OrganizationExpert6 min read
Part VI. Compliance: proof through logs and traces
Well-placed instrumentation supplies the evidence regulators ask for: the AI Act, NIS2 and GDPR matrix and the separation between content and metadata.
For: architects · finance, risk and compliancePrerequisites: Have read part I; notions of logging and data retention.
22. Regulatory evidence matrix
Section titled “22. Regulatory evidence matrix”An illustrative scenario. An auditor asks the team to show that the assistant records its decisions, that logs are kept for at least six months and that a human can take back control. The team could write a fifty-page file. It can also open its traces: every call is timestamped, every handoff to a human is counted, the retention policy is set.
That is the principle of this part. Good instrumentation produces, as a by-product, the evidence regulation requires, much as a lorry’s tachograph serves both the haulier and the roadside inspector. The matrix below links each obligation to its theme, to whoever bears it and to the signal that proves it.
Let me say it upfront: this is a working matrix, not legal advice. The precise article-by-article mapping must be validated with legal and the DPO, the data protection officer.
Who is concerned?
Section titled “Who is concerned?”The AI Act, the European regulation on artificial intelligence, assigns obligations along two criteria. The first is the role: the provider places the system on the market, the deployer uses it in its own business. A company that buys an assistant and makes it available to its staff is a deployer. The second criterion is the risk level. Most rows in the matrix apply only to high-risk systems. Article 50 is the exception: it covers any system that interacts with people or generates content, whatever its risk level.
A few acronyms from the matrix: the SOC is the security operations center. MTTD (mean time to detect) is the average time to detect an incident; MTTR (mean time to recovery) is the average time to restore service. Tool pinning means pinning tools by fingerprint, as described in Part III.
| Obligation theme | Text and article | Who, for which systems | Signal that produces proof |
|---|---|---|---|
| Logging and traceability: the system automatically records events throughout its lifetime | AI Act, Art. 12 | provider, high risk | Timestamped end-to-end traces with trace_id, tool invocation log, model and pipeline versions in the spans |
| Technical documentation: description of the system, the data, the evaluation and monitoring methods | AI Act, Art. 11 | provider, high risk | Evaluation history, versioned configuration |
| Transparency and information to deployers: instructions for use (expected performance, limits, means to collect and interpret logs) | AI Act, Art. 13 | provider, high risk | Model attribution, controlled content capture, prompt versioning, documented dashboards and thresholds |
| Human oversight | AI Act, Art. 14 | provider, high risk | Human handoff rate, validation points, traced guardrails |
| Accuracy, robustness and cybersecurity | AI Act, Art. 15 | provider, high risk | Quality SLOs, continuous evals, drift detection |
| Post-market monitoring: collect and analyze performance data throughout the life of the system | AI Act, Art. 72 | provider, high risk | Quality scores, drift, cost, alerts |
| Monitoring of operation: monitor the system according to the instructions, report risks | AI Act, Art. 26(5) | deployer, high risk | Alerts, escalation procedures toward the provider |
| Retention of automatically generated logs: at least six months, unless Union or national law provides otherwise | AI Act, Art. 19 (providers) and Art. 26(6) (deployers) | provider and deployer, high risk | Documented retention policy, timestamped trace and log retention set accordingly |
| Transparency toward people: inform people that they are dealing with an AI, mark generated content | AI Act, Art. 50 | provider and deployer, any system that interacts with people or generates content | The obligation belongs to the product, not to observability; traces can only record that the “you are talking to an AI” notice was shown and that content was marked |
| Incident handling | NIS2, Art. 21(2)(b) | essential and important entities | Security alerting, SOC escalation, measured MTTD and MTTR |
| Incident notification: early warning within 24 hours, notification within 72 hours, final report within one month | NIS2, Art. 23 | essential and important entities | Detection timestamp, timeline rebuilt from traces |
| Supply chain security | NIS2, Art. 21(2)(d) | essential and important entities | Tool pinning, mutation detection, authentication failures |
| Data minimization | GDPR, Art. 5(1)(c) | controller | Content opt-in, masking at the collector |
| Data protection by design and by default | GDPR, Art. 25 | controller | No content capture by default, separate retention, traceability of personal data |
Texts: AI Act, Regulation (EU) 2024/1689; NIS2, Directive (EU) 2022/2555; GDPR, Regulation (EU) 2016/679. Articles checked on 2 October 2026. The rows for Articles 11 and 72 and for Article 26(5) come from the provider and deployer table of the course, module 5, merged here on 4 October 2026. NIS2 obligations apply through the national transposition law.
A well-built OpenTelemetry stack provides the technical material for these obligations: logs, performance measurements, history. It does not do everything. You still have to qualify your system and your role, document it, then maintain the setup over time.
One row deserves a comment. Article 50 requires telling people they are talking to an AI and marking generated content. That is a product obligation rather than an observability one; traces can only attest that the information was displayed.
23. Content and metadata separation
Section titled “23. Content and metadata separation”A doctor’s practice keeps two binders. The appointment book (who came, when, for how long) can be consulted freely and is kept for a long time. The medical record is under lock and key, open to few people and subject to strict rules.
I recommend the same rule for an AI application. On one side sits content: what the user wrote and what the model answered. It is sensitive, kept briefly, access-restricted and redacted of personal data. On the other side sits trace metadata: which model, how many tokens, what duration, what error. It takes few distinct values (low cardinality), is kept for a long time and can be used freely. This is the OpenTelemetry external content pattern described in Part I, extended to the whole stack.
That leaves sovereignty, meaning under which law and in which country your data lives. Self-hosting the backends (for example VictoriaMetrics or Prometheus, VictoriaLogs or Loki, Grafana, Langfuse or Phoenix) lets you choose where it resides.
For Monday morning
Section titled “For Monday morning”Print the matrix and bring it to your next meeting with legal and the DPO. Start by settling your role (provider or deployer) and your system’s risk level: depending on the answers, a good share of the rows will not apply to you. For the rest, note next to each one the signal you already produce and the one that is missing.
Revised on 2 October 2026: added the articles of each text and the timeline from Regulation (EU) 2026/1744 amending the AI Act.
Revised on 4 October 2026: merged the AI Act provider and deployer table from the course (module 5) into the matrix, which gains a “who” column, Articles 11 and 72 and Article 26(5); the timeline is given only once, in the box.
Revised on 25 October 2026: page rewritten for a wider audience (opening scenario, roles and acronyms explained), with no change of substance.