I used to have a mild preoccupation, as a civil servant, with working out what meetings cost. It’s the kind of thing you start thinking about during your eighth meeting of the day, while wondering why you were invited to another meeting that could have been an email.
We know this, but we’re bad at remembering it: twelve people around a table for an hour comes at a significant cost. There are salaries, pensions and National Insurance contributions, preparation time, travel and the cost of the tired but expensively located building we are obliged to spend the day in. Then there is the work people were not doing while they discussed the new work that may or may not be needed.
We could unwittingly spend almost as much debating whether to commission a piece of research as the research itself would have cost. But the research was visible expenditure: we’d raise a purchase order and give it a line on the budget. Our time was different. The first attracted scrutiny; the second slipped past almost unnoticed.
This matters when organisations start making the business case for AI.
The first wave of workplace AI has mostly been cheap and personal. Someone uses a general-purpose chatbot to improve an email, adapt content for another audience or escape the tyranny of the blank page. A small charity creates a simple retrieval-augmented generation — or RAG — tool that can search previous bids and approved material. Staff feel the difference immediately, even if the saving never appears in the accounts.
Then the organisation wants to move beyond experimentation. The next tool needs licences, integration, information assurance, training or specialist development. Perhaps it will work with sensitive internal data. Suddenly, someone reasonably asks: what are we getting for the money?
That is when enthusiasm has to become a business case. Unfortunately, lots of organisations are not very good at writing business cases, even before AI gets involved.
Time is money
A sensible AI case starts by recognising that people’s time has a value. That does not mean pretending every ten minutes saved can be removed from next year’s staffing budget. It means acknowledging that human capacity is finite: using it for one thing prevents it being used for another.
The UK Treasury’s Green Book defines opportunity cost as the value of the next best use of a resource. If a pointless meeting prevents the same people from attending a genuinely useful one, the cost is not just an abstract total on a spreadsheet; it is also the useful work that did not happen.
For AI, a credible cost estimate should also allow for the proportion of staff who will really use the tool, the time spent training and supporting them to release those productivity gains, and the fact that some saved minutes will be too fragmented to reuse.
Then there is the distinction between extra capacity and an actual reduction in spending. Finance guidance separates cash-releasing savings, which directly reduce spending, from non-cash-releasing gains, where the same staff can do more or better work but the budget does not fall.
If an AI assistant gives a charity fundraiser three hours back, that has value. It might allow another bid to be submitted, improve oversight of funders or reduce out-of-hours work. It might help manage a rising workload that was going to bite next year but had not yet appeared as a cost. But unless the charity avoids recruitment, reduces external spending or changes a budget, it is not cash in the bank. You might see the impact years later in staff retention or fundraising trends.
Measure the work
The wrong place to begin is: “We would like to buy an AI tool.” Begin instead with: “This task currently takes this long, costs this much and produces these results. Here is what we think can improve.” Establish a baseline before the pilot. Depending on the task, measure:
- Time to an acceptable finished output.
- Direct staff time, including checking and corrections.
- Errors, rework or rejection.
- Volume and cost per completed output.
- Quality and user satisfaction.
“Acceptable finished output” matters. Measuring only the speed of the first draft rewards tools that generate plausible material quickly and leave a human to spend an hour repairing it.
A controlled study of professional writing tasks found that ChatGPT users completed work 40% faster and achieved quality scores 18% higher. But the researchers warned that the tasks did not require the organisational context and factual precision often needed in real work. Meanwhile, the UK government’s large Microsoft 365 Copilot trial reported an average self-assessed saving of 26 minutes a day. A more cautious HMRC evaluation, using a randomised cohort and stricter adjustments, found about 60 minutes a week. That does not mean one result is right and the other is wrong. It means the answer depends on the users, tasks and methods — precisely why your organisation needs its own baseline.
At this point, we have to compare AI with AI
There is another trap to watch out for.
A team tests a specialist product against doing everything manually, finds a time saving and concludes that the product represents good value. A couple of years ago, that might have been enough. But “better than no AI” is no longer the only relevant comparison. The options might now be:
- Continue without AI.
- Use a general enterprise chatbot with good prompts.
- Configure a standard GPT or lightweight RAG tool.
- Buy a well-marketed specialist product.
- Build something bespoke in-house.
If you were paying attention at the beginning, you’ll probably guess that option five is the one whose cost you are most likely to underestimate. Building something yourself requires you to be very honest about the value of your own teams’ time. Just saying.
The specialist tool you saw at a trade fair no longer needs to beat the manual process; it needs to create enough additional value over the cheaper AI option to justify its extra cost and complexity. Yes, it’s nice that suppliers have wrapped general models in specialist data, and workflows. But does producing something that sounds more like a lawyer necessarily justify the premium? Does it perform equally well in all the languages your organisation needs?
Our advice: do not rely on a beauty parade of supplier demonstrations. Test a few possible options on the same tasks, using the same material and definition of a satisfactory result. Compare time, quality, user experience and whole-life cost. For higher-risk uses, add privacy, security and failure testing. Be wary of anyone who will not let you test the tool on realistic tasks in real time.
What decision-makers need
Your boss or finance team does not want a hymn to the transformative power of AI — unless they’re the one singing it. They need a short, testable proposition:
- The problem: What is slow, costly, inconsistent or risky now?
- The baseline: What time, cost, volume and quality do we see today?
- The options: What are the credible non-AI, general AI, specialist and bespoke choices?
- The evidence: What happened when they were tested on comparable tasks?
- The benefits: Which save cash, create capacity or improve quality and outcomes?
- The costs: What is the whole-life cost, including people, data, integration and assurance?
- The plan: Who owns adoption and benefits, and when will we stop, adapt or scale?
Build in some scepticism too. All business cases are prone to optimism: benefits expand, costs shrink and implementation mysteriously gets quicker. Show a cautious, expected and optimistic scenario rather than one heroic return-on-investment figure.
The meeting test
The irony is that some leaders who are least convinced by staff time savings have their own time carefully protected. Assistants arrange meetings, find documents, summarise correspondence and turn rough thoughts into presentable prose. The value of administrative help is already built into their working lives; it just arrives in human form, long-since buried in their core staffing budget.
AI can extend some of that administrative assistance to people who have never had it.
That doesn’t mean every AI product deserves investment. Some general tools perform almost as well as expensive specialist ones; others generate more quality-assurance work. The answer is neither faith nor cynicism; it is measurement and honesty.
Count people’s time even when it is hidden. Compare the new tool with the best realistic alternative. Separate money saved from capacity created.
Because an hour in twelve diaries is still twelve hours. It didn’t become free just because nobody submitted an invoice.
