16 July 2026
Last time, this series changed seats and looked at skills-based L&D from the employee’s side of the profile. Sitting there, something became hard to ignore: almost everything the series has promised — profiles that stay current, feedback that compounds, career paths that assemble themselves from skills data — quietly assumes a layer of machinery doing work no L&D team could do by hand. That machinery has a name everyone is either overusing or avoiding: AI.
So this week the series asks the question directly. Not “is AI coming for L&D” — that debate has been stale for two years — but something more useful: in a skills-based operating model, what is AI actually good for, what can it not carry, and how should a sensible L&D team divide the labour in 2026?
The Assumption Hiding in Every Skills Conversation
Go back through this series and count the moving parts a skills-based organisation depends on: a taxonomy kept current as roles evolve, thousands of profiles refreshed as people learn, gaps mapped against strategy, content matched to gaps, practice generated at the right difficulty, feedback captured in the flow of work. Now imagine maintaining all of that manually. The honest answer is that nobody does — and nobody ever intended to.
This is the assumption hiding inside every skills-based pitch deck: the model only scales because inference, matching, and generation are automated. That is not a criticism. But it does mean the AI question is not an optional add-on to a skills strategy. It is the load-bearing wall, and it deserves the same scrutiny the series has given taxonomies, measurement, and the business case.
What AI Is Genuinely Good At
Three jobs stand out, and they happen to be the three that were breaking L&D teams before automation arrived. The first is inference — reading the exhaust of everyday work and learning activity and suggesting what it says about capability, so profiles drift towards accuracy instead of towards fiction. Paired with analytics that surface the patterns, this is what turns a skills database from a survey snapshot into something closer to a living picture.
The second is matching: person to gap, gap to content, content to moment. This is pure pattern-work across more variables than any human curator can hold, and it is where the “course catalogue” posts of this series quietly get their answer. The third is generation. AI-assisted course creation has collapsed the cost of producing a first draft of learning content from weeks to hours — which is precisely what makes the skills-first content library, built against the taxonomy rather than the calendar, economically possible for teams that are not enterprise-sized.
What AI Cannot Carry
Now the other column of the ledger. AI cannot decide what capability the business actually needs next year — that is strategy, and outsourcing it to a model means optimising towards last year’s patterns. It cannot supply the trust this series wrote about from the employee’s seat; an inference engine that quietly rescores people’s profiles without explanation is the fastest way to lose the honesty that makes the data worth having.
And it cannot replace judgement at the moments that matter. A model can flag that someone’s evidence looks thin; only a human conversation — the kind that structured evaluations and feedback exist to hold — can establish what is actually true and what should happen next. The pattern across every failure story of the past two years is the same: the organisation automated the judgement and kept the admin, when it should have done the opposite.
The Division of Labour That Actually Works
Put the two columns together and a clean rule emerges: let AI propose, let people decide. AI drafts the profile update; the person confirms it. AI suggests the learning path against a training goal; the employee and manager commit to it. AI surfaces who is ready for a stretch move; the career conversation decides whether the person wants it. Every “propose” saves hours; every “decide” protects trust.
The teams getting this right in 2026 are noticeably unglamorous about it. They talk less about AI strategy and more about which specific decisions stay human. That list — written down, shared with employees, honoured in practice — is doing more for adoption than any capability the technology itself has shipped this year.
The Question to Ask Before You Buy Anything
So when the next AI-for-skills demo lands in your inbox, skip the feature tour and ask one question: for each thing this automates, who was doing it before, and who checks it now? If the answer to the first half is “nobody — it simply wasn’t being done,” that is genuine capacity you are buying. If the answer to the second half is “nobody,” walk away.
Everything this series has argued — the taxonomy, the measurement, the practice loops, the feedback rhythm, the employee’s ownership of their own picture — depends on machinery that works and people who trust it. In 2026 you can, at last, have both. But only in that order.
