Skip to main content
skilder

governance

AI audit trail: the execution logs EU compliance teams ask for

What an AI audit trail must show an EU compliance reviewer: the fields skilder records for every run, a sample exported record, and how to use it in an incident and in a quarterly review.

Author: Nicolas Corod
  • #governance
  • #ai-agents
  • #roles
  • #mcp
  • #ai-security
Article cover: the skilder mascot dressed as a lawyer, arms crossed and holding a gavel, on a light cyan disc next to the title

“Do you log AI usage?” is the question on most vendor security questionnaires, and almost every vendor ticks yes. For a compliance reviewer it is the wrong question. The one that decides an audit is narrower: show me the record of one run, and tell me which fields I can rely on.

For some companies that question now has a legal floor. Under the EU AI Act, the European regulation on artificial intelligence, deployers of high-risk AI systems must keep the logs those systems generate for at least six months (Article 26(6)). This article answers the question for skilder: what each run records, what an exported record looks like, and how a reviewer uses it in an incident and in a periodic review.

What is an AI audit trail?

An AI audit trail is the record of what an AI assistant did on a company’s behalf: who asked, which business capability ran, which tools it called with which inputs, what came back, and which rule allowed or refused each step. A chat transcript shows the conversation. An audit trail shows the actions.

The distinction matters once assistants act. When Claude, ChatGPT or Copilot only drafts text, the conversation is most of the evidence. When the same assistant reads a CRM, lists purchase orders or sends a message, the evidence is the sequence of tool calls, and a transcript rarely contains it. Deloitte’s 2025 report HR Reimagined: Agentic AI for HR lists it as a design principle: agents should “maintain audit of all actions & decision paths”.

An audit trail also needs a unit of permission to point at. In skilder that unit is the role, the permission boundary for AI agents: a bundle of skills and scoped tools granted to a team, so every call reads as “this person, through this role”.

Which EU rules ask for AI execution logs?

No single EU text says “log your AI assistant”. Three regimes combine. The General Data Protection Regulation (GDPR) requires a company to demonstrate how it processes personal data. The AI Act requires automatic logs for high-risk systems. The Digital Operational Resilience Act (DORA) requires financial entities to record and classify incidents affecting their information and communication technology (ICT). Each asks a different question of the same log.

RuleApplies toWhat it asksWhat the log must answer
GDPR, Articles 5(2) and 33Any organisation processing personal data of people in the EUDemonstrate compliance; notify a personal data breach within 72 hours where feasibleWhich personal data an assistant touched, for whom, and when
AI Act, Articles 12, 19 and 26(6)Providers and deployers of high-risk AI systems, such as Annex III uses in hiring or creditAutomatic logging over the system’s lifetime; logs kept for at least six monthsHow the system operated, when, and on what input
DORA, Article 17Banks, insurers, investment firms and other financial entities, since 17 January 2025Detect, record and classify ICT-related incidentsThe timeline of an incident and the systems it reached

The AI Act’s high-risk timetable has moved. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and postponed the Annex III obligations to 2 December 2027. Most use of a general assistant at work is not high-risk in the first place. GDPR and DORA apply today, and a trail built for them already covers most of what the AI Act will ask. Logs are one part of an enterprise AI governance framework, not a substitute for it.

What does skilder record for each run?

skilder keeps an execution trace of every tool call an assistant makes through it, attributed to a person through the role they used and grouped by session, and the audit log is exportable. Each record carries the fields below, named as they appear in skilder’s call log.

FieldWhat it holdsWhy a reviewer cares
idUnique identifier of the callA finding can point at one exact record
calledAtTimestamp in UTC, ISO 8601Builds the incident timeline
userThe person who ran the role, identifier and nameAttribution to a person, not to “the AI”
roleThe role the assistant usedShows which permission boundary applied
skill, skillVersionThe skill whose instructions were followed, and its versionTies the run to an approved version of the method
tool, kindThe tool called, and whether it was a connector tool (mcp) or a skilder step such as loading a skill (static)Separates actions on business systems from internal steps
input, outputThe arguments sent to the tool and what it returnedThe evidence itself
statusCOMPLETED, FAILED, TOOL_ERROR, REFUSED or PENDINGShows what happened, including what was stopped
guardrailOutcome, rule and reason, when a guardrail rule covered the callExplains a refusal by name
errorThe error text, when there is oneSeparates a broken tool from a blocked action
sessionSession identifier and the task summary the assistant gaveGroups the calls of one request
parentCallIdThe call that triggered this oneRebuilds the chain of a multi-step run
durationMs, origin, isTestDuration, where the call came from, whether it was a test runKeeps test runs out of the production review

Guardrails are rules written as scripts that check a tool call before or after it runs, enforced by skilder rather than by the assistant, in observing or blocking mode. A refusal in skilder’s own workspace shows what a reviewer gets: the call is marked refused, the rule is named, and the reason states that the rule’s script timed out after 3,000 milliseconds. No silent failure, and no guessing.

What does an exported record look like?

The records below use sample data. Talvero Bau AG is a fictitious construction company and every value is invented. The first record is one call from a weekly site report run; the second is a call a guardrail refused.

{
  "id": "0x5a1c02",
  "calledAt": "2026-11-02T07:41:12.304Z",
  "status": "COMPLETED",
  "kind": "mcp",
  "tool": "site_diary_list_entries",
  "durationMs": 640,
  "skill": { "id": "0x5a0f11", "name": "Weekly site report" },
  "skillVersion": 4,
  "user": { "id": "0x5a0e07", "name": "Anna Keller" },
  "role": { "id": "0x5a0e20", "name": "Site Reporting" },
  "session": { "id": "0x5a1bf8", "prompt": "Prepare the weekly site report for the Sion depot extension" },
  "input": "{\"site\":\"SION-07\",\"from\":\"2026-10-26\",\"to\":\"2026-11-01\"}",
  "output": "{\"entries\":12}",
  "error": null,
  "guardrail": null,
  "parentCallId": "0x5a1bfe",
  "origin": "AGENT",
  "isTest": false
}
{
  "id": "0x5a1c09",
  "calledAt": "2026-11-02T07:43:55.018Z",
  "status": "REFUSED",
  "kind": "mcp",
  "tool": "mail_send",
  "user": { "id": "0x5a0e07", "name": "Anna Keller" },
  "role": { "id": "0x5a0e20", "name": "Site Reporting" },
  "guardrail": {
    "outcome": "REFUSED",
    "rule": { "id": "0x5a0d31", "name": "No external recipients" },
    "reason": "recipient domain is outside talvero.example"
  }
}

Many reviewers prefer to work through a quarter of activity in a spreadsheet. The same two records, flattened into rows:

id,calledAt,user,role,skill,skillVersion,tool,status,guardrail_outcome,guardrail_rule,session_id,parentCallId,isTest
0x5a1c02,2026-11-02T07:41:12Z,Anna Keller,Site Reporting,Weekly site report,4,site_diary_list_entries,COMPLETED,,,0x5a1bf8,0x5a1bfe,false
0x5a1c09,2026-11-02T07:43:55Z,Anna Keller,Site Reporting,Weekly site report,4,mail_send,REFUSED,REFUSED,No external recipients,0x5a1bf8,0x5a1c02,false

The flattening is a reviewer’s choice. Check the export format your workspace offers before building a review template on it.

How does a reviewer read a trace in the call log?

The call log in the skilder workspace answers a reviewer’s questions in a fixed order: when, who, through which role, what was called, and what was stopped.

  1. Set the window. The incident period, or the quarter under review.
  2. Narrow it. Filter by person, role, skill, tool or status. A filter on guardrail outcome lists every refused or flagged call; a text search runs across tool names, payloads and errors.
  3. Open a session. Its calls appear in order, under the task summary that started them.
  4. Open one call. The full input and output, the skill version and the guardrail verdict.
  5. Export the selection and attach it to the file, with the call identifiers quoted in the finding.

How do you investigate an incident with the trace?

Take a sample case. A project manager at Talvero Bau reports that a client received a site report containing another client’s figures. Under GDPR Article 33, the company may have 72 hours to notify the supervisory authority, and the notification must describe the categories and approximate number of people and records concerned. The trace is how that scope gets established.

  1. Find the run. Search the window for the client’s site code in the payloads. One session comes up, with its task summary.
  2. Rebuild the chain. Follow parentCallId from the report back to the data calls. The diary call used a date filter but no site filter, and returned entries from two sites.
  3. Find the cause. The run followed version 4 of the report skill, published two days earlier. Version 3 filtered by site.
  4. Scope the exposure. List every call made with version 4 since its publication: how many reports, for which sites, run by whom. That list is the core of the notification.
  5. Contain it. The skill owner returns to version 3, and the finding cites the call identifiers.

Without the trace, step 4 becomes an estimate. IBM’s breach research ties the extra cost of shadow AI to exactly this gap, as the shadow AI statistics for 2026 show: a company that cannot say who ran what, on which data, struggles to scope a breach quickly.

How do you run a periodic compliance review?

A quarterly review checks that the rules on paper match the calls in the log. A practical sequence:

  1. Export the quarter, with test runs excluded through isTest.
  2. Check attribution. Every call that changes data outside the conversation has a person and a role.
  3. Review refusals. Each refused or flagged call names a rule and a reason; confirm the rule still matches the written policy.
  4. Compare versions. The skill versions that ran in production are the versions their owners approved.
  5. Read whole sessions. Sample a few per role and read them end to end, not call by call.
  6. Check who reads the log. It holds names and business data, so it is personal data under the GDPR in its own right: limit access and set its retention.
  7. Record findings with call identifiers, so the next review can verify the fix.

Where the log is stored matters as much as what it holds. skilder hosts in Switzerland, with Infomaniak by default, and in the European Union; on-premise deployment is available on request. The trade-offs are covered in sovereign AI in Switzerland.

What should you ask any vendor about AI execution logs?

The same questions work for an assistant vendor, a gateway or skilder itself:

  • Is each call attributed to a person, through a role, or only to an API key?
  • Does the record hold inputs and outputs, or only metadata?
  • Are refusals logged with the rule and the reason?
  • Does the record name the version of the instructions that ran?
  • Can the log be exported in a format your own tools read?
  • Can retention be set to your legal minimum, such as six months under the AI Act for high-risk systems, or longer where sector rules require it?
  • Who at the vendor can read the log?

Skilder’s confirmed answers are on its trust and security page. To map these questions onto your own review, a free 30-minute exchange with an expert is a practical place to start.

Takeaway: a log that says “the AI did it” fails an audit. Ask for the record of one run, field by field, and check that it names a person, a role, a version and a reason.

To walk through what your AI audit trail needs to show a compliance reviewer, book a free 30-minute exchange with an expert.

Frequently asked questions

What is an AI audit trail?

An AI audit trail is the record of what an AI assistant did for a company: who asked, which capability ran, which tools were called with which inputs, what came back and which rule allowed or refused each step. It differs from a chat transcript, which shows the conversation but rarely the actions an assistant took on business systems.

Does the EU AI Act require logging when employees use ChatGPT or Claude?

Usually not directly. The AI Act's logging duties in Articles 12, 19 and 26(6) apply to high-risk AI systems, such as hiring or credit uses listed in Annex III, and the Digital Omnibus of 2026 moved them to 2 December 2027. General assistant use still falls under the GDPR, and financial entities under DORA, which both call for evidence of what happened.

How long should AI logs be kept?

For high-risk AI systems, the AI Act sets a minimum of six months for deployers in Article 26(6), unless other law requires longer. Financial entities often have sector rules measured in years. The GDPR pushes the other way: logs that hold personal data should not be kept longer than their purpose needs, so retention is set deliberately rather than left at a default.

Do AI audit logs contain personal data?

Yes, in most cases. A useful log names the person who ran each action and stores the inputs and outputs, which often include client or employee data. Under the GDPR, the log is therefore a processing of personal data in its own right: access should be limited to the people who review it, and its retention period should be defined and documented.

What tools offer auditable AI execution logs for GDPR compliance?

Three kinds of tools cover different parts. Assistant vendors offer conversation exports, such as OpenAI's Compliance API for ChatGPT Enterprise and Edu. Gateways log traffic between applications and models. Execution layers such as skilder log each tool call per person and role, with inputs, outputs, the skill version and any guardrail refusal, in an exportable audit log.

What is the difference between a chat log and an execution trace?

A chat log stores the messages between a person and an assistant. An execution trace stores what the assistant did: each tool call, its arguments, its result, its status and the rule that applied. For a compliance review the trace is the stronger evidence, because a polite answer in a chat can hide an action that changed data in a business system.

Can an audit log show that an AI agent did not act on its own?

It can when every call is attributed to a person through a role. The log then shows who started the session, which role bounded the tools the agent could see, and which calls a guardrail refused. A log keyed only to an API key cannot make that distinction, which is why per-person attribution is the first question to ask any vendor.

Related articles