Every organisation adopting AI is drawing, whether it knows it or not, a boundary: which decisions pass to the machine and which remain with people. The thesis of this article is that the boundary exists regardless — the only choice is whether to draw it deliberately or let it emerge by inertia, one default at a time. And boundaries that emerge by inertia have an unpleasant property: nobody approved them, so nobody answers for them.
The invisible delegation
Algorithmic delegation rarely arrives with a signature. It arrives by accretion. A system suggests, then pre-fills, then decides except for exceptions, then decides. At each step human review thins, because reviewing costs and the system “works”. Eighteen months on, nobody in the business can say where the suggestion ends and the decision begins.
The phenomenon has a technical name — automation bias, the tendency to trust automated output beyond the evidence — but its root is organisational, not psychological: if human review is not designed, measured and defended, it goes extinct. The human oversight the EU regulation requires for high-risk systems reflects exactly this understanding: oversight is not a gesture, it is an architecture (EU AI Act, Article 14). The UK has reached the same ground by its own route: the Data (Use and Access) Act 2025 rewrote the UK GDPR’s provisions on automated decision-making, permitting more of it while making safeguards mandatory where decisions carry legal or similarly significant effects (Data (Use and Access) Act 2025); the ICO consulted on draft guidance interpreting those provisions in the spring of 2026 (ICO, 2026).
The risk is not that the machine decides. It is that nobody ever decided to let it decide.
A criterion in three axes
Where should the boundary run? Three axes order the problem better than any list of use cases.
Reversibility. A decision reversible at low cost tolerates the system’s error: correct it and learn. An irreversible decision — closing a line, dropping a strategic supplier, refusing credit that triggers an insolvency — requires the last mile to stay human, with the evidence in front of them.
Legibility of the error. Some errors show immediately: a wrong demand forecast collides with the warehouse. Others stay invisible for a long time: systematic discrimination in hiring raises no alarm; it produces only absences. The quieter the error, the more delegation must be accompanied by active measurement, not by trust.
Value of the exception. Systems optimise the typical case. But in many decisions the value sits in the exception: the anomalous customer who anticipates a market, the weak signal that contradicts the time series. Where the exception is worth more than the average, algorithmic standardisation destroys value while it cuts costs.
Crossing the three axes produces a map: full delegation where the error is reversible, visible and poor in exceptions; assisted delegation where one axis inverts; explicit human reserve where all three do. The map must be written, approved and revised — it is board business, as we argue in our analysis of board accountability.
Three terrains where the criterion is tested
Credit. Automated creditworthiness assessment is the textbook case: an error barely reversible for the applicant, barely legible for the lender (the rejected disappear from the data), rich in exceptions that matter (the atypical but solid profile). Not by chance it sits among the domains the EU regulation treats as high-risk. Here full delegation is not an option, and human review must be substantive — with the power and the information to overturn the outcome, not a stamp at the end of a queue of files.
Pricing. Dynamic price revision is usually reversible and its errors read in the sales figures: a natural candidate for full delegation. But with a caution on the exception axis — price communicates positioning, and an algorithmic oscillation that works at the margin can damage the perception of the brand. The delegation holds if the system operates within corridors commercial leadership has approved.
Recruitment. Automated CV screening concentrates all three axes on the wrong side: the error is invisible (the excluded are never seen), barely reversible (the candidate has gone elsewhere), and the exception is precious (the unusual path that brings exactly what is missing). It is the terrain where de facto delegation is most widespread and least declared — the ICO chose it as the first setting for its expectations on automated decision-making (ICO, 2026) — and where an audit of the real review rate almost always brings surprises.
The three recurring errors
The first: confusing human presence with oversight. A reviewer who approves 99.8% of automated outputs at the end of the day is not supervising: he is signing. Oversight is measured by the rate of informed intervention, not by the presence of a name at the bottom.
The second: drawing the boundary once and for all. Systems change with updates, data drifts, uses spread to neighbouring cases. A boundary drawn for the system of 2025 can be absurd for its 2027 version — in either direction.
The third: drawing the boundary without those who will live on it. Policies written far from operations produce two outcomes, both bad: they are ignored, or they are followed to the letter where the letter is wrong. The people who work the process know where the system stumbles; that knowledge is worth more than any framework.
The productivity paradox
Those who fear that retaining decisions means surrendering efficiency are reading the data backwards. McKinsey’s State of AI survey (2025) shows that more than 80% of organisations using generative AI report no tangible impact on enterprise-level results, and that the factor most correlated with impact is the redesign of workflows — undertaken by only 21% (McKinsey, 2025).
The correct reading: value does not come from replacing human decisions with automated ones, but from redesigning the process around a considered division of roles. Execution accelerates; the choice of what to accelerate, and how far, remains the work of senior leadership. It is the heart of what we call decision capital: the more abundant execution becomes, the scarcer the decision becomes.
Who defends the boundary
An unguarded boundary moves on its own — always in the same direction, because every organisational pressure pushes towards more delegation: it costs less, scales better, does not argue back. Explicit counterweights are therefore needed: thresholds beyond which the decision travels back up, periodic audits of the actual rate of human review, and an owner of the boundary who is not the person who bought the system.
Finally, there is the question of the record. Every delegated decision should leave a reconstructable trace: which system contributed, with what data, who reviewed, who could overturn. Not only because both the European framework and the UK’s reformed rules push in this direction for the systems that matter — but because without a trace the boundary is unverifiable, and an unverifiable boundary is a statement of intent. Organisations that document discover in months what the others discover in court.
On this terrain, independence of judgement is everything. Whoever sold you the platform is not the right person to tell you which decisions to withhold from it. Our institute’s Executive AI Advisory work exists for exactly this: sitting beside those who decide, with no incentives tied to adoption.
What this means for decision-makers
Step out of the “adopt or resist” logic: it is the wrong question. The right question is where the boundary runs, and who approved it. Three concrete actions: take a census of the decisions already delegated in practice, including those never formalised; classify them on the three axes — reversibility, legibility of the error, value of the exception; bring the map to the board and turn it into policy with a named owner.
Well-designed algorithmic delegation is a strength: it frees human attention for the decisions that deserve it. Delegation that emerged by inertia is merely a liability waiting for its incident.