Why training-led approaches don't produce capability, and what the research says actually does.
In two years, AI use at work has gone from a curiosity to a mainstream activity. McKinsey's State of AI in 2025 puts 88% of organisations using AI in at least one business function, up from 78% the year before. 76% of employees now use AI at work, up from 30% in 2023. By any measure of adoption, the shift is real and fast.
What hasn't shifted at the same pace is organisational capability.
Only one-third of organisations have begun to scale their AI programmes (McKinsey). Most are still stuck in piloting and experimentation. The 80% of employers who've committed to AI upskilling (WEF) are mostly running training programmes that don't translate into on-the-job change. Adoption is real. Capability isn't being built at anything like the same rate.
This piece is the first of three on what it actually takes for an organisation to build AI capability, written for leaders, business owners, and the senior people whose job it is to make this happen. The argument runs across all three pieces and is worth stating up front:
AI capability is an organisational design problem, not a training problem.
The next two pieces cover the five organisational competencies that change how AI actually gets used, and how to write an AI strategy that holds up against contact with reality. This week's piece is about why training-led approaches don't build capability, and what does.
The training-programme reflex
When AI arrives in an organisation, the default response from most leadership teams looks similar across industries and sizes. It runs roughly like this:
- Buy a tool (Microsoft Copilot, ChatGPT Enterprise, Claude for Work, an industry-specific AI platform).
- Roll it out to staff with onboarding.
- Run an "AI literacy" training programme.
- Maybe hire a Head of AI.
- Wait for capability and productivity gains to materialise.
This is a reasonable response. It's also why most AI programmes underdeliver.
The premise underneath the response is that AI capability is fundamentally about individual skill: that if you train enough people on the tools, capability follows. The premise is wrong, and the data shows it. 80% of employers have committed to upskilling. The number of organisations producing measurable AI-driven business change is a fraction of that.
The reason: capability isn't built by training. It's built by changing how work happens.
What capability actually means here
When researchers and consultants talk about "AI capability," they usually mean one of two things:
- Tool fluency: the ability of individuals to use AI tools effectively
- Organisational capability: the ability of the organisation to consistently produce better outcomes by using AI
These are different problems. Tool fluency is taught. Organisational capability is built.
Tool fluency (knowing how to prompt Copilot, draft with Claude, query an AI agent) is genuinely useful, and McKinsey's data is right that workers in occupations with explicit AI fluency requirements have grown sevenfold in two years. But tool fluency on its own doesn't produce organisational capability. It produces individuals who use AI for their own tasks. What it doesn't produce is workflows redesigned around what AI can now do, or quality structures that catch AI errors before they reach customers, or decision-making about where AI should and shouldn't be used.
Capability is the workflow, the structures, and the decisions, not the individual skill.
This is why "we trained 5,000 people on Copilot" rarely shows up in next year's productivity figures. Trained individuals are operating inside workflows designed before AI existed. The work they now do faster still has to pass through review processes, sign-off chains, and downstream tasks that weren't redesigned around the new pace. The bottleneck moves from generation to everywhere else.
The organisations seeing real productivity returns from AI aren't the ones who've trained the most staff. They're the ones who've redesigned the work itself.
What organisational design actually involves
When I work with leadership teams on this, the conversation usually starts with training and ends with five other things they hadn't planned on changing.
The actual components of building AI capability across an organisation:
Workflow redesign. Where in current processes does AI now do something a human used to do? Where does that change the shape of the workflow downstream? Most organisations are running the old workflow with AI inserted into one step of it. Real capability comes from redesigning the whole sequence around what AI now changes.
Quality structures. AI-generated work fails differently from human work. The errors are plausible-sounding, professional-looking, and subtly wrong. Existing review processes were built for the kinds of mistakes humans make. They don't reliably catch the kinds AI makes. Without redesigned quality structures, AI-generated output finds its way into customer-facing material, financial reports, compliance content, and client deliverables with errors nobody noticed.
Decision rights. Who decides where AI is appropriate? Who decides where it isn't? In most organisations these decisions are happening implicitly, by individual choice, with no organisational view of which decisions actually matter. A clear decision-rights structure prevents both over-use (AI in places it shouldn't be) and under-use (AI not used in places where it should be).
Learning systems. AI tools change every quarter. Workflows you set up six months ago may already be outdated. Annual training cycles can't keep up. Organisations that build real capability replace one-off training with continuous, peer-led, work-embedded learning. The training function itself has to redesign, not just produce more training.
Hiring and progression criteria. What competencies do you now hire for that you didn't two years ago? What gets weighted in promotion decisions? Most organisations haven't updated either. The criteria they're using to identify high-performers were calibrated for pre-AI work.
Cultural permission to experiment. AI capability is built through use. Use requires permission to try things, get them wrong, and try again. In organisations where every output has to be perfect, AI capability stalls: staff use AI quietly, don't share what works, don't surface what doesn't. In organisations where experimentation is expected and supported, capability compounds across teams.
These six are interdependent. You can't redesign workflows without updating quality structures. You can't update quality structures without clear decision rights about where AI is and isn't appropriate. You can't make any of this work without cultural permission to try it. Most organisations are working on one or two of these and treating the rest as separate problems for separate teams.
Why training-led approaches keep underdelivering
The combined effect of training without redesign is what shows up in the McKinsey data: lots of activity, little capability.
A training programme rolled out into an unredesigned workflow produces individuals who can use AI in their own task, but who hit organisational friction every time the AI's output crosses into someone else's task or process. The friction discourages use. Use discourages sharing. The capability never accumulates.
A training programme without quality structures produces AI-generated output that goes through existing review processes, which weren't built to catch AI-style errors. Either the errors get through, or staff stop trusting AI for anything that matters, or the review process becomes such a bottleneck that AI's speed advantage disappears.
A training programme without decision rights produces a thousand small individual decisions about where to use AI, with no organisational view of which decisions matter. Some teams over-use, some under-use, and almost no one is making the calls that actually change business outcomes.
A training programme without continuous learning systems produces capability that's already obsolete by the time it's been rolled out across the organisation. Annual upskilling programmes lose to quarterly tool changes every time.
In every case, the training itself isn't wrong. It's necessary. It just isn't sufficient. Treating training as the answer is treating a symptom.
What this means for how leaders should think about it
The leaders who get genuine returns from AI aren't running larger training programmes. They're treating AI capability as a system-design challenge, one that needs work across workflows, structures, decisions, learning systems, hiring criteria, and culture. Training is one input among six.
Two practical implications, both available to any leader reading this:
First: stop treating AI capability as primarily an HR or training problem. The training function has a role, but most of the work is operational and strategic, not training-led. The leadership team that owns AI capability building usually needs to include operations, technology, the relevant business unit head, and HR, not just HR.
Second: when designing an AI programme, work backwards from the workflow. Pick one workflow. Map what changes when AI does part of it. Identify what else has to change downstream (review processes, sign-offs, hand-offs, decision points). That's the design work. Once you've done it for one workflow, you can do it for the others. Until you've done it for one, no amount of training will produce real capability.
What's next
The next piece walks through the five organisational competencies that actually change how AI gets used, translating the things individuals need into the structures, processes, and design choices that make those individual capabilities work at scale.
The week after: a practical piece on how to write an AI strategy. Not the marketing-deck version. The actual working document leaders use to make decisions about where AI fits, what gets invested in, and how progress gets measured.
This is the work, in my experience, where organisations either compound their advantage or quietly fall behind. Few are doing it well yet. The ones who figure it out in the next eighteen months will be the ones whose AI investments produce the returns leaders are expecting from them.