Regulation

# An incident is a poor time to start collecting evidence

When a serious incident starts, your team needs records it can retrieve. Reconstructing every AI action from memory adds a second problem.

18 June 2026 · 3 min read

Updated 28 October 2026

[← All articles](/blog)

[Respond and explain](#two-jobs)[Choose useful records in advance](#records)[Test retrieval as well as capture](#retrieval)[Preserve records people can check](#verification)[Keep reporting responsibility clear](#responsibility)[Sources and further reading](#sources)

## Respond and explain

During an ICT incident, people need to contain the problem and establish what happened. For financial entities in scope, DORA also brings incident-management and reporting duties. The relevant classification, timing and reporting route need to be understood before the incident.

This is not the ideal moment to discover that three systems disagree about the time and the only person who knows the agent setup is on leave. Their holiday was not the control failure.

## Choose useful records in advance

For an AI workflow, identify the data requests, action proposals, permissions, approvals and reported outcomes that would help an investigation. Decide what can be retained, who may read it and how long it is needed.

Keep coverage explicit. A record from a connected tool can describe what that control observed. It does not establish what an unconnected application did elsewhere. Missing coverage belongs in the investigation, not in a reassuring footnote nobody reads.

## Test retrieval as well as capture

Run a short exercise. Pick one test action and ask the incident team to find the relevant records. Check that they can distinguish a proposed action from an approved action and a reported outcome.

Time the retrieval. Check access permissions and export formats. Identify which dependencies might be unavailable during an incident. If viewing the evidence requires the failed service to recover first, that dependency deserves attention.

## Preserve records people can check

Available for evaluation in Enterprise. Athena seals records for independent verification. A reviewer with the record, compatible verifier and trusted public-key material can check the supported integrity properties offline. Coming to hosted Athena.

Keep ordinary logs and other operational evidence too. A sealed action record may answer one question very well while a network trace or recovery test answers another. An investigation benefits from the right evidence, not from declaring one format the winner.

## Keep reporting responsibility clear

Your incident owners and compliance team decide what happened, whether it meets a reporting threshold and what to tell the relevant authority. Product records support that work; they do not make those decisions.

The useful preparation is straightforward: agree the records, test retrieval, know the gaps and name the people responsible. Then repeat the exercise when the workflow changes. A reporting clock is a poor project manager.

## Sources and further reading

Originally published 18 June 2026. Rewritten on 28 October 2026.

[ESMA: Digital Operational Resilience Act](https://www.esma.europa.eu/esmas-activities/digital-finance-and-innovation/digital-operational-resilience-act-dora)

[Trust and evidence handling](/trust)

Keep exploring

[Explore Athena](/athena)[Trust and verification](/trust)[All articles](/blog)
