AI ROI is the measurable financial return an organisation gets from its AI spend, set against what that spend costs to run.
Almost nobody can produce it. ivee asked 500 UK AI decision makers in August 2026 whether they could prove the return on their organisation's AI spend, and 4% said they track it and can show the numbers. The other 96% could not. Half of the sample had never tried to measure it at all.
The headline findings from the research:
4% of AI decision makers can prove the ROI of their organisation's AI spend. 96% cannot.
50% have never tried to measure it.
75% say either that they have no idea, or that fewer than 1% of their workforce can build and deploy an AI automation.
Asked which costs more money, a thin prompt or a loaded, context-rich one, 80% picked the loaded prompt. That is the wrong answer.
Asked to size the difference, answers ranged from 2x to 100x. In ivee's data the figure is 14x.
Those are not five findings. They are one. An organisation cannot put a number on a return when the people doing the work cannot build the thing that produces it, and the person signing the invoice cannot price the unit they are buying. Below is the full dataset, the method behind it, and the questions the room asked once they saw it.
How many organisations can prove the ROI of their AI spend?
4% can. Asked whether they could prove the ROI of their organisation's AI spend, 4% of 500 UK AI decision makers said they track it and can show the numbers. Every other answer was a version of no.
Answer | Share of respondents |
|---|---|
Haven't tried to measure it | 50% |
Think the value is there, but can't prove it | 32% |
Not seeing a clear return | 16% |
Track it and can show the numbers | 4% |
Respondents could select more than one answer, so the column sums above 100%. Base: 500 UK AI decision makers, surveyed 22 to 25 August 2026.
The 32% is the most revealing number in the set. That group is not disillusioned. They believe the value is real, they are willing to keep spending, and they have nothing to hand a finance director who asks how they know. Belief is doing the work that evidence should be doing, and belief has a poor record here: in METR's randomised controlled trial published in July 2025, 16 experienced developers estimated AI had made them 20% faster on 246 real tasks, while the measured result was a 19% slowdown. The direction was wrong, not just the magnitude.
Why can't AI budget holders prove the return?
Because half of them have never tried. The single largest answer in the dataset is not "we measured and the result was bad". It is 50% saying they have not attempted a measurement at all, which means the question of whether the money worked has never been asked in a form that could produce an answer.
The mechanical reason is almost always the same. Nobody took a baseline. Once a tool is rolled out, the before state is gone, and every number available afterwards describes activity rather than value: seats issued, licences active, people who opened the app in the last 30 days. Those are habit measures wearing a value measure's clothes, and which of the three common measures survives a question from a finance director is a decision worth making before the training is booked, not after.
The 16% who report no clear return are, in one sense, ahead of the 50%. They looked.
How many employees can actually build an AI automation?
Almost none, and most of their leaders cannot say for certain. 75% of AI decision makers say either that they have no idea, or that fewer than 1% of their workforce knows how to build and deploy an AI automation. Fewer than 10% say that half or more of their organisation could do it.
Read that as a capability distribution rather than a training statistic. An organisation with a 1% build rate has a handful of people producing the automations and everybody else consuming them, which caps the return at whatever those few can ship. It also explains the ROI numbers directly: value from AI arrives when work changes shape, and work changes shape when somebody builds something. Seats do not build anything.
The "no idea" half of that 75% is its own finding. These are the people who hold the budget. Not knowing what proportion of the workforce can build with the tools they are paying for is a measurement failure sitting one level above the ROI measurement failure, and it is the cheaper of the two to fix. An AI skills assessment takes five to eight working days and produces the denominator that every later ROI calculation needs. It also tends to find that the average describes nobody, which is the point of running it at team level rather than reading the national picture in the UK AI skills gap and assuming it applies evenly.
Do AI budget holders know what they are paying for?
Four in five have the direction of the cost backwards. Asked live which costs more money, a thin prompt or a loaded, context-rich one, AI decision makers split 80/20, and the 80% picked the loaded prompt.
They are not wrong about tokens. They are wrong about the unit of account. Per call, a longer prompt does cost more, because every token sent is billed as input. Per completed piece of work, the thin prompt costs more, because it does not produce the answer and the difference gets paid for in turns.
The pricing makes this concrete. On Anthropic's published API pricing, Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, a five-to-one ratio, and a cached input token reads at a tenth of the base input price. So the expensive half of any exchange is the half the model writes, and the cheap half is the half you can front-load and cache. A short prompt saves money on the cheapest, most cacheable component and spends it on the most expensive one, repeatedly.
An 80/20 split on a two-option question is not a knowledge gap at the edges. It is a five-to-one majority holding the wrong model of their own cost base while approving spend against it.
How much more does an under-specified prompt cost?
14 times more, on average, in ivee's data. A thin prompt rarely lands first time, and the cost is not the prompt, it is the back-and-forth needed to repair it: each correction resends the whole conversation as input and buys another round of output at five times the input rate.
Asked to estimate that multiplier themselves, AI decision makers gave answers from 2x to 100x. Two orders of magnitude of spread, on a cost their organisations are already paying every day. The 2x answers are budgeting for a rounding error. The 100x answers cannot plan at all.
What makes this the most tractable finding in the set is that it is a writing problem, not a procurement one. Nothing needs renegotiating. The same task, specified properly on the first attempt, costs a fraction of the same task discovered through six rounds of correction, and the skill transfers across every model an organisation uses. It is also the one place where tooling does the work for you: ivee's prompt improver exists because a 14x range is too wide to leave to individual habit.
What AI decision makers are actually asking
The questions asked during the live session map the confusion more precisely than the poll numbers do. Three patterns came up repeatedly, all from people who hold AI budgets.
On what a token even is. "Is a token the same as a Claude usage credit?" "Is your token cost on top of the subscription per user costs?" "How does creating a skill or a project or persona reduce token usage? Because it is still having to read through the instructions before it does anything, so is that like a giant prompt?" These are not naive questions. They are the questions of someone trying to reconcile an invoice.
On vendor pricing and the exit. "Once you automate loads and then the price goes up, can you swap vendor of the AI software?" One attendee put it more sharply: "The LLM vendors are the same as cloud vendors - they want you to easily spin up resources, but they make it difficult to track. It is like a drug dealer: sell it cheap at the start, then ramp up afterwards." Another had seen it before: "We had the same issue, and it continues today, about cloud and people jumping onto the bandwagon without understanding the reality."
On who is accountable when a non-specialist builds something. "Will the marketing team, although they can make the applications they need, think about security and compliance? ISO 20000, ISO 27001, ISO 9000. Will the AI guarantee that the application it builds will be safe to use?" And, from someone with a decision sitting in their inbox: "Someone has requested access to an automation tool and I haven't allowed it yet, as I do not know the wider risks to the business."
Every one of those is a governance question dressed as a technical one, and none of them is answered by a usage dashboard.
What to do if you hold the AI budget
Four things, in this order, and the order matters more than the speed.
Take the baseline before the next rollout. Not after. The before state is the only part of an ROI calculation that cannot be reconstructed later, and 50% of this sample no longer has one.
Count builders, not seats. Establish what proportion of your workforce can build and deploy an automation today. If the honest answer is "no idea", that is the same answer three quarters of your peers gave, and it takes about a week to fix.
Price outcomes, not calls. Measure the cost of a completed piece of work, including the corrections. A per-call view makes the thin prompt look cheap and hides a 14x multiple.
Write the exit before you automate. The question "can you swap vendor" gets expensive in exact proportion to how much you have automated before somebody asks it.
None of that requires a new platform. It requires deciding what you are measuring before you spend, which is the step 96% of this sample skipped.
How this research was conducted
ivee surveyed 500 UK AI decision makers between 22 and 25 August 2026: budget owners, heads of AI adoption and C-suite founders at organisations actively building with AI. Respondents were drawn from ivee's AI leaders network and its AI leaders forum, whose membership includes senior AI leadership at organisations such as Microsoft. The job title mix is predominantly heads of AI adoption, followed by CEOs and founders, followed by functional heads.
The same 500 people were polled live during ivee's webinar "The hidden cost of AI: what every budget holder needs to know about tokens", held in August 2026. The token literacy findings and the verbatim questions above come from that session. This is one sample measured with two instruments, not two independent samples, and the figures should not be added together.
Some questions permitted multiple answers, which is why response shares can exceed 100%. The sample is UK-only and skews towards organisations already building with AI rather than those considering it, so it describes the state of measurement among committed adopters, not the national picture. For that, the Department for Science, Innovation and Technology publishes AI Adoption Research (DSIT 2026/003, January 2026), which tracks how far AI has been adopted and scaled across UK businesses.
Journalists, analysts and anyone wanting to cite these figures are welcome to. Attribution to ivee, with the sample size and the August 2026 date, is all we ask.
Where this leaves you
If you can already show the numbers on your AI spend, you are in the 4% and the rest of this page is confirmation. If you cannot, the useful question is not whether AI is working, it is which of the three failures you have: no baseline, no builders, or no idea what a completed piece of work costs. They need different fixes and they are usually diagnosed in an hour.
ivee runs that diagnosis with leadership teams before recommending any training. Book a call and bring whatever numbers you do have.
