How to Measure If an AI Tool Saves Money
Share
Most AI savings are reported as impressions. Impressions are how organisations end up paying for twelve subscriptions nobody uses.
Direct answer
Time ten instances of one named task before you start and ten after, with the same person, and count the outputs that needed heavy rewriting. Compare the monthly saving against the monthly cost including checking and learning time.
Based on the AI and Automation Practitioner standard (ST1512), a Level 4 apprenticeship for people already in work. Funding rules described here are the England position and are set by the funding body.
Potentially free for you
The learner never pays. Employers potentially pay nothing either, and where the apprentice is aged 16 to 24 the government funds the whole cost even if the employer does not pay the levy.
Closing any gap. A levy transfer from another employer covers 100% of the cost. Unused transfer allowance is not carried forward at the end of the tax year, and it must be agreed before the apprentice starts.
From 1 August 2026 the government stopped adding the 10% top-up to new funds entering levy accounts, so there is less in the system than there was.
Part of our guide to AI for UK Business: Where It Pays and Where It Does Not.
The measurement, in full
Pick one named task. Not a role, not a department, one repeating task with a clear start and end.
Time ten instances before you change anything. Rough minutes are enough. Precision matters less than the number existing.
Change one thing, then time ten more with the same person doing them.
Count separately how many of the ten needed heavy rewriting. That is your real quality cost and it is where reported savings usually evaporate.
The numbers that mislead
Self-reported time savings. People consistently overstate them, particularly when they advocated for the tool.
Licence usage statistics. Opening a tool is not using it, and vendors report the first as if it were the second.
Output volume. Producing more drafts is not the same as producing more finished work, and it can create downstream checking cost.
Satisfaction scores. People enjoy new tools. Enjoyment and saving are unrelated and only one of them is on the invoice.
Costs to include on the other side
The subscription, obviously, but per person actually using it rather than per licence bought.
Checking time. If verification went up, that is a cost and it belongs in the calculation.
Learning time, once. Six hours to learn a tool has to be earned back before anything else counts.
Support time. Somebody answers the questions, and in small organisations that somebody is usually already busy.
There is more on this in our guide to AI for UK businesses.
When the saving does not appear
Wrong task. This is the commonest cause and it looks like a tool failure. Try a second task before concluding anything.
No habit change. People kept working the old way and used the tool occasionally. This is a training issue, not a tooling one.
Checking cost ate the saving. Real, and it means the task needed too much verification to be a good candidate.
The bottleneck was elsewhere. If quotes take three days because one person approves them, a faster draft changes nothing.
A worked example
A ten-person firm times first-line email replies. Ten instances take 118 minutes, so roughly 12 minutes each.
After thirty days with a drafting prompt, ten instances take 61 minutes. Two of the ten needed heavy rewriting.
That is a 48% raw saving, reduced to roughly 40% once the rewrite time is added back. Across 40 replies a week it is about three hours.
Three hours a week against a subscription of a few pounds a month is a clear result. The number is what made it clear, not the impression.
Costing it properly in a UK business
Price the time at a fully loaded rate including employer National Insurance and pension, not at the salary rate. It changes the answer.
Check whether the subscription is quoted in sterling and whether VAT is included, because both affect the comparison.
If you pay the apprenticeship levy, note that capability funded from the levy does not compete with the software budget.
Present the saving in hours as well as pounds. Hours are what a manager can act on and pounds are what finance needs.
Reporting it honestly
Give the number with the rewrite rate alongside. A 40% saving with a 10% rewrite rate is a good result and worth saying plainly.
Report what did not work as well. A report with no failures reads as marketing and gets discounted accordingly.
Keep it to one page. Anyone deciding on a renewal needs the number, not the journey.
Repeat the measurement in six months. Savings decay when habits slip, and nobody notices without a second measurement.
Key takeaways
- Ten instances before, ten after, same person, one named task.
- Count heavy rewrites separately, because that is where savings evaporate.
- Include checking time, one-off learning time and support time as costs.
- Self-reported savings and licence usage statistics both mislead.
- Repeat the measurement in six months, because savings decay.
Frequently asked questions
How do I measure AI return on investment?
Time ten instances of one named task before and after with the same person, count heavy rewrites separately, and compare against the full monthly cost.
Why do reported AI savings look so large?
They are usually self-reported and exclude checking time. People overstate savings, particularly when they advocated for the tool.
What costs should I include?
The subscription per active user, additional checking time, one-off learning time, and internal support time.
What if there is no saving?
Usually the wrong task rather than the wrong tool. Also common is no habit change, which is a training issue rather than a tooling one.
Is licence usage a good measure?
No. Opening a tool is not using it, and vendors report the first as if it were the second.
How often should I measure?
At the start, at thirty days, and again at six months. Savings decay when habits slip and nobody notices without the second measurement.
What is a good result?
A 20% to 50% time saving on a named task with a low rewrite rate. Anything above that on a first attempt is worth double-checking.
How should I report it?
One page, the number, the rewrite rate, and what did not work. Reports with no failures get discounted by anyone experienced.
The funded route, in figures
Figures are the England position. Levy-paying employers usually draw the whole cost from the levy they already contribute, so they pay nothing extra. Employers below the £3 million pay bill threshold pay nothing for an apprentice aged 16 to 24. For an older apprentice the government funds 95 per cent, and a levy transfer from a larger employer can cover the rest. Eligibility and funding rules are set by the funding body and can change.
Where to go next
If you want measurement and improvement work assessed towards a qualification, the funded route uses exactly this kind of evidence.
We help UK employers and employees access government-funded AI training. Training is delivered by third-party providers. We do not guarantee a place, a qualification, a job or any particular outcome. Eligibility and funding rules are set by the funding body and can change. Always check the current position before you commit. See our disclaimer.