AI saved time. What changed in the budget?

A practical distinction between released capacity, better service and money the business can actually avoid spending.

A finance director receives an artificial intelligence (AI) business case with an attractive calculation: minutes saved, multiplied by transactions, multiplied by an hourly rate.

The operations director recognises the improvement. People can finish part of the work faster. Yet the team still has the same employment costs, customers still need support, and the new system adds expenses of its own.

Both observations can be true. The task has improved. The claimed financial saving has not yet been demonstrated.

In my previous Artificial Mind Insights edition, I examined who owns the workflow after a successful pilot. This edition takes the next decision: what kind of value can that owner credibly commit to?

Name the benefit before pricing it

Released capacity gives people time they may be able to use differently. A better service outcome means something has improved for the recipient of the work. A cash saving requires expenditure to fall against a credible comparison.

These benefits can overlap, but they should not be added together automatically. The same hour cannot both disappear from paid staffing and remain available to deliver more work.

Nor should every improvement be forced into a cost-reduction story. More time for difficult customer cases, more reliable decisions or less sustained overload can be worthwhile objectives. Each needs evidence appropriate to the benefit being claimed.

The management choice is to name the intended benefit honestly, then determine how it will be realised.

What the research can support

Documented public research. TNO's June 2026 account of research in four organisations distinguishes time savings from higher productivity or revenue. What happens to the released time matters. Its findings also show that changes in tasks can affect workload and the quality of work. Four cases provide useful examples, not a representative estimate of what every organisation will achieve. [1]

There is credible evidence of positive operational effects. Brynjolfsson, Li and Raymond's study, published in the May 2025 issue of The Quarterly Journal of Economics, examined the introduction of an AI assistant among 5,172 customer-support agents. It reported a 15% average increase in issues resolved per hour, with substantial differences between workers. That is evidence about a particular operational measure in a particular setting, not a transferable promise of a 15% budget reduction. [2]

Evidence also needs updating. In February 2026, METR explained that selection effects complicated its follow-up research on developer productivity. Its earlier finding of slower completion with early-2025 tools should not be presented as a general verdict on current AI tools. The follow-up did not establish a clean, universally applicable replacement percentage either. [3]

My conclusion is practical: use external research to identify what to investigate, then base the commitment on the work and costs that can actually be observed in your organisation.

An illustrative budget conversation

Illustrative example. All figures below are invented assumptions, not a client result or a forecast.

A service team handles 1,000 comparable requests each month. Assume the current process takes 12 minutes of staff work per request. That is 200 staff hours.

Before adding AI, the team improves its intake form and removes unnecessary re-entry. Suppose that brings the complete handling time to 10 minutes. Monthly staff effort would then be about 167 hours.

An AI-assisted version is expected to reduce it to eight minutes, including human checking and normal correction work. Assume only 60% of requests are eligible and actually handled that way. The remaining 40% stay on the improved ten-minute process.

That gives 600 requests at eight minutes and 400 at ten minutes: about 147 staff hours. Compared with the improved process without AI, the incremental reduction is 20 hours per month. The wider reduction from the original process is about 53 hours. Attributing all 53 hours to AI would overstate its contribution.

At an assumed loaded labour rate of €50 per hour, those incremental 20 hours represent €1,000 of capacity value. They are not automatically €1,000 available to remove from the budget.

If salaries and staffing stay unchanged, the direct payroll saving in this example is zero. The team may still create value by improving service or absorbing demand, but it must show what happened to that capacity.

Suppose instead that the team can genuinely avoid 20 hours of paid external support at €50 per hour without reducing quality or shifting work elsewhere. That would represent €1,000 of avoided expenditure. With assumed incremental AI operating costs of €600 per month, the ongoing net benefit would be €400 before one-off implementation and transition costs. Those costs and the period over which the benefit persists still matter.

The operations lead owns the service decision. Staff help establish whether the work remains manageable. Finance verifies whether an expense really disappears. The technology team supplies the full incremental cost of operating the capability.

Make room for a result that disappoints

The arithmetic is only as useful as its assumptions.

If just 30% of requests use the AI-assisted process, the same two-minute improvement releases ten hours. Even if every hour genuinely avoids €50 of external expenditure, the €500 avoided cost would be below the assumed €600 monthly operating cost, before implementation costs.

If checking and correction bring the AI-assisted handling time back to ten minutes, the incremental time advantage disappears. The tool might still improve quality, but that would be a different benefit requiring different evidence.

Request complexity, demand, staff experience and changes elsewhere can also explain an apparent improvement. Comparing an unusually difficult month with a quieter one does not isolate an AI effect.

For a real service, use sufficiently comparable work and include the time spent correcting mistakes, dealing with exceptions and supporting the tool. Record whether customers return with unresolved issues. Keep sensitive data within approved systems, and retain human authority over consequential responses and exceptions.

Do not make staffing commitments from a short pilot average while quality, demand and sustained adoption remain uncertain.

Bring one claim to the next review

Choose one material benefit already claimed in an AI proposal. Ask its owner to bring the observed baseline, the realistic improvement without AI and the evidence for the additional AI contribution.

Then ask one financial question: which expenditure would actually change?

If the answer is none, the proposal may still deserve support. State the intended service, capacity or employee outcome and judge it on that basis. If the answer identifies an avoidable cost, test the mechanism and include the costs required to sustain it.

The useful decision is whether to invest, adapt or stop on the basis of a benefit the organisation can explain and verify.

If your next investment decision rests on a time-saving claim, I am available to discuss a focused review of the business case, the work involved and the evidence still needed. Any further assessment would have an agreed scope and deliverable.

Milo Izzinosa — Founder & AI Transformation Lead, Artificial Mind Solutions; Founder & Enterprise Transformation & Execution Lead, eXcelleRes.

Sources and limitations

1. TNO, 10 June 2026 — AI biedt veel kansen voor productiviteitswinst, maar gevolgen voor werknemers vragen om aandacht. Four organisational cases plus literature; context-specific evidence, not a general savings benchmark.
2. Brynjolfsson, Li and Raymond — Generative AI at Work, The Quarterly Journal of Economics, May 2025. Study of the staggered introduction of AI assistance; operational productivity is not identical to cash savings.
3. METR, 24 February 2026 — We are Changing our Developer Productivity Experiment Design. Primary research update describing selection and measurement limitations.

Sources checked on 15 October 2026. The numerical service example is hypothetical and independent of these studies. It illustrates the distinction between benefit types; it is not a return-on-investment forecast, an accounting determination or evidence of delivered client results.

Originally published on LinkedIn on 15 October 2026 in Artificial Mind Insights. Read the original LinkedIn article.

Find out more about AI ValueForge or discuss a focused business value assessment.