AI didn't make the call — but it shaped it. That distinction is dissolving faster than most organisations are prepared for.
A mid-sized logistics firm in the Netherlands recently dismissed a warehouse worker after an AI-flagged performance review recommended termination. HR signed off. The manager countersigned. No one interrogated the model’s inputs. Six weeks later, an employment tribunal asked a simple question: who decided? Three people pointed at the software. The software, naturally, pointed at no one.
This is not a story about AI going rogue. It is a story about humans using AI to quietly redistribute blame — and discovering that tribunals, regulators, and colleagues are not particularly sympathetic to that move.
When decisions were made by individuals, accountability was uncomfortable but locatable. When they were made by committees, it diffused — but slowly, visibly, through meeting minutes and paper trails. AI introduces a third mode: decisions that feel individual because one person clicks “approve,” but are shaped by a model that aggregated the preferences, errors, and blind spots of whoever built and trained it.
This is the diffusion trap. Responsibility spreads so thin across the chain — developer, procurement lead, line manager, HR business partner — that no single node feels genuinely responsible. Everyone contributed; no one owns it. The result is not shared accountability but dissolved accountability.
In practice, this means that when things go wrong, organisations default to one of two equally useless postures: blame the tool (“the system recommended it”) or blame the individual who clicked approve (“you should have checked”). Neither produces learning. Neither produces justice.
Most AI governance frameworks invoke “human oversight” as the solution. The instinct is correct; the implementation is usually theatrical. Oversight that consists of a busy credit analyst spending forty seconds reviewing an AI-generated loan recommendation before approving it is not oversight. It is liability laundering.
Genuine oversight requires that the reviewer understand, at minimum, what inputs drove the output, what the model cannot see, and what its known failure modes are. In a well-run legal operations team reviewing AI-drafted contract summaries, that standard is achievable — the tool is narrow, the domain is bounded, the reviewer is expert. In a marketing team using a large language model to segment audiences for a sensitive financial product, it is considerably harder. The model is opaque, the reviewer may not know what questions to ask, and the timeline pressure is real.
Organisations that conflate these two scenarios in a single “AI oversight policy” are writing fiction.
Finance is having a quiet accountability crisis. Algorithmic trading, AI-assisted valuations, and credit decisioning systems now sit inside processes where individual traders and analysts still carry personal regulatory obligations. The FCA and SEC have not yet decided how personal liability attaches when a licensed individual relied in good faith on a model they were instructed to use and were not trained to audit. That ambiguity is not a gap in the future — it is a live professional risk today.
HR faces a variant of the same problem with a sharper ethical edge. Hiring tools that screen CVs, flag flight risks, or recommend performance improvement plans are making consequential decisions about people’s livelihoods. When a line manager challenges a recommendation and is told by the system vendor that the model is proprietary, and by their own legal team that documentation is limited, they are being set up to fail. The decision will be attributed to them. The inputs that drove it remain invisible.
In IT, the accountability question is often misdirected entirely. Engineers who deploy and maintain AI systems are frequently asked to bear accountability for business decisions their tools informed — because they are closest to the mechanism. This misunderstands what accountability means. Technical custodianship is not decision ownership.
Procurement teams are negotiating AI vendor contracts that say remarkably little about what happens when the model contributes to a bad outcome. Indemnification clauses are narrow. Audit rights are limited. Explainability requirements are absent or toothless. The vendor delivers the software; the liability stays with the buyer.
This is partly the vendor’s leverage in a hot market. It is partly the buyer’s failure to ask the right questions. Most procurement processes were designed for software that executes instructions reliably — not for systems that generate probabilistic recommendations whose reasoning is not fully transparent even to their creators. The contractual vocabulary has not caught up with the technical reality.
Until it does, organisations are absorbing risk they have not priced, in processes they do not fully understand, for decisions they may have to defend under legal standards that were written for humans.
Here is the uncomfortable insight that most accountability frameworks resist: the question is not “who owns the decision” in the abstract. It is “who benefits from the decision, and who bears the consequences if it goes wrong.”
In the Netherlands tribunal case, the worker bore the consequences. The organisation benefited from the efficiency of automated performance management. The vendor benefited from the contract. The manager who countersigned suffered some reputational damage. The model was retrained with a note in a changelog.
That distribution — where the least powerful party absorbs the worst outcome, and the most powerful parties absorb the least — is not an accident of AI. It is a reproduction of existing power structures through a new mechanism. Accountability frameworks that focus only on process (did someone review it? was there a policy?) while ignoring distribution (who actually got hurt? who actually gained?) will keep producing the same result.
The point of assigning accountability is not to find someone to punish. It is to create the conditions under which decisions improve. That requires honesty about who shaped a decision, who rubber-stamped it, and who was never really in a position to object. AI makes that chain longer and less visible. The answer is not shorter chains. It is better lighting.