You are not deciding whether your people use AI. They already are. You are deciding whether you can see it.
Employees are already using AI at work in numbers that make the do-nothing option imaginary, and a meaningful share of them hide it. That produces three costs that never show up as a line item: inconsistent quality nobody traces back, sensitive information going into tools you have no relationship with, and the same problem being solved eleven different ways across one department. The business case that works internally is not about productivity percentages. It is about naming a specific recurring workflow, a specific quality standard, and a specific thing you will be able to measure in ninety days. There is also a real limit worth knowing: a short primer on prompting does not produce judgment, and one field experiment found the group given exactly that did worst on the task the AI could not do.
A marketing manager I spoke to earlier this year described her team’s AI setup like this: two people on personal ChatGPT accounts, one using Gemini because it came with her Google login, one refusing on principle, and one who “definitely uses something but won’t say what.” Nobody had been told not to. Nobody had been told anything.
She wasn’t running a company without AI. She was running a company with five uncoordinated AI policies, none of which she’d written and four of which she couldn’t see.
You are not choosing between AI and no AI. You are choosing between managed and unmanaged.
That reframing matters because “falling behind competitors” is a cost you can always defer. It has no date on it and no invoice attached, so it loses to whatever is on fire this month. The costs of unmanaged use, by contrast, are already being paid. They’re just not being counted.
Some sense of scale is useful here, and the most careful work on this comes from a study run by Melbourne Business School researchers with funding from KPMG, covering 48,340 people across 47 countries, fielded between November 2024 and January 2025. Two findings from it stay with me. More than half of employees said they’d made mistakes in their work because of AI. And a similar proportion said they hide their AI use and present AI-generated work as their own.[1] The fieldwork is now some way behind us, which is worth holding in mind, though nothing since suggests the direction has reversed.
Put those two together and you get the actual problem. Not that people make mistakes with a new tool, which is normal and survivable. That they make mistakes with a tool they’ve decided not to mention, which means the mistake never gets traced back to a cause, and the same mistake happens again next month in a different team.
The public version of this is instructive because it happened to an organisation with every possible resource. Deloitte Australia produced an assurance review for the Australian Department of Employment and Workplace Relations under a contract worth A$440,000. The published document contained fabricated academic references and a fabricated quotation from a Federal Court judgment. A corrected version was issued in September 2025, disclosing that Azure OpenAI had been used in preparing it, and Deloitte agreed to repay part of the fee.[2]
I’d be careful about how that’s used, because the easy reading is wrong. That wasn’t untrained junior staff. It was a governance and review failure at a firm that sells AI advisory work. Which makes it a stronger argument, not a weaker one: if the review step can fail there, “our people are sensible” is not a control.
| The cost | What it looks like on the ground | How you’d detect it |
|---|---|---|
| Untraceable quality drift | Work that’s fine but slightly generic, and nobody can say when it started | Pull ten pieces of client-facing output from this quarter and last year’s. Read them side by side |
| Hidden use | People using AI and not saying so, which means no shared learning and no error trail | Ask in a one-to-one, not a survey. “What are you using, and what do you not trust it with?” |
| Eleven versions of one solution | Five people each maintaining their own prompt for the same recurring task | Ask three people in the same role to show you how they do the same job |
| Data leaving without a route back | Client material pasted into a tool your organisation has no agreement with | Check what’s actually installed and logged in, not what’s on the approved list |
| Paid seats nobody opens | Licences bought in a wave of enthusiasm, used by four people | Pull the usage figures from the admin console. This one takes ten minutes |
A detection register for the costs that never appear as a line item. Every check in the right-hand column can be done this week without a budget. Framework, not measured data.
When leaders picture an AI skills gap, they usually picture people who don’t know how to use the tools. That group exists, and they’re the easiest group to help. They’re not where the money is going.
The expensive gap is people who use AI confidently and can’t tell when it’s wrong. Confidence and accuracy come apart, and nothing in the experience of using these tools tells you which side of the line you’re on, because a wrong answer arrives with exactly the same polish as a right one.
There’s a field experiment that makes this uncomfortably concrete, and it deserves a moment because it cuts against the article I’m writing. Researchers working with 758 consultants at Boston Consulting Group randomised them into three groups: no AI, AI, and AI plus a short prompt-engineering overview. On tasks inside the model’s capability, the AI groups did substantially better. On a task deliberately placed outside it, the pattern inverted, and the group that had received the prompt-engineering primer produced the lowest share of correct answers of all three, below the group with no AI at all.[3]
Read that carefully before you use it as ammunition either way. The “training” involved was a brief overview of prompting, which is a long way from an upskilling programme, so it tells you nothing about whether training works. What it shows is narrower and considerably more useful: tool training on its own doesn’t produce judgment, and it can leave people more willing to trust an output they should have questioned.
Train the work, not the tool. A prompting session teaches buttons; the job teaches judgment.
Which is why “we ran an AI lunch and learn” and “our people are capable with AI” are further apart than most leadership teams assume. And it’s why the gap looks different depending on who you’re asking about, so it’s worth being specific rather than talking about “employees”.
| Role | What they’re actually stuck on | What would move them |
|---|---|---|
| Content marketer | Output that’s fast and sounds like everyone else’s. They can produce, not differentiate | Working on their own brand rules and a quality bar they can test against, not more prompts |
| Finance analyst | Can’t tell a plausible number from a checked one, and the stakes make them avoid it entirely | Verification practice on real files, and a clear list of what AI never touches |
| HR business partner | Genuine and correct nervousness about personal data, so they use nothing at all | A decision rule about what data goes where, which converts anxiety into a boundary |
| First-line manager | Their team uses AI, they don’t, so they can’t set a standard for work they can’t assess | Enough hands-on time to review output credibly. This group is skipped most often |
| Senior leader | Approving spend on something they’ve never used for real work | One real task, done themselves, start to finish. It changes the questions they ask |
Role-specific diagnosis. Generic AI training aimed at an undifferentiated “employee” reads as generic because it is. Framework, not measured data.
The first-line manager row is the one I’d argue hardest for. A manager who can’t assess AI-assisted work can’t hold a standard, and without a standard the quality question quietly becomes a taste question. That’s how a team ends up with five people producing work of five different qualities and nobody able to say so.
This section is usually where an article tells you everyone else is racing ahead and you’re about to be left behind. That’s not what the evidence shows, and overstating it would make the rest of this less useful.
What’s actually happening is more interesting: movement, from a low base, with most organisations still not there. ISACA’s 2026 pulse poll of more than 3,400 digital trust professionals found a third saying their organisation trains all employees on AI, up from around a fifth the year before.[4] Worth flagging that ISACA sells AI certifications, and that its respondents are audit, governance and security specialists rather than a cross-section of employers, so read it as what risk professionals see rather than as a national average.
Two things follow from that, and they point in opposite directions.
The competitive-threat version is weak. Two thirds of organisations still aren’t training everyone, so nobody is being left behind this quarter, and a leadership team that hears “everyone else is doing it” will correctly notice that isn’t true.
The opportunity version is much stronger. The gap between organisations on this is currently enormous and mostly invisible from the outside, which means the difference between a competitor whose marketing team ships twice as much usable work and one whose team quietly wastes a day a week is not something you’ll see in their announcements. You’ll see it in what they produce, about eighteen months later, and by then it isn’t a gap you close with a workshop.
There’s also a pattern in who moves first that’s worth naming. It’s rarely the whole organisation. It’s one function with a leader who picked a workflow, and it spreads sideways because other teams want what that team has. That’s a much easier internal case to make than a transformation programme, and it’s how most of the successful rollouts we work on actually started.
The comforting story about shadow AI is that it was a 2023 problem, that sanctioned tools have since arrived, and that it’s on its way out. The data available doesn’t support the ending.
Netskope’s platform telemetry, drawn from customer traffic between June 2025 and July 2026, found that just over half of enterprise AI users work only in organisation-managed AI applications, while roughly three in ten use only personal, unmanaged apps and a further slice use both. Their own summary of the trend line is that the shift away from shadow AI stopped around March 2026 and has since drifted slightly back the other way.[5] This is a security vendor measuring a problem it sells the solution to, and it’s observed traffic rather than a survey, which cuts both ways: harder to over-report, easier to draw a flattering boundary around.
The reason this persists after you’ve bought sanctioned tools is worth understanding, because the usual response makes it worse.
People don’t use a personal account to be difficult. They use it because the sanctioned tool doesn’t do the thing they need, or it’s three clicks further away, or they tried it once when it was worse and never went back. Responding with a stricter policy and a blocking rule moves the behaviour rather than stopping it, and it removes the last bit of visibility you had.
| Ask, in a one-to-one | If the answer is | It means |
|---|---|---|
| “What do you use AI for in a normal week?” | “Nothing really” | Either genuinely nothing, or they don’t trust the question. Ask the next one anyway |
| “What’s the last thing it got wrong?” | A specific, detailed story | They’re using it a lot and thinking about it. This person should be helping design the standard |
| “What do you not trust it with?” | “I don’t put client stuff in” | Good instinct, no rule. Give them the rule so it survives a deadline |
| “Which account are you signed in with?” | Hesitation | Personal account. Don’t react to it. This is the moment that decides whether you ever hear the truth again |
| “What would make the approved tool worth switching to?” | Anything concrete | Your actual roadmap. Shadow AI is usually a product gap wearing a compliance costume |
A discovery script for managers. Run it as a conversation, not an audit; the fourth row is the one that determines whether the other four ever get honest answers again.
The practical move is boring and it works: make the sanctioned route better and closer than the unsanctioned one, tell people plainly what may and may not go into any AI tool, and treat someone admitting to personal-account use as information you’re glad to have rather than an incident. Training is where all three of those get delivered at once, which is the least glamorous argument for it and probably the truest.
Most internal cases for AI training fail in the same way. They lead with a productivity percentage from a vendor study, a finance director asks how that number would show up in this company, and there’s no answer. The meeting ends politely.
A case that survives contact does three things instead:
It also has to be honest about the causal chain, because leadership teams have heard the transformation pitch before. Training builds capability, meaning somebody can now do the thing. Workflow design gives that capability somewhere to land, a real recurring task it slots into. Reinforcement turns it into behaviour through repetition and a manager who actually checks. Adoption is what you see when the new way of working sticks, and it’s the outcome of the first three rather than a lever you can pull on its own. Skip the middle two and you get a well-reviewed workshop and no change, which is the most common outcome and the reason budget holders are sceptical.
On what to measure, resist the metric that’s easiest to pull. Logins and active users tell you about access, not about adoption. The lens worth using runs use, then persistence, then impact.
| Section | Worked example, marketing team of six |
|---|---|
| The workflow, named | The monthly channel performance report. Currently one day of one person’s time, every month, and nobody enjoys it. |
| What good looks like | Same structure as today, every number traceable to the source export, written in our voice, ready for the director by the second working day. |
| What we’re asking for | Two half-day sessions on this specific workflow, six people, plus four weeks of a named internal owner running it live. |
| Use, at 30 days | All six have produced one report this way. Evidence: the reports exist. |
| Persistence, at 60 days | It’s still being done this way without anyone reminding them. Evidence: month two happened on time with no prompting. |
| Impact, at 90 days | The report lands on day two instead of day six, and the director acts on it before the mid-month review rather than after. Evidence: two decisions that were previously made without it. |
| What we’re not claiming | Any headcount saving. Any percentage productivity figure. Any effect outside this one workflow. |
| The kill criterion | If month three still needs chasing, we stop and say so rather than extending it. |
A filled-in ninety-day case using the use, persistence and impact lens. The last two rows are the ones that get a sceptical finance director on side, because they show you have a way of being wrong.
If nobody can name what changed, no training happened. Decide the evidence before you spend.
Given the field experiment above, the shape of what you buy matters more than the amount.
| Buy this | Be wary of this |
|---|---|
| Sessions built on your team’s own recurring work, using your own files | Generic tool walkthroughs with sample data |
| Practice at spotting wrong output, with real examples that fooled someone | Prompt libraries handed over with no practice attached |
| A named quality bar the team helped write and can test against | “Best practice” guidance nobody can check compliance with |
| Manager sessions, so the people setting standards can assess the work | Training the doers and skipping the people who review them |
| Follow-up built in, four to six weeks later, on real output | A single event with a satisfaction score attached |
Shaped around the finding that a short prompting primer did not produce judgment, and that the group given one performed worst on the task the model could not do. Framework, not measured data.
Teams can genuinely start this internally, and if you have someone with the time and the credibility to run it, start there tomorrow. Where structured training earns its money is in speed and in coverage: it shortens the learning curve, it builds the capability across a whole team rather than concentrating it in one enthusiast who then becomes a bottleneck, and it forces the work into the room instead of leaving people to translate generic advice into their own jobs afterwards. Our corporate workshops are built around exactly that, and if you want the version that keeps working after the sessions end, we’ve written about building an upskilling programme that doesn’t go stale and about moving an organisation from awareness to fluency.
Here’s where I’d start this week, and it costs nothing. Pick one team. Ask the five questions from the shadow AI section in your next round of one-to-ones. Write down what you learn about which tools are actually in use and what people don’t trust them with.
You’ll have a better business case by Friday than any slide deck could give you, because it will be about your own company.
Because the do-nothing option does not exist. Employees are already using AI at work in large numbers, and a meaningful share do not tell anyone, which means the choice in front of a leadership team is between managed use and unmanaged use rather than between AI and no AI. Unmanaged use produces costs that never appear as a line item: quality drift nobody traces to a cause, sensitive material going into tools the organisation has no agreement with, and five people separately solving the same recurring task. None of those show up in a quarterly review, which is exactly why they persist. The competitive argument is real but weaker, because most organisations still have not trained everyone.
Three things, in roughly this order. Use fragments, so different people in the same role work to different standards and nobody can say which is right. Use goes underground, because people who suspect they are doing something unsanctioned stop mentioning it, which removes the error trail. And errors stop being learnable, because a mistake caused by an unchecked AI output gets attributed to carelessness rather than to a process gap, so the same mistake recurs elsewhere. The public example worth knowing is Deloitte Australia’s assurance review for an Australian government department, which contained fabricated academic references and a fabricated court quotation and led to a partial refund of a A$440,000 contract. That was a review failure at a large firm, not an untrained junior, which is what makes it a caution rather than a reassurance.
In substantial numbers, yes, and it does not appear to be going away. Netskope’s platform telemetry covering June 2025 to July 2026 found roughly three in ten enterprise AI users working only in personal, unmanaged AI applications, with the trend away from shadow AI stalling around March 2026 and drifting slightly back. That is a security vendor measuring something it sells a solution for, so treat the framing with care, though it is observed traffic rather than self-report. The more useful finding for a manager is behavioural: people reach for a personal account when the sanctioned tool is worse, further away, or was disappointing once. Blocking moves that behaviour rather than ending it.
Do not lead with a productivity percentage from a vendor study, because the first question will be how that number would appear in this company and there is no good answer. Name one recurring workflow, state what good output looks like for it, and commit to what you will be able to show in ninety days. Measure use at thirty days, persistence at sixty, and impact at ninety, rather than logins and active users, which measure access rather than adoption. Include what you are not claiming, such as headcount savings, and include a kill criterion that says when you would stop. A case with a stated way of being wrong is far more persuasive to a finance director than one without.
The risk is not that a competitor announces something. It is that the gap becomes structural. Capability, workflow design and reinforcement compound slowly, so a team that started a year ago is not one year better, it is one year into having standards, shared prompts, a quality bar and managers who can assess AI-assisted work. That is not something a workshop closes afterwards. The second risk is that unmanaged habits harden. Retraining someone out of a workflow they built themselves and trust is considerably harder than teaching a workflow to someone who has not yet invented their own.
The five sources in this article were each verified against the organisation that produced the research rather than against secondary coverage, with sample sizes, fieldwork dates and funding stated in the text where they affect how the figure should be read. Vendor and consultancy funding is disclosed for the Netskope, ISACA and KPMG figures, and the BCG involvement in the field experiment is noted, although that finding runs against the interests of everyone who funded it. Several verified figures were deliberately left out. METR’s randomised controlled trial finding that experienced developers were slower with AI is genuinely interesting, but METR published an update in February 2026 saying the effect has likely reversed and become very difficult to measure, and citing the original number honestly would need more qualification than this article has room for. Two further figures, from the World Economic Forum’s Future of Jobs Report and BCG’s 2026 AI at Work survey, were verified and then cut to keep the research density down. One widely repeated statistic, that 78% of AI users bring their own AI tools to work, is frequently attributed to the 2026 Microsoft Work Trend Index and does not appear in it; it is a 2024 figure, and it is not used here. No claim about the proportion of unused AI licences appears anywhere in this article, because no primary source for one could be opened.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.