The problem, defined

What is the accountability gap?

The distance between what an AI agent reports it did, and what can actually be shown to have happened.

The accountability gap opens when an agent can act on your behalf but cannot demonstrate what it did. It's not a shortage of logs — most agent stacks produce more logs than anyone reads. It's that the logs, the confirmations, and the summaries are all written by the same AI agent that took the action, so a customer, an accountant, or an auditor asking "did this actually happen, and was it right?" has no independent record to consult. The gap is the distance between a claim of completion and evidence a third party can check without trusting the party that produced it.

Why now

Agents act. Reports are self-written.

Chatbots produced text; the worst that could go wrong was reading it wrong. Agents produce actions: a record gets written, an email gets sent, an approval gets granted, a payment gets routed. The step from answering to doing is the step at which the account-of-what-happened stops being something a human can verify by reading a screen and starts being a claim about a change in the world. And the only thing writing that account is the agent that made the change. That was manageable when the number of AI-initiated actions per company per day was small enough to spot-check; it stops being manageable at the volume every serious deployment now aims for. Governance frameworks and audit policies were written when the AI's output was the thing to review; they are not sized to a world where the AI's actions are.

The four questions

What a business cannot currently answer.

Each of these is a question an owner, an accountant, or an auditor asks after the fact — and that most AI stacks cannot answer with evidence.

What actually ran?

Not what was requested, and not what the tool reported — which steps genuinely executed, in what order.

Under whose approval?

Which named person authorised this, and were they shown the version that was actually used?

Did the world change as reported?

The AI agent says it wrote the record. Does the record exist, and does it match?

What silently didn't happen?

The step that was skipped leaves no log line and raises no alarm. Nothing looks wrong.

What closes it

Four mechanisms, not four policies.

Enforcement

Rules inside the execution path

Your rules are what the engine runs under, not a note stored beside the action. A quote that breaks them doesn't get flagged for review — it stops, and asks you.

Binding

Approval tied to a version

A person approves a specific version, and the approval is bound to a fingerprint of it. Change a line afterwards and the approval no longer covers it.

Independent verification

Checked by something that didn't do the work

A read-only account reads your system of record and compares values, line by line — and reports Confirmed, Doesn't match — needs you, or Not confirmed yet.

A record built for questions

Answerable in one lookup

What ran, under which of your rules, approved by whom, with what result — permanent, and answerable without reconstructing events from a diary written by the protagonist.

The honest note

Closing it is architecture, not policy.

A policy document describes what should happen. It has no way to stop the action that shouldn't. The gap closes only where the mechanism sits between the agent and your systems — which is a decision made in how the thing is built, not in what it promises.

The burden of proof should sit with the system — not with you.