WHITEBOXX / GUIDE
Auditable AI: how to check
what an AI system did.
An AI system is auditable when someone who was not involved can reconstruct what it did: which data it used, which steps it took, what it produced and who approved it. This guide sets out what to record and how to test it.
What is auditable AI?
Auditable AI is an AI system that keeps enough records for a reviewer to check a result after the fact. The reviewer should be able to see the request, the data sources and their dates, the steps and tools the system used, the output, and the people who started or approved the work.
Auditability is different from explainability. Explainability asks why a model produced an answer. Auditability asks what the whole system did, so that a result can be checked, repeated or challenged.
Why it matters now.
AI agents now take steps, call tools and pass results to other systems. Each step is a point where something can go wrong. In a 2026 survey of 204 security leaders by SecurityHQ and Datadog, only 49% said they could fully reconstruct every action an agent had taken over the previous 30 days. 45% had had an AI-related incident where it was unclear who was accountable.
In an IBM study of more than 3,200 executives in Europe and the Middle East, about 30% said they expect to need real-time monitoring and audit trails of agent decisions and actions by 2027.
THE RECORD
What an audit trail should record.
| Record | What it contains | Question it answers |
|---|---|---|
| Request | Who asked, when, and the exact instructions | What was the system asked to do? |
| Data | Each source used, with its version or retrieval date | Which data was the result based on? |
| Steps | The tools called, the queries run and the calculations made | How did it get from the data to the result? |
| Output | The result as delivered, and any intermediate results | What exactly did it produce? |
| Review | Who checked or approved the result, and when | Who is accountable for it? |
| Limits | The tools, data and budget the agent was allowed to use | Could it have done something it should not? |
CHECKLIST
A checklist for auditable AI.
- 01
Give every run an ID
Link the request, data, steps and output to one record.
- 02
Store source dates, not only source names
Data changes. A result can only be checked against the data as it was.
- 03
Record calculations, not only answers
A reviewer should be able to redo the arithmetic.
- 04
Name an owner for each workflow
One person should be accountable for each workflow’s results.
- 05
Set limits before the run
Decide which tools and data an agent may use, and record those limits.
- 06
Test reconstruction
Ask someone who was not involved to rebuild a past result from the records alone.
Common gaps.
Many teams log the conversation but not the data. A transcript shows what the agent said, not what it read. Others keep only the latest version of each source, so older results cannot be checked once a source is revised.
A third gap is recording the model but not the tools. In agent systems, many errors come from the data and tool steps rather than from the model.
How Evidence OS records each run.
Evidence OS records each workflow run with its inputs, step progress and deliverables. Execution agents work within the tools and budget set for each assignment, and their results carry the sources and calculations behind them.
See what a run record contains