When the Decision Goes Wrong, They Won't Come for the Algorithm

Let me start with something that doesn't get said often enough.

You cannot sue a bot. You cannot court martial an algorithm. You cannot hold a piece of software to account in a tribunal, a coroner's inquest, a regulatory review, or a courtroom.

When an AI-influenced decision causes harm, the accountability doesn't follow the technology. It follows the human who acted on it.

Think about the decisions being made right now, across every sector, in every kind of organisation, that are being shaped — partly or significantly by AI outputs.

A clinician reviewing a diagnostic recommendation from an AI-enabled imaging system. A social worker whose case management platform has flagged a risk level. An HR manager who used an AI screening tool to shortlist candidates. A construction project manager who signed off a safety assessment generated by a digital twin.

None of those people did anything wrong. They used the tools available to them. They applied their professional judgement. They made a decision.

But here is the question I want you to sit with.

If that decision is reviewed in two years' time by a regulator, a coroner, a tribunal, a civil claim — what is your record of what you knew, what the tool told you, what you considered, and what you decided?

Not what you remember. Not what a colleague recalls. Not what can be reconstructed from an email chain.

What is the record, created at the time, with the information you had available of the judgement you exercised?

For most people, the honest answer is: there isn't one.

The tools arrived faster than the documentation standards. AI is being used to inform consequential decisions across healthcare, education, financial services, social care, construction, and public sector delivery — and the professional frameworks for recording human judgement in AI-influenced contexts simply haven't kept pace.

The gap isn't in the technology.

The gap is in the record.

There is a concept in law that matters here. It is called contemporaneous evidence. In simple terms: a record made at the time, before the outcome was known, before anyone was looking. Those records carry significant weight in review contexts because they cannot be shaped by hindsight. They reflect what the decision-maker actually knew and actually concluded — not what they wish they had known, given what happened next.

When AI influences a decision, three problems emerge that a contemporaneous record directly addresses.

Retrospective reconstruction. When something goes wrong, organisations piece together what happened from memory and incomplete email trails. That picture is assembled after the fact, under pressure, knowing the outcome. It is a very different thing from a record made in the moment.

Diffused responsibility. When AI is involved, it can become genuinely unclear who owned the decision. Without a clear record of who exercised judgement and on what basis, accountability disperses — and dispersed accountability in a review context is not protective. It is dangerous.

Hindsight bias. Post-incident reviews judge past decisions by present outcomes. A decision entirely reasonable given what was known at the time may look very different assessed against what subsequently happened. The only protection is a record that establishes what information was available — and what was not — at the point the decision was made.

None of this is about blame. It is about the reality of what happens when AI-influenced decisions are reviewed — and the fact that the humans involved are almost always underprepared for that moment.

The technology will not protect them. The vendor whose platform they used will not protect them. This is not simply about compliance frameworks, governance policies, or responsible AI principles. Because beneath all of those sits an individual professional making a decision — and whether they will be able to explain that decision if they are ever asked to.

What protects people is evidence. Specifically: what they knew, what they considered, and what they decided — recorded at the time they decided it.

The algorithm won't be in that room.

The software vendor won't be in that room.

The prompt won't be in that room.

The model won't be in that room.

The individual who exercised judgement will.

And one day, the question may not simply be:

"Why did you decide that?"

But:

"What evidence exists of what you knew, what you considered, and why you decided it?"

Because when decisions are reviewed, memory is rarely enough.

Records matter.

Louize Clark is the founder of AI Policies UK. She works with organisations and individuals navigating AI governance, data compliance, and the human accountability questions that sit underneath both. aipolicies.uk