AI Can Do Your Basics. That's the Problem.

Part 2 of the Cracks in the Ladder series. Everyone's investing in AI capability. Almost nobody is investing in protecting the human intelligence that AI is quietly replacing. Not because they don't care - it's because the skills most at risk are the ones hardest to identify.

Not every routine task is just a routine task.

Some tasks produce an output and that's all they do. Formatting a report, scheduling a meeting, cleaning a dataset into a standard template – these are pure production. Offloading them to AI is pure efficiency gain. But other tasks that look equally routine are doing double duty: they produce an output and they train the person doing them for harder work downstream. The accountant who manually explores raw data before it's been cleaned is not being inefficient – she's building the anomaly-detection instinct she'll need when the model breaks. The manager who writes a first draft from a blank page rather than editing an AI-generated one is doing generative cognitive work – struggling with structure, discovering what she actually thinks – that builds the capacity to think clearly under pressure. The developer who writes a database query from scratch rather than accepting an AI-generated one is learning where performance bottlenecks hide – intuition she'll need when a production system fails at scale.

This distinction – between tasks that only produce and tasks that also develop – is the one that matters most for how organisations adopt AI. And it is the one almost nobody is making.

The management theorists Rittel and Webber drew a useful line: some problems are "tame" (well-defined, solvable, with clear criteria for success) and others are "wicked" (ill-defined, entangled with values and politics, with no clean solution). AI is built for tame problems. It matches or beats humans on pattern recognition, classification, information retrieval, statistical analysis – anything with defined inputs, a known method, and a verifiable output. Wicked problems – sensing that a high-performer is about to quit before any data says so, knowing when to push back on a CEO's pet idea – are where human judgment earns its keep.

The natural move is to hand the tame work to AI and free up humans for the wicked work - navigating ambiguity, breakdown, conflict, novelty, and non-routine judgment that AI cannot handle independently. That logic has a hole in it. The tame zone is not just where humans produce. It is where they learn how to handle wicked work.

The development pipeline hidden inside routine work

Ask any senior leader how they developed judgment – the ability to read a room, to sense when data is telling the wrong story, to hold a line in a difficult negotiation – and they won't point to a training program. They'll point to years of doing work that didn't feel like judgment-building at the time.

Consider what this looks like in practice. A junior HR business partner joins a large manufacturing organisation. For the first two years, much of her work is routine: she sits in on disciplinary hearings as a note-taker, processes employee grievances using standard templates, runs exit interviews with a checklist, compiles monthly attrition dashboards from raw HRIS data, and drafts talking points for her senior HRBP's skip-level meetings.

None of this looks like leadership development. It looks like administrative work – the kind AI could do faster and with fewer errors. But each of these tasks is quietly building something.

The disciplinary hearings teach her to read body language under tension – who is defensive, who is withholding, how the room shifts when the union representative speaks. The grievance processing exposes her to patterns: which managers generate repeated complaints, which sites have systemic issues dressed up as individual conflicts. The exit interviews, repeated dozens of times, train her ear for the difference between someone leaving for a better offer and someone leaving because they've given up. The attrition dashboards force her to sit with raw numbers long enough to notice when a spike in one department doesn't match the story the department head is telling. The talking points require her to anticipate what a skip-level audience will push back on – which means modelling someone else's perspective before she's senior enough to be in the room herself.

Three years in, a crisis lands on her desk. A plant head with thirty years of tenure is accused of creating a hostile work environment. The evidence is ambiguous. The union is watching. The plant's production targets are already strained. Her senior HRBP is on leave. She has to make a call: escalate formally and risk a protracted standoff that disrupts production, or attempt an informal resolution and risk being seen as soft on misconduct.

She handles it. Not because she was trained for this specific scenario, but because every signal she needs to read – the union's posture, the plant head's defensiveness, the workforce's mood, the gap between what the data says and what people are saying – is a signal she has been unconsciously learning to read for three years through work that looked like paperwork.

Now imagine if she had heavily relied on AI to begin with - where AI handled the note-taking, the grievance templates, the exit interview summaries, and the attrition dashboards. Her visible output from day one might have been indistinguishable from – even better than what it was when she did everything manually. But when the crisis comes, the instincts aren't there. She wouldn’t have sat across from sixty people and learned to hear what isn't being said. She'd been reviewing AI-generated summaries, not doing the cognitive work that builds the human intelligence.

Michael Polanyi called this tacit knowledge: we know more than we can tell. It has to be built through direct, repeated, often tedious experience with the work itself. Nonaka and Takeuchi's knowledge-creation model identifies the mechanism: through socialisation (learning by shared experience with others who already have the tacit knowledge) and internalisation (learning by doing, until explicit knowledge becomes embodied skill). Both modes require the person to be immersed in the work. You don't learn what "anomalous" looks like by reading a report on anomalies. You learn it by making sense of the data and situations yourself, over months and years, until the pattern recognition becomes second nature.

And crucially, the cognitive scientist Robert Bjork showed that the difficulty is the point. Conditions that make learning slower and more effortful in the short term – what he termed desirable difficulties (focus on desirable) – produce more durable and transferable knowledge in the long term. When AI removes the struggle, it removes the mechanism through which the human intelligence is built.

The zone where AI replaces human effort is the same zone where humans develop the intelligence that AI cannot replicate. The next logical question is: can AI not accelerate expertise development rather than undermining it?

Ericsson's research on deliberate practice suggests that what matters for building expertise is not sheer volume of repetition but structured engagement at the edge of competence, with immediate feedback. An AI that surfaces anomalies for a junior HRBP to interpret, rather than either doing the full analysis or leaving her to grind unaided, could theoretically compress the novice-to-competent transition. But this only works if the AI is deliberately designed as a training scaffold – if someone has decided in advance which parts of the task the human must still do (the ones that build human intelligence), and which can be safely handed off (pure AI efficiency gains).

That decision requires knowing which parts of the task carry developmental weight – and that is exactly the kind of knowledge that is hardest to identify and articulate. The senior HRBP doesn't know her grievance-pattern instinct exists until a situation trips it. The finance director who spots a buried risk in a balance sheet can't list the micro-signals she reads – she just reads them.

Which means that even in organisations that want to get this right, the default is not careful scaffold design. It is wholesale task delegation: "AI can do this, so let AI do this." The efficiency logic is driving adoption – not always because organisations don't care about development, but because the development logic requires solving a problem that no one has yet solved.

The system rewards the wrong choice – even when people know better

There is a third dimension that makes this harder still. Even when an individual manager does recognise that a routine task is building something in her junior, the system around her rarely rewards protecting it.

James March's work on organisational learning draws a sharp distinction between exploitation (refining and optimising what you already know) and exploration (building new capabilities whose payoff is uncertain and delayed). Organisations systematically favour exploitation. The incentives are immediate, measurable, and legible to leadership: faster turnaround, lower headcount, higher output per person. Exploration – keeping a junior analyst on manual reconciliations because it builds judgment she'll need in five years – has none of those properties. The payoff is invisible, the timeline is long, and the manager who protects the developmental task will look less efficient than the one who automated it.

AI adoption amplifies this bias. When the tool is available and the efficiency gain is obvious, the manager who says "I want my junior to keep doing this by hand for developmental reasons" is swimming against a current that runs through performance metrics, budget reviews, and leadership expectations. She may be right. But the system is not set up to reward her for being right – at least not until the cost of the missing judgment shows up years later, in a crisis that lands on someone else's desk.

This means the problem is not only that organisations can't identify which tasks carry developmental weight – it is also structural: even when they can, the incentive architecture pushes toward automation anyway.

Why L&D's current toolkit was built for a different problem

Organisations are spending significant resources helping people build AI capability – tools training, prompt engineering workshops, AI fluency programmes, co-pilot rollouts. That investment makes sense. But there is almost no corresponding investment in the other direction: protecting the human intelligence that AI adoption quietly displaces.

The asymmetry is understandable, because the traditional L&D toolkit was designed for named competencies: what a person in this role should be able to do. The effectiveness depended upon how well learning translated into performance outcomes. The definition of competency – the knowledge, skills, and abilities to do a task – has now become fuzzy because of how much AI can deliver. It is easier for L&D to build expertise in how to use AI tools and continue measuring for performance (hello, Kirkpatrick!), rather than sitting with the uncomfortable question of identifying what constitutes as human intelligence for each role.

This is not an indictment of L&D's tools. It is a recognition that those tools were built for stable role architecture which might be insufficient in protecting expertise and human intelligence through a period of rapid task redistribution and role redesign.

Closing the gap requires new instrumentation – ways to trace judgment backward from the moments where it visibly matters and map which upstream tasks and exposures built it over time. That instrumentation is what Part 3 will build.

Where to start before the framework arrives

You don't need a full methodology to begin. Here are three questions worth asking before any task in your function gets automated:

1. If a junior person in this role never did this task because AI handled it from day one, would they still develop the judgment to make the high-stakes call five years later?

   
   

2. Is there a senior practitioner who currently does this task and whose judgment the team relies on in ambiguous situations – and if so, can you trace any of that judgment back to the years they spent doing this exact work?

   
   

3. If you automate this task, who loses the opportunity to watch an experienced person do it – and what were they learning by watching?

If any of these questions surface a "probably not" or an "I'm not sure," that task needs protection, not automation. Not forever, and not entirely – but it needs someone thinking about it as a developmental asset, not just a production cost.

That's the minimum. Part 3 will build the full framework.

Photo by Yaroslav Shuraev: https://www.pexels.com/photo/man-and-woman-with-prosthesis-in-futuristic-surroundings-7688749/

Notes

1. Rittel, H.W.J. & Webber, M.M. (1973). "Dilemmas in a General Theory of Planning." Policy Sciences, 4(2), 155–169. Rittel and Webber distinguished "tame" problems (well-defined, with testable solutions) from "wicked" problems (ill-defined, value-laden, with no definitive solution). The distinction maps onto where AI performs well and where human judgment remains necessary.

2. Polanyi, M. (1966). The Tacit Dimension. University of Chicago Press. Polanyi's claim – "we know more than we can tell" – explains why expertise built through hands-on practice cannot be replicated by AI outputs or codified in training materials. It also explains why identifying the capabilities at risk from AI substitution is so difficult: they are, by nature, resistant to explicit articulation.

3. Nonaka, I. & Takeuchi, H. (1995). The Knowledge-Creating Company. Oxford University Press. Their SECI model identifies four modes of knowledge conversion; socialisation (tacit-to-tacit through shared experience) and internalisation (explicit-to-tacit through learning-by-doing) are the two modes most disrupted when AI substitutes for hands-on practice.

4. Bjork, R.A. (1994). "Memory and Metamemory Considerations in the Training of Human Beings." In J. Metcalfe & A. Shimamura (Eds.), Metacognition: Knowing About Knowing. MIT Press. Bjork's research demonstrates that effortful, slower learning conditions – which he termed "desirable difficulties" – produce more durable retention and better transfer to new contexts than easier, faster conditions.

5. Ericsson, K.A., Krampe, R.T. & Tesch-Römer, C. (1993). "The Role of Deliberate Practice in the Acquisition of Expert Performance." Psychological Review, 100(3), 363–406. Ericsson's research shows that expertise develops through structured practice at the edge of competence with immediate feedback – raising the possibility that AI could enhance developmental practice if deliberately designed to do so, but also implying that unstructured AI offloading bypasses the conditions expertise requires.

6. March, J.G. (1991). "Exploration and Exploitation in Organizational Learning." Organization Science, 2(1), 71–87. March showed that organisations systematically favour exploitation (refining current efficiencies) over exploration (building new capabilities with uncertain, delayed payoffs). AI adoption amplifies this bias: the efficiency gains from task automation are immediate and measurable, while the developmental costs are invisible until a crisis reveals the missing judgment years later.