The Decision Without a Decider
How enterprise AI inherits a problem older than itself
A failure surfaces. Investigation begins. The trail leads back to a decision that must have been made — except, examined closely, no one made it.
The team followed the framework. The framework was approved by governance. Governance was advised by the consultancy. The consultancy applied the methodology. The methodology was developed against industry benchmarks. The benchmarks were derived from prior frameworks. The decision exists. The artefacts that produced it exist. The decider does not.
This pattern is older than enterprise AI. The arrival of AI does not introduce it. It inherits it, amplifies it, and adds an additional layer of opacity at exactly the point where the system was already most uninterrogable.
The recent suggestion, prominent in tech commentary, that the "biggest challenge" of enterprise AI is reproducibility — the ability to produce the same answer to the same question every time — is an instructive misframing. Determinism is presented as a technical problem awaiting a technical fix. The actual function of the corporate demand for determinism is something else entirely.
The grammar of accountability displacement
Before any code was deployed, the displacement was already in place.
A senior decision is made. Asked to defend it later, the executive does not say "I judged this to be correct." The executive says: the framework recommended this approach, the analysis supported this conclusion, the precedent from peer organisations confirmed it, the methodology required these steps. Each citation moves the locus of decision somewhere else.
The framework is not authored — it is canonical. The methodology is not chosen — it is applied. The benchmark is not selected — it is consulted. Each artefact in the chain produces decisions, but each artefact is also a structure that was built specifically to require no individual decider.
This is the corporate function of process. Process is not primarily about consistency or quality, although these are real benefits. Process is, structurally, an accountability shield. The decision becomes the output of a procedure rather than the output of a person. When the decision is later examined, the procedure is examined first. Only if the procedure is found to have been violated does an individual come into focus. If the procedure was followed and the outcome was bad, the conclusion is that the procedure must be revised — not that anyone in particular failed to think.
This grammar predates AI by decades. Strategy reports cited as the basis for restructuring decisions. Six Sigma processes cited as the basis for quality decisions. Risk frameworks cited as the basis for capital allocation decisions. Methodology after methodology, each functioning as a place to deposit the decision before anyone is asked who, exactly, decided.
Robodebt: a case study in determinism without a decider
The Australian Government's Online Compliance Intervention programme — known publicly as Robodebt — was operated from 2016 to calculate and recover alleged welfare overpayments from approximately 470,000 citizens. The mechanism averaged annual reported income into fortnightly periods and matched the result against welfare payments received. Where the algorithm computed a discrepancy, a debt was raised and pursued.
The system was deterministic. The same inputs produced the same outputs every time. Reproducibility was not the issue.
The issue was that the algorithmic premise — averaging annual income into uniform fortnightly buckets to identify overpayment — produced false debts at scale. People with seasonal work, casual employment, or income that varied month to month were systematically miscalculated as having received more than they were entitled to. The recipients of these debt notices were, by design, presumed to owe what the system had calculated until they could prove otherwise. The burden of disproof fell on people whose lives the demand had already destabilised. Multiple deaths were subsequently linked to the programme, including several confirmed suicides where the debt notice was a factor.
When the failure became impossible to ignore politically, the explanations followed a recognisable shape. The system had calculated. The algorithm had identified. The matching had produced. The outputs were technical, the consequences were administrative, the burden was on the recipients. Where the doctrine of "the system decided" met the question of "who decided to deploy this system, and to keep deploying it after the warnings", the answer fragmented across years of meetings, briefings, and advice notes that no single party had owned in full.
The Royal Commission of Inquiry, reporting in 2023, found that the scheme had been unlawful from the outset and that senior officials and ministers had been aware of legal advice raising serious doubt about its validity. Accountability that should have surfaced in 2016 surfaced in 2023 — seven years and an unknown number of destroyed lives later.
What Robodebt demonstrated was not that automated systems can fail. That has never been controversial. What it demonstrated was that automated systems are particularly suited to a specific organisational need: the production of decisions that, when they fail, cannot be traced to a decider in time for the consequences to be prevented. The determinism of the system is not the bug. It is the feature that makes the displacement work.
Reproducibility, restated
Return to the argument that reproducibility is the great challenge of enterprise AI.
A reproducible system produces the same output every time. This is, on its face, a virtue. But reproducibility carries no information about correctness. A reproducible system can be reproducibly wrong. A reproducible system can encode a flawed premise and apply that flawed premise to every subsequent input with perfect consistency. The output is consistent. The output is also, in many cases, catastrophically misaligned with reality.
What reproducibility does provide is something different: defensibility. When the system gives the same answer every time, the answer can be cited. When the answer can be cited, the answer can be relied on. When the answer can be relied on, the answer can be acted on. And when the action turns out to be wrong, the chain of citation supplies the artefact that procedural defence requires: this was the system's output, applied as designed, in line with the methodology, consistent with prior runs.
This is why the demand for reproducibility is loud in corporate AI discourse, and the demand for verifiable correctness is comparatively quiet. Correctness requires a human to take responsibility for evaluating each output against ground truth. Reproducibility does not. Reproducibility allows the output to be treated as fact and the human to be repositioned downstream of it.
The repositioning is the part the discourse rarely names. The expert is no longer the author of the analysis. The expert is the quality reviewer of the model's analysis. Domain knowledge is described as "more valuable than ever" — not because expertise is being elevated, but because expertise is being redirected. Authorship moves to the model. Verification moves to the human. Accountability, formerly held by the expert, now floats between the two — accessible from neither side when the consequences arrive.
The asymmetric direction of investigation
When a system fails, the investigation moves in a predictable direction. Downward.
The technician on shift is examined first. Then the engineer who approved the release. Then the auditor who signed off. The processes are reviewed. The training records are pulled. The shift logs are scrutinised. Where the failure can be located within an individual's actions or omissions, the inquiry tends to settle there.
Investigations rarely run upward with the same intensity. The budget decision two years prior that reduced the team headcount by thirty per cent. The executive directive that fixed an impossible delivery timeline. The board-approved metric that incentivised the shortcut. The strategic choice to defer modernisation in favour of return-on-equity for the next quarter. These decisions sit upstream of the failure. They are usually causally closer to it than the technician's actions on the day. They are also, structurally, harder to investigate. There is no shift log for a budget meeting. There is no training record for a strategic priority. The artefacts that exist for senior decisions — minutes, board packs, advice notes — are written to be unfalsifiable. They record decisions in language that distributes them across the room.
This asymmetry is not the result of bad faith on the part of investigators. It is the result of where the artefacts are dense and where the artefacts are sparse. Investigation flows toward the dense documentation. The technician's day is densely documented. The executive's strategy is sparsely documented. The investigation goes where there is something to read.
AI determinism amplifies this asymmetry. The model's output is densely logged. The decisions to deploy the model, to define its training data, to set its objective function, to accept its known failure modes, are usually not. Investigations after AI-related failures tend to focus on the model — its prompt, its tuning, its outputs, its hallucinations — and rarely surface the executive decision that placed the model in the path of the consequence in the first place. The dense documentation crowds out the sparse documentation. The decisions that mattered most are the ones least available to be examined.
The argument that AI's principal challenge is reproducibility is best read as an unintended confession.
Reproducibility addresses the problem of decisions that cannot be defended. It does not address the problem of decisions that are wrong. The corporate utility of an output that can be reliably cited is large. The corporate utility of an output that is reliably correct is also large, but it is harder to produce, requires sustained expert engagement, and does not yield the procedural cover that the citable-but-wrong output yields. Given the choice, the system gravitates toward the cheaper of the two.
Determinism is presented as solving the AI problem. It is, in fact, the AI version of an older corporate operation: the production of decisions whose decider cannot be located when the consequences arrive. The framework, the methodology, the benchmark, the consultancy report — and now the model — each one a structurally similar artefact that allows the decision to exist while the decider does not.
The novelty in the AI case is that the displacement now operates faster, scales further, and adds a layer of opacity that even the procedural artefacts of the past did not match. A board can read a strategy report. A board cannot meaningfully audit a model. The artefact has retreated one further step from interrogation, and the decider one further step from accountability.
The next time an organisation announces it has "deployed AI for decision-making", the question worth asking is not whether the model is accurate, fair, or aligned. Those are real questions, and they are also the questions the system is set up to absorb without producing accountability.
The harder question is the one the system is structured to evade.
Which decision, exactly, can no longer be traced to a human decider?
If the answer is "many", the deployment is not a technical achievement. It is an accountability operation. The model is performing the function the framework, the methodology, and the consultancy report performed before it — only faster, more opaquely, and with fewer fingerprints left behind.