Accountability Without Commitment

Something went wrong. Investigators ask who decided. The answer, under examination, is nobody — or everybody. Opinions were exchanged, models ran, emails circulated, but no moment exists where a named authority converted reasoning into an irrevocable commitment. Accountability is demanded anyway, and lands on whoever is most visible or most junior.

In software systems

Systems emit recommendations, alerts, and draft states that trigger action without a commitment event. A monitoring dashboard turns red; an on-call engineer acts. The action is logged as a configuration change, not as a decision with evidence and authority. A pipeline deploys because “the build passed” — no DeploymentApproved event names who accepted the residual risk.

When failure occurs, postmortems reconstruct intent from Slack threads and CI logs. The engineer who clicked deploy absorbs blame, though the actual commitment — accepting production risk — was never formalised. Conversely, models auto-escalate without human commitment and organisations claim “the algorithm decided,” obscuring that no principled boundary crossing occurred. The log contains operations, not decisions. Accountability queries have no anchor event to attach to.

In human organisations

Committees discuss without deciding. Action items accumulate; ownership is diffuse. “The team agreed” substitutes for a named authority and timestamp. When a product launch fails, the VP asks which decision was wrong — and discovers the launch proceeded through momentum, not commitment. Marketing assumed product sign-off; product assumed executive approval; executive assumed the committee had decided.

Performance reviews punish individuals for outcomes that emerged from uncommitted collective drift. The organisation has accountability rituals — RACI matrices, approval chains — but the actual commitment moments are skipped under schedule pressure. The matrix describes who should decide; the record shows who was blamed after the fact. Decisions exist as atoms in theory and as fog in practice.

In socio-technical systems

AI-assisted workflows make the gap visible. A loan application receives a model score, a rules pass, and an analyst comment — then status changes to approved with no discrete commitment event linking authority, evidence snapshot, and threshold applied. The customer receives funds. Default occurs. Legal discovery requests the decision record and receives a folder of inputs with no committed output.

Regulators expect named accountability at catastrophic severity. The organisation points to the analyst; the analyst points to the model; engineering points to product policy. The socio-technical stack produced an outcome without a commitment atom. Retrospective attribution becomes political: whoever lacks documentation wins the dispute by default, or whoever lacks power loses it. Neither outcome reflects who actually held authority at the moment that mattered.

Structural causes

Epistemic tier collapse

Outcomes are recorded and acted upon without crossing a commitment boundary. Status fields, dashboards, and “approved” flags imply L0 decisions that never occurred as events — only L2 inputs and operational side effects.

The commitment boundary pattern exists precisely to give accountability an anchor: every irreversible consequence traces to a named authority at a recorded moment.