In brief
- Most organisations can say what their AI cost and what it saved. Far fewer can say what it is doing to judgement, expertise, and how dependent they have become.
- Those effects are real and documented - but they arrive later than the savings, and get recorded somewhere else entirely. So the trade is made without being seen.
- They run both ways. The same deployment can build human capability or quietly spend it down, and this year's numbers look identical either way.
- The assessment is a structured way to look at both halves together, across five areas, and produces a report you can check, disagree with and act on.
The core assessment takes 15-20 minutes. If you also complete the optional financial section - the fuller version - allow 30-45 minutes. It is free while being tested, and nothing commercially sensitive is required. The rest of this paper explains why the question is hard to answer, and what the assessment does about it.
Three questions
Think about your most significant AI deployment. Then answer three questions about it.
What did it make possible?
What did it make unnecessary?
What did it make harder to do without?
Most organisations can answer the first. Many can answer the second - it is usually in the business case.
Very few can answer the third.
What you already measure
By now your organisation can likely tell you, for any significant AI deployment: what it cost, what it saved, whether it is faster, whether quality held up, whether customers noticed, whether people are using it, and whether it passed governance.
That is a reasonable set of measures, and a competent transformation team will have them.
What those measures were not designed to tell you
In each pair above, the first question is a fair one and most organisations can answer it. The second is not a harder version of the first - it is a different question altogether, and nothing that answers the first will answer it.
Why this is difficult to see
There are two separate reasons, and they need different responses.
Some of it nobody measures
No organisation tracks whether a review step produces real scrutiny or simply records an approval. None tracks whether the expertise needed for difficult cases is still being built. These are not numbers filed in the wrong place. They do not exist.
And some of it is measured, but in a way that hides the connection
Consider how the two halves of a single decision travel through an organisation.
- The saving appears this year, in a named budget line, attributed to the AI initiative. It is clear, early and easy to report.
- The consequence appears later - as an incident, a recruitment cost, a slower response on an unusual case, a business continuity problem. Each is recorded. Each is attributed to the incident, the labour market, the team, the IT estate.
Both halves are visible. They are just never visible together. By the time the second half arrives, nobody connects it back to the decision that produced it.
This is not a failure of competence. It is how organisations are structured. Finance sees one part, Operations another, HR another, Technology another. The connection is nobody's job.
Three things worth knowing
These come from published research. Sources are available in the accompanying evidence paper.
1. A human in the loop is not the same as human judgement
In a study of UK general practitioners using clinical decision support, the system improved decision accuracy in 13.1% of prescribing cases. There was a net improvement in performance.
But in 5.2% of cases, clinicians switched from a correct prescription to an incorrect one after the system gave them poor advice. The same tool that helped more often than it harmed also caused some good decisions to become bad ones.
The study also found that less experienced clinicians switched their decisions more often. That matters here for a particular reason: if AI simultaneously removes the routine work through which people become experienced, the two effects compound.
2. The same tool can build capability or consume it
A growing body of research distinguishes two ways people use AI: one in which the person delegates the thinking, and one in which the person does the thinking and uses AI as support. These patterns appear to have very different effects on what people retain and develop.
What that means for leadership: it may be the way AI is used that determines the outcome, not how much it is used. Two organisations with identical adoption rates and identical productivity gains could be moving in opposite directions - and no adoption metric would tell them apart.
3. People are poor judges of their own AI productivity
In a controlled study, experienced software developers using AI tools took longer to complete tasks than those without. They expected to be substantially faster. Afterwards, they still believed they had been faster.
Perceived productivity and measured productivity are not the same thing - and where they diverge, the gap tends to flatter. Where your evidence that AI is working relies heavily on what people report, it is worth testing those reports against something measured.
Two situations worth recognising
You automate your customer service front line
Cost per interaction falls. Resolution times improve. Satisfaction holds. Every measure says this worked.
Meanwhile, the straightforward cases stop reaching your people. Only the difficult ones do. Your experienced staff are fine - they learned on the easy cases years ago. Your newer staff never get them. In three years, when those experienced people have moved on, the question is who handles the difficult cases, and where they learned to.
Nothing here is an argument against the automation. The question is whether anyone decided this would happen, or whether it simply did.
If that capability has to be rebuilt later, the organisation pays for it - through recruitment, training, slower handling, or some combination of the three.
You introduce AI into a professional decision process
Decisions get faster and more consistent. A human reviews each one before it goes out - a proper control, documented, auditable.
The reviewer has thirty seconds. They have a throughput target. The output is confident and usually right. They approve 99% of what they see.
Ask whether that step is oversight or a record of approval. Then ask how you would know the difference.
If the control fails, the cost is not the cost of the review step. It is the cost of the decision the review did not challenge.
What the assessment does
It asks a structured set of questions across five connected areas, and produces a report.
What you get back
Where the organisation stands on each area. What the assessment found, with the answers that triggered each finding, so you can check it or disagree with it. Each also shows what the situation looks like if left as it is, and who would need to approve doing something about it.
Where you supply financial figures, the report sets out which potentially material costs and benefits sit outside them - so the reported return can be read in its wider context, and so you can decide whether any are worth putting a number to.
If several people take it
Invite four or more colleagues at different levels and we compare the answers. Where you agree, you have something firmer than any one person's view. Where you disagree - particularly between those who make decisions and those closest to the work - the gap itself becomes the useful thing. Both accounts can be accurate from where each person sits. The assessment makes those differences visible rather than averaging them away.
What it is not
Being clear about this matters more than the pitch.
It does not replace your AI strategy, financial analysis, risk review or governance - those remain necessary, and this sits alongside them. It does not assume your transformation team has missed something obvious; they are probably already doing much of this work, and what is hard is connecting it across functions.
It puts no pound value on human capability - it identifies where capability may be changing, and what that is worth is a question for your figures, not ours. It does not tell you AI is bad for your organisation: capability can go up as easily as down, and the assessment is built to find either. There is no single score, because a number would hide exactly what you need to see. And it is not an evaluation of individual employees.
What taking part involves
Deliberately simple. There is nothing to install, nothing to prepare, and no need to assemble data in advance.
1 · You complete the assessment
Online, in one sitting, answering from your own experience of AI in the organisation. The core assessment takes 15-20 minutes. If you also complete the optional financial section - the fuller version - allow 30-45 minutes. Answer for what actually happens, not what is intended.
2 · You get your report straight away
On screen as soon as you finish, ready to save or print. Nothing to wait for.
3 · Colleagues can take it too
Where four or more people across different levels complete it, we compare the responses and produce a second report showing where views align, where they diverge, and what that difference may be telling you.
4 · We go through the findings together, if useful
A conversation rather than a presentation. The questions worth asking are usually: are the expected benefits being realised, what costs are visible and which may not be, where is time being saved and where is work being created, are different teams taking overlapping or fragmented approaches, and what should be measured before the next investment decision.
5 · You tell us what worked
What was useful, what was not, and what was missing. That is the part we most need.
Where this stands, and the invitation
Being straight with you
This assessment is being tested. It is built on a documented framework and the evidence base behind the questions is real, but the assessment itself has not yet been validated against the thing that matters most - whether it changes decisions. That is what the current trial is for.
We are working with a small number of organisations already using or introducing AI, who are willing to take it, say what worked and what did not, and tell us whether it showed them anything.
The organisations that benefit most from AI may not be those that deploy the most technology. They may be those that learn fastest about what it is actually doing to them.
The test is simple
- Did it show you something about your AI transformation that you were not already seeing?
- And did it change, or sharpen, a decision?
If the answer to both is no, we would rather know.
Take a look at the assessment:
capacity2grow.com/human-flourishing-ai-assessment
I would be very happy to explain the approach, answer questions, and discuss whether a trial would be useful for your organisation.
Firooz Firoozmand · Capacity2Grow
AI value: what gets missed in the business case?
Suppose an organisation estimates that an AI programme will create £5m of annual gross benefit. That might look like a very successful transformation. But let us make some of the costs visible.
Still a substantial benefit. But notice what has not been priced - and that is where the business case gets more interesting.
The figures above are illustrative assumptions, not industry benchmarks. In a real assessment, the organisation's own financial and operational data should be used.
Beyond the business case: what happens to capability?
The £2.5m calculation is useful. But it still does not tell the whole story. What happens to the things that are difficult to put a reliable value on?
These should not simply be given invented pound values. They are potential value at risk - things that may matter to the organisation but are harder to quantify reliably.
If people stop developing certain capabilities today, what happens to the organisation's capability in year three or year five? The effect could compound over time.
So the question is not simply "Should we automate?" It is: