way to challenge the account. Leave consent unclaimed where it has not been established.

7. Best feasible alternative

Compare the strongest feasible options, including a better manual process, staffing or training. What could the same resources achieve? Compare costs, burdens and intended benefits. Why choose this option?

8. Review and decision status

Status: proposed / testing / decided / changed / withdrawn. Record reasons, remaining uncertainty and whose decision this is. Name actual measures for time, cash, quality, workload and access, their evidence source and recorder. Set a review date and thresholds for changing or withdrawing the plan. What result would overturn the case?

Related illustration: the appointment model.