Picture a system administrator at a mid-sized local government agency. He has been running the organisation’s infrastructure for eleven years. He knows where every skeleton is buried: the legacy application that refuses to run on anything newer than Windows Server 2016, the undocumented firewall rule that keeps a critical integration alive, the network segment that was set up before anyone thought to write it down. For a decade, troubleshooting a novel fault meant sitting with the problem, reading logs, building a mental model, calling on pattern recognition accumulated through hundreds of earlier incidents.
Now he uses AI for most of that. He describes the problem in natural language, gets a diagnostic path, follows it. The fault gets resolved. Often faster than before. His manager has noticed his ticket closure rate improving. She thinks he is operating at peak performance.
What she cannot see is the rate at which he is no longer doing the cognitive work that built the performance in the first place. The mental modelling. The log-reading. The slow, frustrating process of ruling out hypotheses. He is borrowing the output of that work from an AI system, and the loan is accumulating interest at a rate neither of them can currently measure.
This is the story that very few people are telling about AI in workplaces. The conventional fear is that AI will replace experienced workers. The actual risk in most organisations, where AI is being adopted by the people already employed, is something different and more subtle. AI is changing what it means to be experienced. Not by displacing experienced people, but by lending them competence faster than they can build it, until the borrowed portion of their daily work outweighs the part they still own.
And the senior people, the experts, the trusted hands, are the ones running up the largest balances. The people you trust most are quietly accumulating the most borrowed competence in your business.
What the Deskilling Research Has Shown for Forty Years
The aviation industry has been studying this since the 1980s, when autopilot started taking over more of the flying. Pilots who flew on autopilot for most of their hours got rusty at the manual tasks autopilot was doing for them. Not just slightly rusty. Measurably worse, in ways that mattered when autopilot disengaged and the aircraft needed a human who actually remembered how to fly it.
Worth a listen: Tim Harford explored this phenomenon in depth in an episode of Cautionary Tales with Andrew Wright, tracing the autopilot dilemma through aviation disasters and drawing a direct line to why semi-autonomous vehicles may be the most dangerous category of technology we have ever deployed.
More recently, a five-year case study of an accounting firm by researchers at Aalto University in Finland found the same pattern in white-collar work. Accountants who used cognitive automation tools became progressively unable to perform the tasks the tools were doing for them. The deskilling was almost invisible to managers, because the work kept getting done. It only became visible when the firm decided to discontinue the software. The competence the staff had been borrowing was called in, and the firm discovered how little of it had ever been built.
The pattern holds in medicine. A 2025 scoping review documented measurable deskilling in radiology, pathology, and clinical decision-making among physicians using AI tools. In one striking example, the transition to AI-assisted cervical screening reduced cytologist case volumes by 70 per cent and forced consolidation from 48 laboratories down to 8. The next generation of cytologists will not have the same opportunity to develop tacit pattern recognition that built their predecessors’ expertise. The borrowing is intergenerational.
Behind these findings is a longer tradition. Polanyi’s 1966 work on tacit knowledge established the foundational point that experts know more than they can tell. The implicit knowing that constitutes real expertise gets built slowly, through engagement with hard cases. It cannot be transferred verbally. It cannot be downloaded. It cannot be borrowed. It can only be built. AI lets us produce work as if the building had happened. The work product is real. The building is not.
Why Experts Carry the Largest Balances
My initial intuition said experts should be protected from AI’s downsides because they know enough to spot AI’s mistakes. I was surprised to discover the research suggests the opposite, for two reasons.
The first is that experts have more to lose. A junior staff member using AI to do work they could not yet do alone gains a productivity boost. A senior staff member using AI to do work they used to do alone is, by definition, no longer doing that work. Over time, they lose the practice that maintains their expertise. The clinical literature calls this the use-it-or-lose-it problem, and it is well documented in domains from surgery to radiology.
There is a parallel here that any software developer will recognise immediately. The concept of technical debt describes what happens when a development team takes shortcuts to ship faster, choosing quick solutions over correct ones. The code works. The product ships. But the shortcuts accumulate. The codebase becomes harder to maintain, harder to extend, harder to understand. The debt is invisible until it becomes a crisis: a security vulnerability that cannot be patched without rewriting half the system, a performance problem that cannot be optimised because nobody understands what the original code was doing, a new hire who cannot onboard because the architecture exists only in the head of the senior developer who wrote it in 2019.
Borrowed competence is technical debt for human expertise. The shortcuts are cognitive. The codebase is a person. The accumulation is just as invisible, and the crisis arrives in the same way: not gradually, but suddenly, when something fails, and the person who was supposed to know how to handle it discovers that the knowledge has quietly atrophied. The sysadmin who has not read raw logs in eighteen months is carrying the human equivalent of a legacy codebase he no longer fully understands.
The second reason is more uncomfortable. Experienced staff have the social authority to use AI without oversight. They are the people whose work other staff trust without checking. In most organisations, with thin layers of review, the senior engineer, the lead planner or the principal analyst is often the final reviewer of their own work. When AI is doing more of that work, the verification chain that catches errors is shorter and weaker, and the person at the end of the chain is the same person who used the AI in the first place. The borrowing is unsupervised. The lender is not asking for collateral.
This is not hypothetical. A Boston Consulting Group field experiment showed that on tasks outside the AI’s capability frontier, even elite consultants with degrees from Harvard, Yale, and Oxford were 19 percentage points less likely to get the right answer with AI assistance than without. Their expertise did not protect them. In some respects, it set them up to trust AI advice that they should have questioned. The most experienced people had the largest implicit credit lines and the least scepticism about drawing on them.
The Structural Problem This Creates
The borrowed competence problem is not evenly distributed across organisations. Three structural features make it land harder in some contexts than others.
The first is bench depth. An organisation with twenty engineers has structural redundancy when AI introduces an error. An organisation with three engineers does not. The cost of one expert quietly losing competence is much higher when there is no second expert to catch the drift. Local government, primary healthcare, environmental management, mining services, and smaller professional services firms all tend to operate with thin expert benches where each senior person carries disproportionate organisational knowledge. The borrowed portion of their daily work matters more, precisely because they matter more.
The second is the profile of the work itself. High-stakes technical judgment in specialised domains, exactly the kind of work that defines senior roles in most organisations, is also exactly the work where AI training data is sparsest and where AI confidently produces plausible-but-wrong advice. The Boston Consulting Group research framed this as the jagged frontier: AI performs brilliantly inside its capability boundary and fails badly outside it, and the failure is invisible because the output looks confident and coherent either way. Senior experts working at the edge of their domain are working at the edge of AI’s domain too. This is where borrowed competence is most expensive when it is called in.
The third is workforce age. Many industries are already facing a wave of retirements that will remove substantial tacit knowledge from organisations, regardless of AI. AI accelerates the underlying problem by lending competence to people who would otherwise be building it through on-the-job experience alongside those retiring colleagues. The two trends compound. The lenders may keep extending credit; the capability is leaving the building regardless.
Three Things You Can Do This Month
There is no good news in pretending the problem does not exist. There is genuinely good news in what comes next, and I will write about that in the coming weeks. For now, three practical responses any organisation can begin immediately.
Mark which work is high-stakes and which is routine. Not every task carries the same borrowing cost. Tasks that cannot afford errors need explicit verification protocols. Routine tasks can run faster with AI and lighter checking. Most organisations currently treat all AI use the same way, which is to say, with no protocol at all. They are extending unlimited credit on undifferentiated terms.
Build deliberate practice into senior roles. The most experienced staff need protected time to do the work AI is now doing for them, periodically, even though it is slower. This is the same logic as a commercial pilot doing manual landings in a simulator, even though autopilot would do them better on most days. It is also the same logic as a responsible development team scheduling time to pay down technical debt, even when the product is shipping fine. Without that practice, the capacity to take over when AI fails atrophies invisibly. The point is to keep the borrowed portion of the work small enough that the loan can always be repaid.
Make AI use visible to the organisation. You cannot manage what you cannot see. Most organisations have no reliable picture of where AI sits in their workflows, who is using it, or what decisions are being made on its output. A simple AI register, updated quarterly, is a foundational governance step. It is also the thing that tells you how much your organisation is actually borrowing, and from whom.
None of this is an argument against AI. AI in the right hands, used in the right way at the right time, is the most powerful productivity and capability tool we have ever had access to [period]. The right hands and the right way are the parts we have not yet figured out. The cost of figuring it out badly will fall hardest on the people most of us think are the safest: the senior expert, the trusted manager, the people at the top of the professional tree.
As a trainer who spends his days signing off on competence statements for the qualifications my students earn, I find the difference between borrowed competence and built competence to be the most important question in my professional life right now. I find writing helpful in reflecting on the challenges we are facing and the importance of getting this right. Thank you for reading.
A final note on attribution. This piece was developed with Claude (Anthropic) acting as a research partner, checking articles, reminding me of where I heard or read the sources and providing me feedback on how I could improve the structure of my arguments. The argument, the structure, and the final voice are my own. I share this practice with you because, if I am going to write about the difference between borrowed and built competence, I need to show how I make that difference in my own work.