How to tell whether your AI rollout actually stuck

Your low usage number is measuring the wrong thing

The University of Melbourne and KPMG surveyed more than 48,000 people across 47 countries, with nationally representative sampling. Their finding should stop any executive looking at a disappointing adoption dashboard:

57% of employees say they hide their use of AI and present AI-generated work as their own.

Nearly half admit using AI in ways that contravene company policy, including putting sensitive company information into free public tools. Only 47% have received any AI training. Only 40% say their workplace has a policy on generative AI at all.

The Federal Reserve found the same shape in US data.

41% of workers report using AI at work, against 18% of firms that had adopted it as of the end of 2025. The Census Bureau’s own April 2026 working paper on AI diffusion, drawn from nationally representative data collected between November 2025 and January 2026, states it plainly: worker use can occur without firm adoption, and firm adoption can occur without worker use. The firm-level number has since moved into the high teens and low twenties depending on the reference period; the gap between workers and firms is what matters here, and it has not closed.

Read those together and the low usage dashboard means something different from what it appears to mean. The work is being done. It is being done on personal accounts, outside your systems, by people who have concluded that mentioning it is not in their interest. You are not looking at low adoption. You are looking at the part of adoption you can see.

Which is the worst possible position: paying for seats, getting none of the governance, none of the learning, and none of the compounding, while the exposure accrues anyway.

And depth is where it actually breaks

Even inside the firms that have properly adopted, use is shallow.

The Census Bureau’s paper, authored by economists including John Haltiwanger and Census Chief Economist Lucia Foster, found that among firms using AI at all, 57% use it in three or fewer business functions, most often sales and marketing, strategy and IT. Among firms with worker-level use, 65% restrict it to three or fewer tasks, mostly writing, document analysis and search.

And frequency is thinner still. The Federal Reserve puts daily generative AI use at 12% of workers against 41% reporting any use at all. Roughly one in three people who touch it uses it every day.

A weekly user and a daily user cost you exactly the same. Seat count cannot see the difference between them, and the difference between them is the entire return.

Three quarters of executives think they are measuring this

Here is the squeeze, from four independent studies of different populations.

Wharton’s Human-AI Research group, with GBK Collective, surveyed around 800 US enterprise decision-makers at companies above 1,000 employees, tracking whether they measure ROI at all and whether they see it.

Then look at what they measure. Among those measuring, the most common metric is employee engagement and productivity at 47%, which is a self-report. Only 36% link AI spend to KPIs aligned with business goals.

McKinsey’s 2026 survey of 1,719 participants across 97 nations found the EBIT-attribution figures above. Gartner, surveying HR leaders, found 88% reporting their organisations have not realised significant business value from AI tools, on a small base of 114.

Three quarters of executives believe they are measuring the return and getting one. About one company in sixteen can attribute a material share of EBIT to it. Both of those cannot be true.

The time-saved number is not a return

There is one specific trap worth naming, because almost every buyer has walked into it.

Gartner surveyed 2,986 employees in July 2025 and found 62% reporting AI saved them time, averaging about 1.5 hours a day in AI-relevant roles. That is the number a vendor puts in front of a CFO.

The same survey found only 7% of organisations provide any guidelines on how to use the time AI frees up, and only 42% of employees know how to identify where AI would improve their work.

Time saved with no plan for the time is not a return. It is slack. It shows up in a survey and it does not show up in a P&L, which is precisely why 62% of employees report saving 1.5 hours a day while 88% of HR leaders report no significant business value. Both findings are true at once and there is no contradiction between them.

“Empowering employees is not enough and has no significant effect on likelihood of exceeding revenue goals.”

— Sari Wilde, Practice Vice President, Gartner HR practice

Buying everyone a licence and hoping is not a rollout. It is a purchase.

What a scorecard actually measures

There is no standard framework for this, which is worth saying out loud. What exists is either a vendor’s telemetry dashboard, which counts logins, or a consultancy’s maturity model, which counts opinions. Nobody publishes a departmental baseline-and-remeasure method, and the Federal Reserve has documented that even national statistical agencies disagree on how to measure AI adoption because of question framing and materiality.

So here is what a Rollout Scorecard measures, and the order matters more than the list.

First, a baseline, taken before anything is configured. How long the department’s core documents take to produce today. How often they come back for correction. What the exception rate is. Captured on the current process, in the current week, before a single account is provisioned. Once configuration starts, this becomes unrecoverable, and every number afterwards becomes an opinion.

Then, department by department, six weeks in:

• Depth, not breadth. How many distinct tasks each department runs through it, against the national picture where two thirds of firms stay inside three.
• Frequency. Daily against weekly against occasional, per department, since one in three users nationally is a daily user and that is where the return lives.
• The specific artefact. How long the 8D, the RFI response, the job order rewrite, the SLA report takes now against the baseline, measured on real output rather than on how people feel about it.
• What got retired. Which manual step is gone. A step that survives alongside the new one has doubled the work rather than replaced it.
• What moved off the shadow estate. Whether the people who were using personal accounts are now inside your governance, which is the one metric that turns an exposure into an asset.

Five measures, per department, against a baseline. That is the whole product, and the reason it is a document rather than a dashboard is that a dashboard cannot tell you what was true before you started.

Why department by department, and why not just training

Forrester’s 2026 research names siloed adoption within functions as one of the reasons enterprises are still chasing returns three years in. Adoption is not a company-level property. Sales gets there and quality does not, and the company average conceals both.

The manufacturing evidence is sharp on why. PwC and the Manufacturing Institute surveyed 102 HR and operations leaders and found only 19% of firms offer any AI-related training, 49% saying existing programmes do not prepare frontline employees for AI-driven role changes at all, and, most usefully, 45% attributing unsuccessful AI initiatives partly to frontline leaders not being sufficiently included in design or rollout. Integration reached 2% of frontline leaders and 8% of frontline workers.

But here is the part that separates a real answer from the standard one. The obvious conclusion is “invest in training,” and the evidence is less flattering to that than you would expect. A 2026 meta-analysis in the Journal of Information Science covering 693 papers and more than 221,000 respondents found that perceived usefulness and performance expectancy dominate as predictors of actual technology use, and that facilitating conditions such as training and support are consistently weaker predictors. The effects have not changed across decades of technological advancement.

People use a tool when they believe it makes them better at their own job. Training helps them do a thing they already want to do. It does not create the wanting.

Which is why a rollout that works is organised around each department’s own highest-friction document rather than around a curriculum, and why the scorecard measures the artefact rather than the attendance.

Where the first step goes

A baseline is cheap and it expires. Every week a rollout runs unmeasured, the ability to prove what it did degrades, because the comparison you needed was the week before it started.

Neurony Wire runs this department by department, on a baseline taken first, from $15,000: a technical readiness assessment, configuration and a pilot, sequenced rollout waves, and the Scorecard at the end. Actual usage data pulled from the tool itself, measured against a before that was captured properly.

For scale, put that next to what is already committed. A hundred seats at typical enterprise pricing runs past $15,000 inside a year, and roughly half of installed enterprise software goes unused. The measurement costs less than the licences it is measuring, and it is the only part of the spend that tells you whether the rest of it worked.

If you have not bought anything yet, the sequence is the same one described in the difference between a pilot and a proof-of-concept: baseline, live data, metrics set in advance, and a real decision at the end.

If you already bought the seats, the baseline for those is gone. The recoverable position is to pick one department, baseline it now on its current real performance, and treat that as the starting line. It is a worse starting point than the one you had six months ago, and it is considerably better than the one you will have in another six.

This is the same pattern across every industry that runs the physical world: the licences land, the demo goes well, and nobody writes down what things looked like before.

The question to ask on Monday

Not “how many people are using it.” That number is visible, comfortable, and measuring the wrong thing.

Ask instead: which department can show me a document that takes less time than it did in March, and how do we know what it took in March?

If nobody can answer the second half, there is no scorecard, and there never was one.