Table of Contents
If your organization uses a high-risk AI system in the EU, start logging now. The EU AI Act says deployers must keep automatically generated logs, when those logs are under their control, for a suitable period and at least six months, unless another law requires longer [source: EU AI Act].
TLDR: If you deploy high-risk AI, you need clear logs, clean audit trails, and a six-month record retention plan. Think of an HR tool that ranks 1,000 job applicants. You should be able to show who used it, when, what version ran, and what decision followed. Six months is about 183 days, so do not set logs to vanish after 30 days and hope nobody asks.
What “deployer” means
A deployer is the person or company using an AI system under its own authority. Not the vendor. Not always the builder. The user.
If your team buys an AI tool for hiring, credit checks, school admissions, medical triage, or worker management, you may be the deployer. That means you get chores. Sorry. Compliance is not glamorous.
The funny part? Many teams already have logs. They just cannot find them. Or they sit in five tools. Or they are named something like “final final audit export v7.” It drives me crazy when a system can score a person in two seconds but takes three days to explain how it did it.
The big logging rule
For many high-risk AI systems, deployers must keep the logs that the system creates automatically. The key words are under your control. If you can access them, export them, or store them, treat them like records.
Your logging setup should answer simple questions:
- Who used the AI system?
- When did they use it?
- What input was used?
- What output did the system give?
- Which version of the model was active?
- What human decision happened after the AI output?
- Was there an error, override, or alert?
That list is not just paperwork. It is your safety net. If someone asks why an AI system flagged a loan application, you need more than “the computer said no.”
What an audit trail should show
A log is a snapshot. An audit trail is the story.
It should show the chain of events from input to final action. For example:
- A user uploads a job applicant’s profile.
- The AI tool gives a match score of 72%.
- A recruiter reviews the score.
- The recruiter rejects the suggestion.
- The reason is recorded: “Missing required license.”
- The final decision is saved.
This matters because the EU AI Act cares about human oversight. A human must not be decorative. They need to understand enough to check the AI, question it, and stop it when needed.
Six-month retention, made simple
The six-month rule means you should keep relevant logs for at least six months. Not 30 days. Not “until storage gets expensive.” Not “until someone clears the folder.”
Set your retention period to 183 days minimum. Many teams choose 12 months to be safe. But do not keep everything forever for no reason. Privacy laws still matter. More data can mean more risk.
A simple retention policy might say:
- High-risk AI logs: 12 months.
- Security access logs: 12 months.
- Raw personal data: shortest useful period.
- Incident records: longer, if legal or insurance teams require it.
Honestly, it feels like every tool has a different default. One keeps logs for 7 days. Another keeps them forever. Another charges extra for export. That is how teams get caught out.
Tools that make this easier
You do not need a giant system on day one. You need a system that works.
Useful tool types include:
- SIEM tools: Good for security logs and alerts.
- GRC platforms: Good for policies, controls, and evidence.
- Data warehouses: Good for storing structured AI usage records.
- Ticketing tools: Good for incidents, reviews, and approvals.
- Model monitoring tools: Good for drift, errors, and performance issues.
- Access management tools: Good for proving who did what.
The magic is not the logo on the software. The magic is consistency. Every AI action should leave a trace. Every trace should have an owner. Every owner should know what to do if something smells wrong.
Minimum fields to record
Start with a small template. Make it boring. Boring is good here.
- Log ID
- Date and time
- User or team
- AI system name
- Model version
- Input type
- Output summary
- Confidence score, if used
- Human reviewer
- Final decision
- Override reason
- Error or incident flag
Do not store sensitive data unless you need it. If a reference number is enough, use that. Keep personal data locked down. Use role-based access. Add encryption. Track exports too.
A small use case
Picture a bank using AI to help review small business loans. The tool reviews 5,000 applications per month. It flags 8% for extra checks. A loan officer makes the final call.
The bank should keep logs showing the application ID, risk score, model version, officer review, and final decision. If a customer complains four months later, the bank can rebuild the record. No panic. No mystery folder. No awkward silence.
Quick checklist for deployers
- Find every high-risk AI system in use.
- Check what logs each system creates.
- Confirm who controls those logs.
- Set retention to at least six months.
- Record human review and overrides.
- Test log exports before you need them.
- Limit access to sensitive records.
- Run a monthly spot check.
Bottom line: logs are not just tech clutter. They are proof. They show that your AI system was used with care, watched by humans, and kept on record long enough for questions later. That may not sound thrilling. But when an audit arrives, boring proof feels wonderful.