An audit trail is a time-ordered record of the actions taken in a system: who acted, what they did, to which object, when, from where and with what result. It exists so that events can be reconstructed afterwards by someone who was not there. An accountant uses one to trace a ledger figure to its invoice; a security team, to find which account deleted a file. In advertising, the most familiar example is the change history in an ad platform, which shows who paused a campaign or cut a budget. The newest use is the oversight of AI agents, software that now edits campaigns and spends money with no person typing the change.

What a record contains

The building block is the event record. Most designs capture the same handful of fields: an identity for the actor (a user, a service account or an agent), a timestamp, the action, the object acted on, the outcome, and often the values before and after. Some add the channel, such as the web interface or an application programming interface (API).

Three properties separate an audit trail from an ordinary log. The first is attribution: each entry has to point to an accountable actor, not just a process. The second is completeness: actions that matter cannot bypass the record. The third is integrity. A record that its subjects can edit proves little, so audit trails are kept append-only, written to separate storage or protected with cryptographic hashes, where each entry carries a digest of the one before it and any later edit breaks the chain. Retention, how long entries are kept, decides how far back anyone can look.

The design principle dates to the US Department of Defense's Trusted Computer System Evaluation Criteria, the "Orange Book", first issued in 1983 and reissued as DoD 5200.28-STD in December 1985. From class C2 upwards, it required that the system "be able to create, maintain, and protect from modification or unauthorized access or destruction an audit trail of accesses to the objects it protects." Regulated industries followed. The US Food and Drug Administration's 21 CFR Part 11, published on March 20, 1997, required "secure, computer-generated, time-stamped audit trails" for electronic records that replace paper ones. NIST Special Publication 800-92, released in September 2006, set out how to collect, store and analyse such logs.

Change history in ad platforms

Every major advertising platform keeps an audit trail of account edits, usually called change history. Google Ads shows changes to accounts, campaigns and ad groups for the past two years; "data older than 2 years won't be available," according to its help centre. The User column shows the email address of whoever made a change through the interface. Changes from automated systems may appear under names such as "ads-dartsearch-budget" or "Google Ads system", and API changes from third-party tools may show "Google Ads API" or the tool's name. In January 2026, Google added a Go To button that jumps from an entry to the edited campaign or ad group, according to Search Engine Roundtable.

The programmatic view is narrower. The Google Ads API's ChangeEvent resource returns old and new field values but only for the past 30 days, with a maximum of 10,000 rows per query, according to Google's developer documentation. A second resource, ChangeStatus, looks back 90 days and lists which resources changed but not what they changed to. Google also says ChangeEvent "may not include every change seen in the web client's Change History." An agency auditing a client account through the API therefore sees a shorter and thinner trail than one using the interface.

Microsoft Advertising and Meta keep equivalent records. Prevention sits alongside the record. Google's multi-party approval requires a second administrator to approve adding or removing users and changing roles on accounts with more than three administrators, with 20 days to act before a request expires, according to its help centre. Read-only roles and API users are exempt.

Accountability under the GDPR

European data protection law turned record-keeping into a legal duty. Article 5(2) of the General Data Protection Regulation (GDPR), applied from May 25, 2018, states that the controller "shall be responsible for, and be able to demonstrate compliance with" the processing principles. That clause, known as the accountability principle, shifts the burden. A company has to show it complied, which in practice means keeping evidence: consent logs under Article 7(1), records of processing activities under Article 30, and an internal register of breaches under Article 33(5).

Article 30 is often mistaken for an event log. It is a register of processing operations, listing purposes, data categories, recipients and retention periods, not a record of individual actions. Organisations with fewer than 250 employees are partly exempt. The European Commission's Omnibus IV package, proposed on May 21, 2025, would require such records from small and mid-cap firms, defined as those under 750 employees, only where processing is likely to pose a high risk, according to law firms Eubelius and Arthur Cox. Arthur Cox noted in August 2025 that adoption might not come before the end of 2026. Whether that change has since been adopted could not be confirmed for this article.

Logging duties in the AI Act

The EU AI Act, Regulation (EU) 2024/1689, made logs a design requirement for high-risk AI. Article 12(1) states that such systems "shall technically allow for the automatic recording of events (logs) over the lifetime of the system." The logs must support identification of risks, post-market monitoring by the provider and operational monitoring by the deployer. For remote biometric identification, Article 12(3) sets minimum fields: the start and end time of each use, the reference database checked, the input data that produced a match, and the people who verified the results.

Article 19 requires providers to keep the logs under their control for at least six months, and Article 26(6) places a similar duty on deployers, unless other EU or national law, such as data protection law, sets a different period. After the omnibus amendments provisionally agreed in May 2026, these duties apply from December 2, 2027 for stand-alone Annex III systems, such as those used in hiring and credit scoring, and from August 2, 2028 for AI built into regulated products. Breaches of operator obligations can draw fines of up to 15 million euros or 3% of worldwide turnover, whichever is higher.

The Act prescribes no log format, and no final harmonised standard yet exists, according to a Help Net Security article from April 2026 by João Marques, founder of the logging vendor Asqav.

Agents and the audit trail

Software agents changed the question from who clicked to which program acted on whose authority. Ad tech vendors now sell audit trails as a condition for handing agents budgets. Yahoo DSP's agents, launched in January 2026, execute changes with human approval and let advertisers audit and override their actions. PubMatic's guardrails, announced on August 5, 2026, log what each agent did, when, under which parameters and whether a human was involved, with escalations approved by one-time passcode. Google Cloud said in October 2026 that its Gemini agents send every action to an audit trail attributed to the agent rather than to a person.

Platforms are writing logging into access rules. Google's developer policies of August 31, 2026, which ban programmatic proxies from the Ads API, require hosted services seeking a secondary interface review to "maintain immutable transaction logs that map every automated programmatic action directly to a verified end-advertiser account", available to Google on request. Inside the auction, IAB Tech Lab's Agentic Real Time Framework makes OpenTelemetry tracing mandatory for containers.

Limitations and disputes

Completeness fails in practice. In October 2024, a Google Ads specialist working for a subcontractor changed a client account without authorisation, unpinning headlines and adding assets, and the edits did not appear in change history. In November, agency leaders reported further cases; Ginny Marvin, Google's Ads Liaison, said Google "did not properly follow our processes for ensuring approvals are granted before making changes." Google's help page itself warns that changes made by Google representatives during consultations are sometimes not listed. In February 2026, PPC Land noted that the Google Ads support form authorises specialists to edit accounts to reproduce issues.

Verification is a second problem. Most agent audit trails are vendor-asserted: the party being audited also writes and stores the record. Privacy is a third. Logs that capture identifiers are personal data, which sits awkwardly with the GDPR's storage limitation principle, and UC Berkeley's agentic AI risk profile acknowledges that comprehensive agent logging may function as continuous surveillance. Finally, a trail records actions, not reasons. A model's stated reasoning may not reflect what actually drove its output, so a complete log of an agent can still leave its decisions unexplained.

Audit trail, audit log, change history and audit

An audit log is the stored data; the audit trail is the sequence it lets an investigator follow. The terms are often used interchangeably. Change history is a product feature, one platform's audit trail of configuration edits, usually without logins, reads or data exports. An audit is the examination itself, carried out by an internal team, an outside firm or a regulator; the trail is its evidence. Observability data, such as traces and metrics, is built to debug systems and may be sampled or discarded, which makes it a poor substitute for a record meant to prove accountability.

Recent developments

Regulators are now asking how agents should be logged. The UK Information Commissioner's Office opened a call for evidence on agentic AI on October 8, 2026, and, as PPC Land reported today, it puts 30 questions to agent builders, including on structured logging that records agent identity alongside actions, tamper-evident records and audit trails. The call closes on November 20, 2026.

Timeline

  • August 15, 1983: US Department of Defense issues the Trusted Computer System Evaluation Criteria (CSC-STD-001-83)
  • December 1985: TCSEC reissued as DoD 5200.28-STD, with audit trails required from class C2
  • March 20, 1997: FDA publishes 21 CFR Part 11 requiring time-stamped audit trails for electronic records
  • September 2006: NIST publishes SP 800-92, Guide to Computer Security Log Management
  • May 25, 2018: GDPR applies, including the Article 5(2) accountability principle and Article 30 records
  • August 1, 2024: EU AI Act enters into force with Article 12 logging duties for high-risk systems
  • October 15, 2024: Unauthorised Google Ads rep changes reported missing from change history
  • May 21, 2025: Commission proposes Omnibus IV, narrowing Article 30 records to high-risk processing for firms under 750 employees
  • January 2026: Google Ads adds a Go To button to change history
  • January 6, 2026: Yahoo DSP launches agents that execute changes with human approval
  • May 2026: Provisional AI omnibus deal moves Annex III high-risk duties to December 2, 2027
  • August 5, 2026: PubMatic announces agent guardrails with audit trails and passcode approvals
  • August 31, 2026: Google developer policies require immutable transaction logs for secondary interfaces
  • October 8, 2026: Google Cloud details agent-attributed audit trails; ICO opens agentic AI call for evidence
  • November 20, 2026: ICO call for evidence closes
  • December 2, 2027: AI Act logging and log-retention duties apply to Annex III high-risk systems

Summary

Who. System operators keep audit trails; auditors, security teams, advertisers and regulators read them. In advertising, Google, Microsoft, Meta, DSPs such as Yahoo and SSPs such as PubMatic maintain them for accounts and, increasingly, for AI agents. EU and UK data protection authorities and AI Act market surveillance bodies enforce the legal duties.

What. A time-ordered, attributable and tamper-resistant record of who did what in a system, when and with what result, kept so events can be reconstructed and responsibility assigned.

When. The computer security requirement dates to the Orange Book of 1983 and 1985; the FDA added it for electronic records in 1997; the GDPR's accountability principle applied from May 2018; AI Act logging duties for Annex III systems apply from December 2, 2027.

Where. In accounting ledgers, operating systems, cloud consoles, ad platform change histories, API change reports and the action logs of AI agents.

Why. Without a reliable record, no one can show who changed a campaign, prove compliance with data protection law or hold an automated agent's operator responsible for what it did.