"Don't operate the equipment. Equip the operator." Danny Wallace, Somebody Told Me 

I read that line and stopped what I was doing. Not because it's a clever turn of phrase (though it is), but because it names, in eight words, the exact mistake I keep watching organisations make with AI.

A leadership team decides they need to "do something about AI." They convene a working group. The working group produces a shortlist of platforms. Procurement gets involved. Six months later there's an enterprise licence, a launch email, and a training video nobody watches past the two-minute mark. The equipment has been operated. The operator has not been equipped.

This isn't a piece about whether AI is a tool or an agent. I've made that argument elsewhere, and I stand by it: something that makes decisions and generates novel ideas without asking permission first is not the same category of thing as a hammer. But that argument, however true, has become a kind of comfortable distraction. Organisations can debate the philosophical status of AI indefinitely while completely failing to do the much duller, much more urgent thing, which is building the judgement of the humans who now have to work alongside it.

The failure isn't conceptual. It's budgetary. And it's happening in your organisation right now, whether or not anyone's noticed.

The tool-first reflex

Watch where the money goes in almost any AI rollout and you'll find the answer to what the organisation actually believes matters. Overwhelmingly, it goes to the equipment. Licences, seats, integrations, a shiny dashboard for the board to look at. The operator gets an onboarding email and, if they're lucky, a lunchtime webinar or a one-off training day with an outside ‘expert’ and then they are left to their own devices.

The numbers bear this out starkly. Skills England's own research found that only 21% of UK workers feel confident using AI at work (as of January 2026), and fewer than four in ten UK businesses have given staff what they'd call comprehensive AI training, despite the investment pouring into the tools themselves. Perhaps more troubling still, in my own anecdotal work with organisations across the UK and beyond, only c.50% of employees could actually check whether an AI-generated output was accurate. Not use the tool. Check whether the tool was right. That's not a training gap. That's a judgement gap, and it's the one that matters.

I've written before about the comprehension gap that opens up when nobody in the room understands what the AI has produced well enough to defend it or fix it when it breaks. This is the organisational version of the same failure, scaled up. More than two-thirds of UK firms now believe their own staff are using AI tools nobody's approved, quietly, off to the side, because the approved route was too slow, too locked down, or simply never explained why it mattered. Shadow AI isn't rebellion. It's what happens when you equip nobody and then act surprised that people find their own equipment.

None of this is really about AI, if we're honest. It's about a much older organisational habit of confusing procurement with capability. Buy the CRM, assume the sales process improves. Roll out the new curriculum, assume teaching gets better. Install the platform, assume the culture follows. It rarely does, and AI has simply made the gap between the two impossible to ignore, because the tool itself is now capable of producing plausible, confident, fluent nonsense at a speed no previous piece of office software ever could.

Engelbart and the argument that got skipped

In October 1962, an engineer at the Stanford Research Institute called Douglas Engelbart published a report called Augmenting Human Intellect that's aged into something like scripture for anyone who cares about this stuff, and is still mostly unread by the people making AI decisions today. Engelbart's whole career was built on a single, deceptively simple argument: the point of a computer was never to replace human thinking, but to augment it. His report described this as increasing "the capability of a man to approach a complex problem situation", and that phrase does a lot of work. Not replace the man. Not remove the problem. Increase his capability to approach it.

He kept saying versions of the same thing for the next fifty years, because almost nobody built what he actually meant. In a later interview, reflecting on why the world hadn't caught up, he put the stakes in blunter terms: 

"We need to think about how to boost our collective IQ." Douglas Engelbart

Not individual output. Not tool adoption figures. Collective IQ, the thing an organisation actually has more or less of depending on whether its people have been equipped to think harder with the tool in front of them, or simply to click it faster.

What's easy to miss, reading Engelbart now, is that he wasn't describing a piece of software. He was describing a research programme, one that assumed you had to deliberately design the human side of the system with as much rigour as the machine side. He called this "bootstrapping": you don't just install the tool and hope capability follows, you build the capability alongside it, on purpose, as its own discipline. Sixty-odd years later, most organisations still skip this step entirely. They'll spend a fortune selecting the right AI platform and not one hour asking what capability, specifically, they're trying to build in the people who'll use it.

This is where I think the education sector, for all its faults, occasionally gets something right that the corporate world keeps missing. A good teacher training programme doesn't hand a newly qualified teacher a scheme of work and walk off. It builds their judgement, over years, through observation/coaching, feedback, and the slow accumulation of professional confidence. Nobody would dream of calling that "onboarding." Yet that's precisely the model most organisations use for AI, a technology that requires more judgement, not less, than a lesson plan.

Engelbart's insight, translated for 2026, is uncomfortable but simple. If your AI rollout increases the capability of the machine and leaves the capability of the person exactly where it was, you haven't augmented anything. You've just added an expensive intermediary between the person and the problem they were already trying to solve.

Illich and the tool that quietly takes over

If Engelbart tells you what equipping the operator should look like, Ivan Illich tells you what happens when you don't bother. His 1973 book Tools for Conviviality draws a distinction that's become more useful, not less, in the fifty years since he wrote it. Illich separates tools that extend a person's autonomy and judgement, which he calls convivial, from tools that quietly take that judgement away and replace it with dependency on whoever designed the tool.

Illich's test isn't about what the tool can do. It's about who remains in charge of the meaning of the work once the tool is in use. As he put it, when a person is mastered by his tools rather than the other way round, "the shape of the tool determines his own self-image". Read that back in the context of an organisation rolling out AI without building judgement alongside it, and it stops being a 1970s philosophy quote and starts being a fairly precise diagnosis. The person hasn't been augmented. Their sense of what good work looks like has quietly been outsourced to whatever the model produces.

“This blindness is a result of the broken balance of learning. People who are hooked on teaching are conditioned to be customers for everything else. They see their own personal growth as an accumulation of institutional outputs, and prefer what institutions make over what they themselves can do. They repress the ability to discover reality by their own lights. The skewed balance of learning explains why the radical monopoly of commodities has become imperceptible. It does not explain why people feel impotent to correct those profound disorders which they do perceive.” Ivan Illich

Illich made a version of this same argument in his earlier book, Deschooling Society, which we reference in the first Edufuturists book, Pick ‘n’ Mix Education, and it lands just as hard on corporate training as it did on classrooms. "Most learning is not the result of instruction," he wrote, and anyone who has sat through a mandatory AI awareness webinar will recognise exactly what he meant. Instruction can hand someone a manual. It cannot, on its own, hand them the judgement to know when the manual is wrong. That only comes from the "unhampered participation in a meaningful setting" he set against it, actual practice, on actual work, with actual stakes attached, which is precisely what a launch email cannot provide and a compliance course rarely bothers to.

This is, I think, the sharpest possible test for whether an AI rollout is equipping the operator or simply operating them. Ask the person using the tool: if the AI's answer looked wrong, would you know, and would you feel able to say so? If the honest answer is "I'd probably just go with it," you haven't built a convivial tool. You've built a radical monopoly on judgement, and it happens to run on a server you don't own.

There's a particular danger in professional and educational contexts here, because both run on exactly the kind of fluent, confident-sounding output that AI produces effortlessly and that humans are, frankly, terrible at second-guessing. We're trained to read confidence as competence. A hedging colleague gets questioned; a smooth one gets believed. AI has no relationship whatsoever between its confidence and its accuracy, which means an unequipped operator is walking directly into the one blind spot most likely to catch them out.

What equipping actually looks like

It would be easy to leave this at the level of theory, so let's get concrete, because there's UK evidence of organisations doing this properly, in enough detail to see exactly what "equipping" costs and looks like in practice, rather than the sanitised version that survives into a press release.

KPMG UK's problem started before it had chosen anything. From late 2022 onwards, its own staff, roughly 17,000 people across audit, tax and advisory, were already using generative AI tools on their own initiative, before the firm had settled its governance, its guidelines, or a single training pathway. That's the honest starting point most organisations won't admit to: the operators equip themselves first, badly, and the organisation finds out afterwards. What KPMG did next is the more interesting part. Rather than commission a training course, it made itself what it calls "client zero", testing every tool and every governance approach on its own staff before it ever advised a client to touch it. It built an AI-assisted learning coach, Spark, directly inside Microsoft Teams, so guidance sat where the work already happened rather than in a separate portal nobody opens twice. It ran an internal apprenticeship, open to staff at every grade including senior leaders, where the test wasn't a certificate but a live piece of business analysis. One participant used it to cut a piece of analytical work from 200 hours to 4. That's not a training statistic. That's a person who was equipped to do something they genuinely could not do as well before, on a task with a client's name attached to it.

What KPMG resisted is just as telling as what it built. It didn't reach for a single AI 101 course rolled out top-down. It ran a leadership-sponsored series called Summer of AI, repeated every year since 2023, precisely so the message never became a one-off memo, and it let "tea and talk" sessions spring up informally among staff rather than mandating them from HR. Judgement, in other words, was built socially, in the flow of real work, with visible senior buy-in, not delivered once from a slide deck and considered done.

Contrast that with something much smaller and considerably less glamorous. Cast Consultancy is a 90-person construction consultancy with offices in London and Edinburgh, working in a sector that has historically lagged well behind on digital transformation. Rather than run generic AI awareness sessions, the kind that produce a tidy training log and no change in behaviour whatsoever, the firm picked one specific, unglamorous, high-stakes task: Gateway 2 submissions under the Building Safety Act 2022, the regulatory paperwork required before higher-risk residential buildings can proceed, where a missed inconsistency in the documentation carries genuine consequences. Around 10 to 15 project managers were trained directly against that one workflow, using a tool built for exactly that purpose, with the provider running sessions in-house so staff could raise real examples from their own live projects rather than a hypothetical demo case.

The detail that matters most here is what happened after the training technically finished. Cast Consultancy built a standing feedback loop with the tool's provider, so staff using it on real submissions kept feeding back what the tool got wrong, and that feedback changed how the tool worked. The operators weren't just equipped to use the machine. They were put in a position to argue with it, correct it, and shape it, which is Illich's convivial test in practice rather than in theory.

What both examples share is a refusal to treat "training" as an event with an end date. Equipping the operator, done properly, looks like a standing capability programme tied to a real task with real consequences, sustained by visible leadership involvement, not a launch week followed by silence. Read more about these and other examples in the UK government report, Skills for AI: What works for AI upskilling in the UK published in June 2026.

The leadership question nobody wants to answer

The question that actually separates organisations equipping their people from organisations merely operating equipment is: who, specifically, in your organisation has been given the authority to say "I don't trust this output, and I'm not using it"?

That’s not the authority to raise a ticket but the authority to override the tool, on the spot, without needing to justify it upward first. If that authority doesn't clearly exist, if it's assumed rather than named, then no amount of prompt-writing workshops will equip anyone for anything. You'll have trained people to operate the equipment more fluently. You will not have equipped an operator.

This is a leadership decision before it's a training decision, and it's the one most organisations quietly avoid, because naming who has override authority means admitting the tool can be wrong, publicly, in a way that a launch email never does. It's far more comfortable to keep the conversation at the level of "upskilling" than to sit down and decide, deliberately, what decisions in your organisation should never be handed to a model without a human who's confident enough to argue with it.

For schools and other education settings, this question has a particular edge, because the people being trained to override the tool are often also being asked to model that same judgement for the young people in front of them. You cannot teach critical evaluation of AI output to a classroom of teenagers if the adults in the building have never been given permission, or the confidence, to do it themselves.

The organisations getting this right, in my experience, tend to share one habit. They treat "we deliberately chose not to automate this" as a strategic decision worth documenting, not a failure of ambition. That's the clearest sign an organisation understands the difference between the equipment and the operator. One is optional. The other, if you want the equipment to be worth anything at all, is not.

Danny Wallace's line from the beginning works because it refuses the easy binary. It doesn't say don't use the equipment. It says the equipment was never the point. Engelbart said the same thing in 1962, in considerably more words. Illich said it from the opposite direction, warning what happens when we forget it. Sixty years and one philosophical argument later, we're still buying the machine and hoping the operator turns up equipped by accident.

They won't. Equipping the operator is not a phase of the rollout, and it isn't finished when the training video's been watched. It's the ongoing, unglamorous work of building the judgement that lets someone look at a fluent, confident answer and ask, out loud, whether it's actually right. Organisations that skip this step haven't found a shortcut. They've just quietly agreed to let the shape of the tool decide who they are.

Key Takeaways

1. The AI investment gap isn't technical, it's a judgement gap, and the statistics on staff confidence prove it
2. Buying the platform is not the same as building the capability, however large the budget
3. Engelbart's sixty year old argument about augmentation still holds - augmentation has to be designed deliberately, it doesn't follow automatically from access
4. Illich's tool test is simple and uncomfortable, ask whether the person could catch the tool being wrong, and would feel able to say so
5. Real equipping looks like an ongoing capability programme tied to specific, high stakes tasks, not a one off launch
6. The leadership question that actually matters is who has the named authority to override the tool, and whether that authority is real or assumed

This isn't a case for slowing down. It's a case for spending as much deliberate attention on the person holding the tool as we currently spend on the tool itself.