Most of the AI dashboards we're shown measure the same thing: logins. Active users, prompts per week, licences assigned. Then someone senior asks what the business got for the money, and the room goes quiet.
That question comes up in almost every conversation we have about AI adoption right now. The CEO of a small investment manager told us their Copilot rollout had no metrics for efficiency, only anecdotal gains. A CIO at a family-owned wholesaler had Copilot running across the business and no way to tie it to return. A customer success director at a SaaS company had a CFO pushing back on any spend that came without before-and-after data.
They're far from alone.
The answer doesn't need a complex model. It needs a clear picture of a handful of workflows before AI changes them, an agreed view of what good looks like for each, and the discipline to measure them again afterwards. We've found it helps to think of it as a chain of receipts.
The chain of receipts
A head of transformation in the NHS put it well in a recent conversation. Knowing how much time AI saved wasn't enough for him. He wanted to see where that time went, with a receipt at every step.
That idea turns into a four-link framework, applied one workflow at a time:
- 01
Baseline. What the workflow looks like before anything changes.
- How long does it take?
- Who's involved?
- How is quality judged?
- 02
Behaviour. How the work actually gets done today.
- Which tools are used?
- Is AI already part of it?
- What checks and tacit knowledge does it rely on?
- 03
Benefit. What a good output looks like once AI is in the work.
- Time saved?
- Improved quality?
- Increased revenue?
- 04
Business. How the workflow's gain moves the metrics the business already runs on.
- Cost is the floor
- Quality, customer satisfaction and revenue are the goal
Each link needs evidence from the one before it. Skip a link and the chain breaks, which is usually the moment ROI turns back into anecdote. And the unit that matters throughout is the workflow, not the programme.
Baseline: map each workflow before anything changes
Without a "before" there can be no "after". It sounds obvious, yet most rollouts start with licences and a launch session and only think about measurement when the renewal lands.
The baseline that matters is set per workflow, not across the organisation. Pick three to five workflows that matter to the business, such as month-end reporting, proposal drafting or customer case handling, and for each one write down:
- What it looks like today: the steps, the inputs and the output it produces.
- How long it takes: the end-to-end time, and the hands-on time inside that.
- Who's involved: every role that touches it, including reviewers and approvers.
- How quality is judged: the standard the output is held to, and who decides whether it's met.
That last one is often the hardest to answer and the most useful. A conservation charity we work with was clear from the start that productivity mustn't come at the expense of quality, and that poor data in means poor data out. If you can't say how quality is judged today, you won't be able to show that AI has improved it.
None of this needs to be elaborate. Time a sample of real work, note who touched it, and agree what a good output looks like. That small step makes everything that follows defensible.
Behaviour: understand how the work really gets done
A process map shows how a workflow is supposed to run. Behaviour is how it actually runs, and that's where most of the hidden cost and value sit. For each workflow, find out:
- Which tools are used today: the systems, spreadsheets, templates and workarounds people rely on.
- Whether AI is already part of the process: officially or otherwise.
- The subjective checks: the judgement calls someone makes before the work goes out.
- The tacit knowledge: what an experienced person knows that isn't written down anywhere.
AI is often in the process already. The AI lead at a large UK charity told us colleagues were using Copilot privately because they worried it looked like cheating, or might flag them for replacement. If that use is invisible, your baseline is wrong before you start.
A survey won't reliably surface it either.
Asking "how confident are you with AI?" tells you more about confidence than capability. Look at the work itself, and assess demonstrated skill.
The checks and the tacit knowledge matter just as much, because they're where quality is won or lost. A regulated professional services firm told us its London and Manchester teams do the same task differently and produce different outputs. A tyre business we work with decided that automated customer reports must still be vetted by account managers, because a report can be accurate and still land badly with a client. These are the parts of a workflow AI can't skip, and the parts your measures need to protect.
Once you know how the work happens today, track how that behaviour changes. Usage on its own can mislead. The same charity AI lead wanted to tell apart the advocates, the people who started then dropped off, and the silent non-adopters, because each group needs something different.
Benefit: decide what good looks like
Before you measure the gain, agree what the gain is meant to be. For any workflow, a good outcome usually takes one of three forms:
- Time saved: the same output, produced faster, measured on real work.
- Improved quality: fewer errors, less rework, better review outcomes or happier customers.
- Increased revenue: more proposals out of the door, faster responses to new enquiries, higher win rates.
These aren't interchangeable, and time isn't always the point. The CEO of one membership body told us AI hasn't saved him any time, but it has significantly raised the quality of his output. A time-only measure would score that as zero.
Employees also report gains that leaders can't see.
Both groups may be right. Time saved by an individual only becomes value when it shows up somewhere the organisation can see it. The charity AI lead put it simply: monitoring that only shows efficiency is incomplete, because what matters is what people do with the time they get back. So for every hour you claim, record where it went, whether that's more cases handled, faster turnaround or work that was previously out of reach.
Business: tie each workflow to the metrics that matter
Success should be judged at the workflow level, not as one overarching number for the whole programme. A finance team, a sales team and a customer service team will each get a different kind of return, and each will show up in a different metric.
One business owner we work with set a single test for his finance team: monthly financials delivered by the 7th instead of the 14th within six months. It's specific, finance already tracks it, and everyone knows what success looks like. Every other team needs a test of its own, tied to a number the business already runs on.
Not all of those numbers carry the same weight, though.
Treat cost saving as the baseline, not the goal. Automation and efficiency are the minimum a workflow improvement should deliver. The real case for AI is what becomes possible once that time and capacity are freed up: better quality, happier customers and more revenue. The NHS transformation lead described the same sequence. Early benefits show up as staff experience and efficiency, then capacity and quality, with cash arriving last, despite the pressure to show cost savings first.
Taken together, workflow improvements should move the business along strategically. If a workflow gets faster but nothing the leadership team cares about moves, the chain has broken somewhere between benefit and business.
Measurement discipline seems to separate the organisations seeing returns.
A simple way to calculate the ROI of AI
Once each workflow has its own measures, you can roll them up. Start with this:
ROI of AI = (value of benefits realised − total cost) ÷ total cost
Value of benefits realised = verified hours saved × loaded hourly cost × realisation rate, plus measured quality gains and any revenue you can fairly attribute.
Total cost = licences + training + the time people spend learning + programme management.
The realisation rate is the honest part. It's the share of saved time that gets reinvested in valuable work. If you can't evidence where the time goes, use a conservative figure and say so.
Here's an illustrative example with made-up numbers. 100 people save a verified 1.5 hours a week across 44 working weeks. At a loaded cost of £35 an hour, that's £231,000 of time. Apply a 50% realisation rate and the benefit is £115,500. If licences cost £36,000, training £20,000 and learning time £21,000, total cost is £77,000. ROI is (£115,500 − £77,000) ÷ £77,000, or 50%.
| Benefit | |
|---|---|
| Verified hours saved (100 people × 1.5 hours × 44 weeks) | 6,600 hours |
| Value of time at £35 an hour | £231,000 |
| Benefit realised at a 50% realisation rate | £115,500 |
| Cost | |
| Licences | £36,000 |
| Training | £20,000 |
| Learning time | £21,000 |
| Total cost | £77,000 |
| ROI: (£115,500 − £77,000) ÷ £77,000 | 50% |
A CFO will challenge the hours and the realisation rate. That's fine, because your baseline and benefit receipts are the answer. And because this example only counts time, it's the floor. Gains in quality, customer satisfaction and revenue sit on top of it.
Common mistakes when measuring AI ROI
Treating usage as return.Logins prove access, and access is only the first receipt.
Skipping the baseline.Without a "before", every "after" is an opinion.
Ignoring the tacit checks.If you don't know how quality is protected today, you can't show AI hasn't eroded it.
Counting time saved without asking where it went.Saved time that disappears into email isn't ROI.
Measuring at programme level only.One blended number hides which workflows are working and which aren't.
Stopping at cost savings.Efficiency is the floor. Quality, customer satisfaction and revenue are where the case is won.
Measurement is the hardest part of any AI programme, and we've yet to meet an organisation that has it fully solved. It's also the best predictor we see of whether a programme survives its next budget review.
Frequently asked questions
What is a good ROI for AI?
There isn't a reliable universal benchmark, because returns depend on the workflows, the starting point and how costs are counted. A better test is whether each workflow's return is evidenced at every link in the chain, uses a conservative realisation rate, and ties to a metric the business already tracks.
Should you measure AI ROI at company level or workflow level?
Start at the workflow level and roll up from there. Success looks different for each workflow and team, so a single programme-wide figure hides where the value is and where it isn't. Tie each workflow to a metric the business already runs on.
Is cost saving the main return from AI?
Cost saving is usually the first and easiest return to evidence, so treat it as the baseline. The bigger returns tend to come from improved quality, customer satisfaction and revenue, which take longer to show but carry more strategic weight.
How do you measure Copilot ROI beyond usage?
Pair your usage dashboard with a baseline of a few specific workflows and measure them again after rollout. Usage shows who has started. Workflow-level time, quality and revenue data show whether that use is producing value.
How do you prove AI time savings to a CFO?
Time real tasks before and after, on the same type of work, and report where the saved hours were reinvested. Present the figures with a conservative realisation rate and spell out your assumptions. Finance teams trust measured samples far more than survey estimates.
How long does it take to see ROI from AI?
Changes in behaviour can show within weeks. Workflow benefits usually follow once people have practised on real work, and business outcomes take longer still. Set expectations for each link separately instead of promising one date for everything.
Where to start
Pick three to five workflows and walk each one through the chain.
Baseline. Do you know how long it takes today, who's involved and how quality is judged?
Behaviour. Do you know which tools are used, where AI already appears, and which checks and tacit knowledge the work depends on?
Benefit. Have you agreed what good looks like for this workflow: time saved, better quality or more revenue?
Business. Which business metric will it move, beyond cost, and by when?
If any link is missing for a workflow, that's your first job.
You can do all of this manually, or you can do it with AcademyAI. We assess each person's AI capability before training starts, build role-specific learning into the flow of work, and evidence the improvement in the work itself. If you'd like to see where your organisation stands, .