The morning after, nobody was accountable
Meta description: The boundary between machine preparation and human judgment has to be explicit — the decision stays with the person who knows the business.
Miklos Roth
The incident review was professional. I have read several of these documents now, and they share a structure: a timeline, a technical analysis, contributing factors, corrective actions. The model’s output is analyzed. The vendor’s documentation is referenced. The team’s procedure is examined. Every element receives its share of explanation, and the explanation is fair, and at the end of the review, the decision that caused the harm belongs to no one. It emerged — from a model, a process, a set of circumstances. Emerged is the word these reviews use for decisions nobody took.
AI can summarize, compare, draft, analyze patterns, and propose alternatives at remarkable speed. But the decision — especially the consequential business decision — must remain with the person who understands the business, its customers, its constraints, and its purpose.
What do incident reviews currently assign?
Reviews assign factors. The model was overconfident; the training data had a gap; the reviewer was under time pressure; the procedure was ambiguous. All true, and all factors. What I notice is the grammar: factors contributed, but no verb of decision attaches to a person. Somewhere in the timeline, a human accepted the model’s output and sent it onward — that acceptance was a decision, and in the reviews I read, it appears as “the output was processed” or “the case proceeded.” The review assigns everything except the one thing accountability requires: a name attached to a choice.
How does “the model suggested” move the name off the decision?
The boundary I draw: a business owner or specialist develops an integrated understanding through years of work. An AI assistant can provide valuable information about components of that picture, but it does not possess the company’s tacit knowledge, its relationships, its constraints, or accountability for the outcome. AI may do substantial preparatory work. Qualified people must interpret results and own consequential business choices.
I do not present this position as abstract superiority. It comes partly from my own mistakes. There was a period when I relied too heavily on AI and pushed my own limits. I had to recognize that being enthusiastic about a tool is not the same as using it in a way that consistently improves judgment.
The phrase does quiet work. Before the model, the decision chain was short: a person assessed and decided. Now there is a middle step — the model suggested — and the suggestion changes the moral geometry of the chain. The person who acted on the suggestion experiences themselves as having followed a recommendation, not as having decided. The person who built the recommendation is a vendor, far away. I have watched responsibility drain into that middle step, and I have noticed that organizations rarely notice the drain, because each individual act of deference is small and reasonable.
Why do principles fail on the morning after?
The false explanation is policy: the company needs a responsible-AI framework, published principles, an ethics board. I have read these frameworks. They are eloquent about values — fairness, transparency, human oversight — and silent about names. Principles operate at the level of the institution; incidents happen at the level of a person and a Tuesday. On the morning after, the principles provide language for the press release and nothing for the question the harmed party actually asks, which is: who decided? A principle cannot answer that. The framework was never built to.
What pattern marks a real accountability gap?
The pattern I use is the reversal test. When an AI-assisted outcome is good, I watch who claims it: the sponsor presents it, the team celebrates it, the leader cites it. Names attach readily to success. When the outcome is bad, I watch the names detach — the outcome becomes systemic, multifactorial, emergent. The asymmetry is the finding. In an organization with real accountability, names attach symmetrically, because the decision has the same owner in both directions. In the gap, ownership is a function of the outcome’s valence, which is to say ownership does not exist.
The appropriate comparison is not a contest between human and machine. A good AI system can provide a person with an extraordinary range of information and ideas. The practical question is how much support improves performance while keeping human responsibility intact.
[AWAITING MIKLOS — scene: second scene]
What question must have a person’s name?
Accountability frameworks multiply categories: roles, responsibilities, committees, oversight bodies. The categories are why the question gets lost. The question itself is almost embarrassingly concrete — it picks one consequential decision and demands one name — and I have watched it cut through a governance architecture that had absorbed months of work. The architecture was not wrong. It was answering a different, easier question.
The question that locates the boundary in practice: how much AI support improves this person’s performance while keeping their responsibility — and their judgment — intact?
PR excerpt (80–120 words, Miklos’s own words): AI can summarize, compare, draft, analyze patterns, and propose alternatives at remarkable speed. But the decision — especially the consequential business decision — must remain with the person who understands the business, its customers, its constraints, and its purpose. I do not present this as abstract superiority; it comes partly from my own mistakes, from a period when I relied too heavily on AI. Being enthusiastic about a tool is not the same as using it in a way that consistently improves judgment.
By Miklos Roth — AI marketing consultant, Budapest. More: his profile.
Neighbors: 10, 26, 35.