Skip to content

OrganizationPractitioner16 min read

2. Compliance and evidence capability

NIS2, DORA, CRA, AI Act and GDPR from the CIO's point of view: a verified timeline, what each text expects from observability, the common evidence foundation and how the CIO, CISO and DPO work together.

For: architects · finance, risk and compliance · executives and CIOsPrerequisites: None.

Reading mode

European texts on cybersecurity and AI have one thing in common. They expect an organization to prove, on request, what happened, when and how it responded. Think of an aircraft’s black box: after an incident, nobody asks the crew to remember, they read the recorder. Observability can play that role for your information system, provided it was designed for it.

The timeline, verified as of 2 October 2026

Section titled “The timeline, verified as of 2 October 2026”

This table is a reference. The texts keep changing: check the revision date before quoting it in committee.

TextReferenceWhat applies and since when
GDPRRegulation (EU) 2016/679applicable since 25 May 2018
DORARegulation (EU) 2022/2554applicable to the financial sector since 17 January 2025
NIS2Directive (EU) 2022/2555transposition due by 17 October 2024. In France, the so-called resilience bill has not been enacted. The Commission referred France to the Court of Justice of the European Union on 8 July 2026 for failure to transpose. In Belgium, the law of 26 April 2024 has applied since 18 October 2024
AI ActRegulation (EU) 2024/1689in force since 1 August 2024; prohibited practices since 2 February 2025; general-purpose AI models since 2 August 2025
Amended AI ActRegulation (EU) 2026/1744 (Digital Omnibus on AI)Annex III high-risk systems postponed to 2 December 2027, Annex I systems to 2 August 2028. Article 50 (transparency) applicable since 2 August 2026. Marking of synthetic content is postponed to 2 December 2026 only for systems placed on the market before 2 August 2026
CRARegulation (EU) 2024/2847reporting of actively exploited vulnerabilities since 11 September 2026; main obligations on 11 December 2027

In France, ANSSI estimates that about 15,000 entities will fall under NIS2, compared with a few hundred under the first directive. The lack of a national law is no excuse for not preparing: customers, insurers and clients in the financial sector already require these measures by contract, as the letter Laurent received shows.

For a long time, keeping logs (the records where systems write down what they do) was a matter of good practice. For a growing part of the European economy, it has become an obligation backed by penalties. For the CIO, that changes three things.

First, keeping the record matters more than having it at hand instantly. Logs stored in a cheap archive that is slow to search are better than no logs at all. Second, the burden of proof shifts: after an incident, it is up to the organization to show what it detected, when and what it did. Third, the dialogue with the authorities can become continuous, between notifications, requests for information and inspections.

The opposite trap exists too. Faced with several texts, it is tempting to keep everything, for a very long time, on the most expensive platform. The bill explodes, GDPR exposure grows and compliance does not move forward. In my view, the right approach is to map what is required, for which scope, for how long and with what level of integrity, then keep only that.

The directive distinguishes two categories of entities, essential and important, according to sector (Annexes I and II, eighteen sectors in total) and size. As a general rule, it targets medium-sized and large companies in the listed sectors, with exceptions that cover certain entities regardless of their size. The first question to ask your lawyer is therefore simple: are we essential, important or out of scope? The answer must be written and dated.

The ten measures of Article 21 and the evidence observability can provide

Section titled “The ten measures of Article 21 and the evidence observability can provide”

Article 21 lists ten measures, from a to j. The table below sets against each one the evidence that logs, metrics (numerical indicators tracked over time) and traces (the detailed journey of a request) can provide. It is a reference table to bring to your meetings with the CISO.

PointMeasure (Article 21(2))Evidence logs, metrics and traces can provide
apolicies on risk analysis and information system securityasset inventory kept up to date by automatic discovery, dated review
bincident handlingalerts, timestamped timelines, tickets linked to traces and logs
cbusiness continuity, backups, disaster recovery, crisis managementbackup success, timed restore tests, logged exercises
dsupply chain securityavailability and errors of supplier services measured, supplier access logged
esecurity in acquisition, development and maintenance, including vulnerability handlingdeployed versions tracked, software bill of materials (SBOM), patch times measured
fassessment of the effectiveness of measuressecurity indicators tracked over time, quarterly reviews archived
gcyber hygiene and trainingpatch and configuration status, training rate tracked
hcryptography and encryptioncertificate inventory and expiry, key rotation logged
ihuman resources security, access control, asset managementaccess logs, review of privileged accounts, departures processed
jmulti-factor authentication, secured communications, emergency communicationsshare of access with MFA, authentication logs, emergency channel tests

The NIS2 Article 21 self-assessment uses this grid to position your organization, measure by measure.

For a significant incident, the directive asks for an early warning within 24 hours, a notification within 72 hours and a final report no later than one month after the notification. These deadlines can only be met if observability provides facts from the moment of detection: a reliable time for each event, the affected scope, the probable vector, the volume affected. Reconstructing a timeline from scattered logs, several tens of hours after the facts, puts the organization in a weak position. That is exactly what the April 2026 simulation showed at Helinord.

Management bodies approve the risk management measures and oversee their implementation. They can be held liable for failures and must follow training. For the CIO, keeping a written record of decisions and trade-offs on detection and logging becomes a form of protection.

For a breach of Articles 21 or 23, the maximum is at least €10M or 2% of total worldwide annual turnover for an essential entity. It is at least €7M or 1.4% for an important entity. The higher of the two amounts applies. Take a mid-sized company with €120M in revenue classified as important: 1.4% would only come to €1.7M. So the €7M amount applies, unless national law raises it, since it is the minimum the directive sets for this ceiling.

DORA applies to banks, insurers, asset management companies, payment service providers, market infrastructures and, indirectly, their IT service providers. Three workstreams affect observability.

The first is a verifiable map of critical functions and the assets that run them. Picture the plans of a building where every pipe carries a label saying which flat it feeds. In observability, those labels are attributes carried by metrics, logs and traces (service, business function, criticality): they link each technical component to the function it serves. Without them, the map remains a document disconnected from reality.

The second is accelerated notification of major ICT-related incidents. The implementing texts set an initial notification within four hours of the incident being classified as major and no later than 24 hours after its detection. An intermediate report and a final report follow. You therefore have to measure the impact, not just the incident: how many customers, for how long, over what scope.

The third covers resilience testing and the oversight of critical providers: a register of providers, continuous measurement of their availability and their errors. A service level written into a contract that nobody measures is not a service level.

CRA, for those who place digital products on the market

Section titled “CRA, for those who place digital products on the market”

The Cyber Resilience Act governs the cybersecurity of products that contain digital elements. Since 11 September 2026, manufacturers must report actively exploited vulnerabilities and severe incidents: early warning within 24 hours, notification within 72 hours, then a final report. This means spotting attacks in production and linking them to a known flaw. It often relies on data reported back from the installed base at customer sites, with everything that implies in terms of GDPR and contracts.

A CIO who buys without manufacturing has no obligation of their own under the CRA. The manufacturer draws up the software bill of materials, or SBOM, a kind of ingredients list for software like the one printed on a package. It keeps that list available to market surveillance authorities. The CIO still has an interest in requesting SBOMs from suppliers and keeping them. It is also worth tracking the versions deployed across the estate and the announced end-of-support dates.

For observability, three requirements matter. High-risk systems must automatically log how they operate (Article 12). The organization deploying such a system must keep those logs for at least six months (Article 26). Finally, the system must be monitored after it is placed on the market (Article 72). The postponement decided in 2026 gives time, not an exemption. Penalties can reach €35M or 7% of worldwide turnover for prohibited practices. The details, signal by signal, are in the evidence matrix of the GenAI method.

A log that contains an IP address, an identifier, an email address or personal request parameters is processing of personal data, from the moment of collection. People often forget this because nobody “decided” to collect that data: it arrives with everything else. With the DPO, you then need to document a legal basis for security logging (usually legitimate interest), justify a retention period for each log category, keep a data processing agreement up to date with each observability vendor and put in writing how transfers outside the European Union are governed.

The safest practice is to pseudonymize as early as possible, meaning you replace the identity with a code. Salted hashing turns an identifier into an unreadable fingerprint that is always the same for the same person: you can still follow a journey without knowing whose it is. Free-text fields, names and addresses are masked before sending. An OpenTelemetry Collector, the relay that receives and transforms data before it is stored, can perform these transformations on the way through.

NIS2, DORA, the CRA and the GDPR ultimately ask for the same six capabilities. You must be able to find critical events, give them a reliable time synchronized across systems, guarantee that audit records have not been altered, keep them for a period suited to each category, produce them quickly on request and get sources that tell the same story. These needs overlap enough to build a single foundation instead of four setups.

flowchart LR
  S["Sources<br/>authentication, admin,<br/>configuration, incidents"] --> C["1. Standardized collection<br/>timestamping, taxonomy"]
  C --> W["2. Immutable storage<br/>write once"]
  W --> R["3. Auditable search<br/>who viewed what"]
  R --> P["Evidence<br/>auditor, authority"]

The first stage is standardized collection. All critical sources arrive in a common infrastructure, each event is timestamped by a reliable time source and tagged with a shared vocabulary: team, application, environment, criticality, event type.

The second is immutable storage, on media where you write once without being able to change anything afterwards, like a bound ledger you cannot tear pages out of. Object storage with locking makes this possible. Immutability is necessary for evidence but not sufficient: see a signed log is not an auditable log.

The third is auditable search: the lookups themselves are traced, so you know who looked at what.

They depend on the type of log, the sector and the recommendations of the authorities (CNIL, ANSSI, ACPR or AMF depending on the case). Only one rule is stable: classify logs by category and justify the period for each one. Authentication, incident and audit logs of the observability platform itself call for long periods. Technical application logs can make do with a short period calibrated on operational needs. Write down your grid, have it validated by the DPO and the CISO, then review it every year.

None of these three roles carries compliance alone. The CIO provides the infrastructure. The CISO defines the policies and answers for security. The DPO protects individuals and limits collection. To get them working together, I recommend a joint committee at least once a quarter plus a one-page RACI, the table that says for each activity who does the work, who decides, who is consulted and who is informed (see the RACI matrix). It must answer precise questions: who decides on retention periods, who signs data processing agreements, who notifies the authority, who keeps the SBOM up to date, who runs the tests.

Add an annual calendar: retention review, review of critical suppliers, restore test, incident simulation, internal audit. Five indicators are enough to follow the whole thing:

  • the share of critical functions covered by complete logging;
  • the share of critical logs in immutable storage;
  • the time demonstrated in simulation between detection and early warning;
  • the share of log categories meeting their retention period;
  • the number of successful archive restores during the year.

The first is maximum retention for everything: the bill multiplies and GDPR exposure gets worse. The second is a single expensive platform for everything. Day-to-day operations and compliance archiving do not have the same needs, like an office and an archive cellar: two tiers cost much less.

The third is partial encryption. You need encryption in transit, at rest and in backups, with managed and audited keys; every forgotten link is a risk. The fourth concerns training AI models on your logs: require by contract that your data is not used to train a vendor’s models without explicit consent.

The last is paper compliance. A policy that is written but not applied mostly documents what the organization knew it should do. Inspections look for evidence of application.

Allow 45 minutes, in pairs, with the CISO if possible.

  1. Have your NIS2 status qualified in writing (and your DORA status if you serve the financial sector).
  2. Fill in the NIS2 Article 21 self-assessment, distinguishing “measure in place” from “measure demonstrable”.
  3. For the two weakest measures, name the missing evidence and the log, metric or trace source that would produce it.
  4. Sort the actions into three months, six months and twelve months, each with an owner.

Three exercises can follow. Capability mapping takes half a day: you place each of the six capabilities of the foundation on the scale absent, in progress, partial, in place. Incident simulation takes a day: you time the path from detection to early warning, judge the timeline you get and how finely you can measure the impact, then do it again six months later. The foundation audit takes a week. For one critical function, you produce the administrator access over several months, the timeline of a past incident and the list of suppliers involved. If this takes more than a day, the foundation is undersized.

To finish, write down your next action: which one, by what date, with whom?

Revised on 2 October 2026: AI Act Article 50 timeline clarified (Regulation (EU) 2026/1744), Belgian NIS2 transposition law added, scope of the CRA for a buyer corrected. The NIS2 penalty ceiling is now presented as a minimum set by the directive; immutability, as a necessary but not sufficient condition for evidence.