An answer is not an audit trail
AI digital forensics examines the records around an AI application to establish what happened during a defined event. A generated answer may describe an action that never completed. Conversely, a background job can complete after the conversation ends. The investigation needs evidence from the application and the system where the consequence occurred.
We begin with a bounded question, a time window and authorised access. Examples include whether a tool changed a customer record, which identity accessed a document, whether an external destination received data, or whether a deployment changed before the first suspicious event. The result may establish a cause, narrow competing explanations or identify a question the retained evidence cannot resolve.
An evidence map for the investigation
| Layer | Records to request where available | Question |
|---|---|---|
| Application | Run/session IDs, release identifiers, prompt-template versions, orchestration logs | Which configuration handled this request? |
| Retrieval | Source versions, document IDs, access checks, ingestion history | Which material entered the context? |
| Tools and identity | Tool arguments/results, delegated principal, scopes, approvals | What authority was exercised? |
| Destination | Audit events, record versions, delivery logs, job completion | What actually changed or left the system? |
| Provider | Customer-visible usage and audit exports; support response | What can the provider substantiate? |
Preserve context, not just content
For each collected item, record the source, collector, collection time, relevant time zone, method and any transformation. Keep original exports separately from working copies, restrict access and use integrity checks appropriate to the agreed process. A screenshot can document an observation but may omit request identifiers or the underlying state. An export can also be incomplete because the queried time range or account permissions were wrong.
Collect the minimum material necessary for the question. Prompt and retrieval records may contain personal data, customer content or credentials. Evidence handling, access and retention are agreed separately from the public enquiry. Do not paste raw logs into the brief builder.
Work with uncertainty explicitly
Classify a conclusion as directly observed, supported by correlation, a hypothesis requiring verification, or unresolved. Record plausible alternative explanations and the evidence needed to distinguish them. Re-running a prompt can test a proposed failure condition, but cannot recreate missing historical evidence or guarantee identical output.
A provider-managed model is not a disk image that every customer can acquire. State which provider records are accessible, which need a support request and which are unavailable. The investigation must not promise recovery of deleted conversations or visibility into model internals.
Useful outputs for engineering and leadership
Engineering needs the affected boundary, reproducible validation conditions and specific changes to test. Leadership needs the confirmed impact, uncertainties, decisions and next update. An agreed deliverable can include both views, an evidence register and a sequence of events. Whether particular evidence or reporting is suitable for litigation is a matter for the relevant legal team and agreed forensic process.
For Microsoft environments, Microsoft’s investigation guidance describes correlating AI activity through identity, time and resource context. Our Copilot response service scopes the records actually available in your tenant.
Discuss your situation
Build an incident brief with non-sensitive context, or request a scope discussion. We confirm people, access, fees and availability before work begins.
