In this article8 sections
  1. The AI Act Is an Engineering Problem, Not Just a Legal One
  2. Step One: Build Your AI System Inventory
  3. Article 12: What Logging Infrastructure You Must Have
  4. Article 9: Risk Management as a Technical Discipline
  5. Vendor Evaluation: What to Demand Before You Sign
  6. Deadlines Your Engineering Roadmap Must Reflect
  7. Human Oversight Is a Feature, Not a Process
  8. References

Every eu ai act cio cto conversation eventually lands in the same place: compliance counsel hands the technical team a PDF, and someone has to turn legal text into infrastructure. That someone is you. The good news is the obligations are specific enough to plan around. The hard part is that most SME tech stacks were not designed with AI governance in mind, so the work is real.

This playbook covers what Article 12, Article 26, and Article 9 actually require you to build, how to evaluate vendor AI systems before you deploy them, and what your monitoring architecture needs to look like. No vague frameworks. Concrete engineering implications.

Step One: Build Your AI System Inventory

You cannot govern what you cannot see. Before any tooling decision, you need a structured inventory of every AI system your organisation deploys. Under Article 26 of the EU AI Act, deployers of high-risk systems carry specific obligations that attach to individual systems, not to AI in general. Article 4 (AI literacy) already applies to every AI tool your staff use. Without an inventory, you have no way to know which obligations apply where.

Your inventory should capture at minimum:

  • System name and vendor

  • Use case and business process it touches (e.g., CV screening in recruitment, credit scoring in sales)

  • Risk classification (prohibited, high-risk, limited risk, minimal risk)

  • Data inputs and outputs

  • Who in the organisation is the designated human overseer

  • Where logs are stored and for how long

Risk classification is where most SME tech leaders get stuck. The high-risk categories are defined in Annex III of the AI Act. They include AI systems used in recruitment and HR management, access to education, essential private services such as credit scoring, and critical infrastructure. If your company uses an AI tool for any of these purposes, it is likely a high-risk system (Article 6(3) has narrow exceptions), and the high-risk regime applies from 2 December 2027.

Keep the inventory in a living document. It is not a one-time audit. New SaaS tools get added quarterly. An employee starts using an AI-powered performance management platform without central procurement. A vendor quietly adds an AI feature to a product you already use. Your inventory process needs to catch all of this.

Article 12: What Logging Infrastructure You Must Have

Article 12 is the most direct technical requirement in the Act. It is a design requirement on the provider: high-risk AI systems must technically allow events to be logged automatically over their lifetime, to support traceability and monitoring. For remote biometric identification systems, Article 12 lists specific items to be logged, such as the period of each use, the reference database, the input data that led to a match, and the people who verified the results.

As a deployer, your duty sits in Article 26(6): keep the logs the high-risk system automatically generates, to the extent they are under your control, for a period appropriate to the intended purpose and of at least six months, unless other law says otherwise. Relying on a vendor's own audit trail alone is risky, because you need to be able to retrieve those logs when you need them.

Practically, this means:

  1. Request log export access from every high-risk AI vendor. If a vendor cannot give you tamper-evident log exports in a structured format, that is a procurement red flag.

  2. Store logs in your own environment. Whether that is your SIEM, a dedicated logging service, or object storage, the logs must live somewhere you own.

  3. Set retention to match the Act. Article 26(6) sets a minimum of six months for deployers of high-risk systems, unless other EU or national law (for example sector rules or the GDPR) requires something else. Choose a period that fits the purpose and align it with your GDPR data minimisation obligations.

  4. Tag log entries with the human decision-maker. Where a human reviews AI output and makes a final call, make that reviewer's identity traceable. It is the simplest way to show your human oversight works. This often means building a thin wrapper or confirmation screen in your internal tooling that writes a structured log entry at decision time.

Article 9: Risk Management as a Technical Discipline

Article 9 requires providers of high-risk AI systems to establish a risk management system. Deployers do not have to build one. Your side is in Article 26: use the system according to the provider's instructions for use, monitor its operation on that basis, and inform the provider (and, where required, the authorities) if you have reason to think it presents a risk.

In engineering terms, this means you need a monitoring pipeline that watches for:

  • Accuracy drift: Is the model's performance degrading over time compared to a baseline?

  • Distributional shift: Are the inputs your users are sending different from what the model was trained on?

  • Adverse outcome rates: Are certain demographic groups receiving systematically different outputs? This is particularly critical in HR and financial services use cases.

  • Human override rates: If users are overriding or ignoring AI recommendations at a high rate, that is a signal the system is not performing as intended.

For SMEs, a full MLOps platform is often out of scope. You do not need Kubernetes and a bespoke feature store. You do need a scheduled job that queries your logs weekly, computes simple statistics, and raises an alert when thresholds are breached. Start simple. The obligation is to monitor, not to build a perfect ML observability stack on day one.

Vendor Evaluation: What to Demand Before You Sign

From 2 December 2027, Article 26 makes deployers of Annex III high-risk systems responsible for using them in accordance with the instructions of use provided by the provider. That means the provider must give you instructions of use. If they do not, or if those instructions are vague, you carry risk you cannot manage.

Before signing any contract for an AI system that touches a high-risk use case, require the vendor to provide:

  • The EU declaration of conformity and CE marking (mandatory for high-risk AI providers under the Act once the high-risk rules apply)

  • The instructions for use required by Article 13, covering intended purpose, accuracy and performance metrics, known limitations, human oversight measures and logging. The full technical documentation is for the authorities; a vendor is not obliged to hand it to you, so ask for a summary of training data sources and performance instead

  • Logging and audit trail capabilities in writing, not just in a sales deck

  • Incident notification procedures: if the system fails or behaves unexpectedly, how and how quickly will they tell you?

  • Post-market monitoring reports: what data are they collecting about the system's real-world performance?

If a vendor cannot answer these questions before contract signature, that tells you exactly what their AI governance maturity looks like. Build vendor AI governance into your standard procurement checklist alongside your existing security and GDPR due diligence.

Deadlines Your Engineering Roadmap Must Reflect

The Act entered into force on 1 August 2024. The key dates for deployers of high-risk systems are:

  • 2 February 2025: Prohibited AI practices and Article 4 (AI literacy) apply

  • 2 August 2026: The Act applies in general, including Article 50 transparency duties; national market surveillance authorities start supervising

  • 2 December 2026: Two new prohibitions apply (non-consensual intimate deepfakes and AI-generated child sexual abuse material)

  • 2 December 2027: Obligations for high-risk AI systems under Annex III apply

  • 2 August 2028: Obligations extend to high-risk AI systems embedded in regulated products (Annex I)

The Digital Omnibus on AI (Regulation (EU) 2026/1744) moved the high-risk dates to these fixed dates. That gives SME deployers a realistic runway: inventory this quarter, logging infrastructure next, monitoring tooling validated well before December 2027.

Under Article 99, infringements of Article 26 deployer obligations can be fined up to €15 million or 3% of global annual turnover. For SMEs and start-ups, the lower of those two figures is the maximum, and authorities must weigh the circumstances of the case. Building the infrastructure early is cheaper than retrofitting it.

Human Oversight Is a Feature, Not a Process

One final engineering principle. Article 26 requires deployers of high-risk systems to assign human oversight to people with the necessary competence, training and authority. Too often, compliance teams interpret this as a policy document. It is an engineering requirement. Your interfaces must make it possible for a human reviewer to understand, question, and override AI outputs.

That means:

  • AI outputs must display confidence scores or uncertainty indicators where technically feasible

  • Reviewers must be able to access the input data that generated a recommendation

  • Override actions must trigger a structured log entry

  • Systems must not present AI outputs as final decisions without explicit human confirmation

Build these patterns into your design system now. They will serve you across every AI tool your organisation deploys.

If you want a clear view of where your organisation stands today, the free 2-minute compliance check at comply.khairos.ai maps your current AI use against the Act's obligations and tells you which gaps to prioritise first.

References

  1. Article 26 of the EU AI Act

  2. Annex III of the AI Act

  3. Article 12

  4. European Commission, AI Act overview

  5. Regulation (EU) 2026/1744 (Digital Omnibus on AI)

  6. Article 9

  7. Article 99

  8. comply.khairos.ai

Not sure where your company stands?

Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.

Am I compliant? Check in 2 minHow Khairos Comply works