Article 13 Instructions for Use That Deployers Can Act On

Article 13 requires high risk AI systems to be supplied with clear instructions for use. These are not optional guidance documents or best practices suggestions. The instructions are binding requirements for how the system must be used. A deployer who uses a high risk system contrary to its instructions for use is not complying with EU AI Act obligations.

Instructions must cover the intended purpose of the system. What decisions is the system designed to make? For a high risk recruitment system, the intended purpose might be to screen résumés and support hiring managers in identifying qualified candidates. The instructions must clearly state this purpose. Instructions must also address what the system is not designed for. The system is not designed to make final hiring decisions without human review. The system is not designed to be used for termination decisions or disciplinary assessments. Clarity about scope prevents misuse.

The instructions must explain the performance characteristics of the system. The deployer must know what accuracy to expect, broken down by relevant groups. If the system is more accurate for some demographic groups than others, this must be stated. If the system’s accuracy is 92 percent overall but only 82 percent for a particular demographic group, the instructions must disclose this. This enables deployers to set appropriate expectations and implement appropriate oversight.

Instructions must specify what data the system needs and in what format. A system requiring specific data fields that are not consistently available across an organisation would create deployment problems. Instructions must explain data requirements so deployers can confirm they have the necessary data infrastructure. Instructions must address data quality. If the system assumes certain data quality characteristics and these are violated (for example, missing values beyond a certain threshold), how should the deployer respond?

The critical requirement is that instructions must specify how the system should be used. This means describing the operational workflow. For a recruitment system, the instruction might be: “The system produces a candidate ranking. Hiring managers must review system recommendations and make hiring decisions based on interviews and qualifications assessment, not solely on system recommendations.” Instructions must specify escalation procedures. What should a deployer do if the system produces an unexpected result? Whom should they contact?

Human oversight requirements must be absolutely clear in the instructions. Which decisions require human review? What does human oversight mean operationally? For a high risk employment system, the instructions might state: “All hiring recommendations must be reviewed by a qualified hiring manager who confirms or overrides the system recommendation. The hiring manager’s decision, not the system’s recommendation, is the decision-maker.” Instructions must describe who is qualified to provide oversight. Can any employee review decisions or must it be someone with specific training or authority?

Instructions must address unusual circumstances. What should a deployer do if system performance appears degraded? What signs indicate a problem? If the system is used in a context where the data distribution is very different from training conditions, what risks emerge? Instructions should help deployers recognise when conditions have changed enough that they should pause system use pending investigation.

Instructions must address monitoring and performance review. How often should the deployer monitor system performance? What metrics should be tracked? If performance drops or patterns change, what should the deployer do? Instructions should specify communication procedures: if problems are detected, whom does the deployer contact for support?

Instructions must address record keeping and logs. The deployer must understand how logs are generated, what information they contain, how long they are retained and how to access them if needed. If regulators investigate, deployers may be asked to produce logs. Instructions should explain this requirement.

Consider practical implementation. A financial services organisation purchases a high risk system for investment decisions. Article 13 instructions for use specify that the system is designed to support portfolio managers in identifying investment opportunities but not to make autonomous investment decisions. The instructions specify that all investment decisions require human review and approval by qualified portfolio managers. Instructions address system accuracy, noting it is 88 percent accurate on average but 79 percent accurate in volatile market conditions. Instructions specify that system use should be paused if market volatility is beyond historical training ranges. Instructions explain that system performance should be reviewed weekly by the risk management team and that any performance degradation should trigger investigation before continued use.

Organisations deploying high risk systems must ensure their staff are actually trained on the instructions. Providing instructions is not sufficient. Deployers must understand them. The financial services organisation conducting training for portfolio managers about the investment system ensures that managers understand which decisions require human oversight, how to recognise when system performance may be degraded and what to do if concerns arise.