What AI makes you gain shows immediately. What it makes you lose does not.

An organisation deploying AI sees a quick, real, measurable gain. Tasks are done faster, documents come out, analyses arrive. None of that observation is false.

What the observation leaves out is the other half of the transaction: what the organisation stops practising in the meantime. That half appears on no dashboard, because it produces no event. It shows up later, all at once, at the moment judgement is required.

The myth of the net gain

A productivity gain reads over a short period and a narrow scope: this file, this team, this quarter. A loss of capability reads over a long period and a wide scope — and it reads only if someone thought to measure it, which nobody does.

The result is not an accounting lie, it is an asymmetry of observation. The organisation sees what it gained because gains produce deliverables. It does not see what it lost because loss produces nothing — until the day it produces an incident.

It is also an asymmetry of accountability. Whoever decides on the deployment is assessed on the gain; whoever discovers the loss is not yet in the role.

Cognitive debt

The term comes from Mohammad Hossein Jarrahi, professor of information science at the University of North Carolina at Chapel Hill, who defines it as the hidden, accumulating loss of skill, judgement and capacity that occurs when an organisation automates in ways that reduce practice, learning and ownership (Cognitive World).

The analogy with technical debt is exact on one point: you borrow time, and the borrowing is often rational at the time. It misleads on another: technical debt can be observed by opening the code, cognitive debt only by putting someone in a position to decide unassisted.

You gain something in the short term that may end in losses in the long term.

Jarrahi describes three forms the bill takes: operational fragility, degraded quality, and blurred accountability. All three are alike in that none is attributed to AI when it arrives.

A signal from medicine

The clearest observation does not come from business. A multicentre observational study published in The Lancet Gastroenterology & Hepatology measured, across four Polish endoscopy centres, the adenoma detection rate in colonoscopies performed without AI assistance, before and after the tool was introduced in those same centres.

The rate fell by six points (Lancet Gastroenterology & Hepatology). The clinicians had not changed; what had changed was what they had grown used to no longer doing themselves. The authors see in it the first real-world observation of automation-induced deskilling tied to a patient outcome.

Two cautions apply before transposing. This is a highly specific clinical act, measured by an indicator of its own; nothing licenses deriving a percentage for another activity. And assisted colonoscopy detects more than unassisted colonoscopy: the tool does what it is asked to do.

What the case demonstrates is narrower, and sufficient: an acquired professional capability, practised daily by experienced clinicians, can degrade measurably within a few months of assistance. That is not a hypothesis about the future of work. It is a measurement.

What erodes, and what holds

BCG surveyed seventy C-suite and senior executives on this phenomenon. One in two says they already observe it in their organisation, and more than six in ten expect it to be a material threat within three to five years (BCG).

The most uncomfortable point in that study is not the volume, it is the overlap: the skills those leaders name as most critical to long-term performance are exactly the ones they see eroding.

• Judgement and decision making. Deciding on incomplete information, and owning it.
• Understanding and framing the problem. Knowing what the question is before looking for the answer — which AI never does for you, since it answers the one it is given.
• Creative thinking. Producing the option that was not on the list.

Conversely, the study identifies relatively protected skills, and what they have in common is worth noting: they are relational skills, not productive ones. Empathy and active listening, which sustain the mentoring AI quietly displaces. Leadership and social influence, which make structured disagreement possible where AI converges. Motivation and self-awareness, which push people to interrogate an answer rather than accept it.

In other words: what holds is what requires someone on the other side.

Protecting friction rather than eliminating it everywhere

The conclusion Maeliom draws is not to slow deployment down. It is to stop treating all friction as a defect to be fixed, and to decide explicitly where it stays.

An organisation already knows how to do this elsewhere. It runs evacuation drills it hopes are pointless, backup restore tests that serve once, double sign-offs that slow things down. Nobody calls that lost productivity: they call it maintaining a capability.

• Name the decisions that are not delegated, and write them down — not on principle, but because an unwritten rule gives way in the first difficult quarter.
• Keep unassisted exercises on critical tasks, at regular intervals, the way a backup is tested.
• Require framing before tooling: the question asked is a deliverable in itself, and it is the one that disappears most quietly.
• Measure capability, not only output. Nobody today can answer “what would we still know how to do without these tools”, and that is precisely the missing question.

None of these moves is technical. All of them assume you have settled what the organisation wants to remain capable of — which is a question of positioning before it is a question of tooling. That is the ground Maeliom Consulting works on: deciding what matters before instrumenting it, rather than the reverse.

The level where this is actually decided

An organisation does not lose its capacity for judgement as an organisation. It loses it one person at a time — and first among those who have not yet had time to acquire it.

That is the subject of the second part of this series: AI should accelerate a skill, never replace it — and the real risk is not where it is being looked for.

Common questions