EU AI Act Article 86: Right to Explanation, Implemented in Practice
EU AI Act Article 86 gives every person about whom a deployer has made a decision on the basis of a high-risk system the right to a clear and meaningful explanation of the role the AI played and of the main elements of the decision. We generate that explanation from the evidence log, not from what the model says about itself.
Alongside the right to complain in Art. 85, Article 86 is the only article of the AI Act that gives an affected person a direct claim against the deployer. It is addressed to the deployer, not the provider. For banks, insurers and HR departments this means: if you prepare a decision with an AI system, you have to be able to explain it afterwards, even when a human made the final click. That exact case, model prepares, human decides, is the gap Article 86 of the AI Regulation closes, because Art. 22 GDPR does not reach it.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. For Article 86, Art. 113 names 2 August 2026 as the date of application; in practice the right only bites once a decision rests on an Annex III system that is operated as a high-risk system. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.
What Article 86 requires
Art. 86(1), scope: An affected person is subject to a decision the deployer takes on the basis of the output of a high-risk system listed in Annex III, except critical infrastructure under Annex III point 2. The decision produces legal effects or similarly significantly affects the person, and the person considers it to have an adverse impact on their health, safety or fundamental rights. That assessment is the person’s own.
Art. 86(1), content: The person has the right to obtain from the deployer clear and meaningful explanations, first of the role of the AI system in the decision-making procedure, second of the main elements of the decision taken. Recital 171 states the purpose: a basis on which the person can exercise their rights.
Art. 86(2): The right does not apply to AI systems for which exceptions or restrictions follow from Union or national law, provided that law complies with Union law, for instance disclosure restrictions in anti-money-laundering.
Art. 86(3): The article applies only to the extent the right is not otherwise provided for under Union law. For solely automated decisions, Art. 22 and Art. 15(1)(h) GDPR take precedence, for consumer credit the Consumer Credit Directive (EU) 2023/2225. Article 86 fills the rest, above all the decision with a human in the procedure.
What is missing: No deadline, no form, no level of detail. The deployer has to decide those itself, before the first request arrives.
The technical example
Two cases we see in projects exactly like this. First: a bank rejects a consumer loan. The application passes through the policy gateway, which pseudonymises, checks scope and data zone, and calls a pre-check model, a high-risk system under Annex III point 5(b), creditworthiness assessment of natural persons. The model returns a class, a score and the features with the largest contribution. In parallel, the credit policy runs as a rules engine: debt-to-income ratio, negative markers, minimum income. A loan officer sees the application, the model recommendation and the rules that fired, and rejects with a reason code. The gateway writes the evidence.
Second: a company rejects a job application. A language-model-based assistant, high-risk under Annex III point 4(a), receives the pseudonymised CV, retrieves the job profile in version 3 and the business unit’s criteria list, and produces a structured check: criterion met or not met, with the passage in the CV it relied on, plus a summary. A recruiter reads both, verifies two of the passages and declines.
Where the obligation bites: in both cases a human decided. Art. 22 GDPR therefore does not apply; Article 86 does. The language model delivered a rationale along with its output, fluent and convincing. But that rationale is a second model output, not a look inside, and nobody can verify whether it describes what drove the assessment. A plausible explanation is not the same thing as a faithful one.
How we implement it
At SIFAMO we treat the Article 86 explanation as a query on the evidence log, not as a text somebody writes or a model generates. What is not in the evidence is not in the explanation. The components are described in the governance layer.
The decision record. Per decision, the gateway writes: the input hash and the list of fields used, model ID and version from the registry, policy version and the rules that fired, the retrieved documents with version and passages, the output, the signals (score, confidence, disagreement between model and rules engine, feature contributions for the scoring model) and the reviewer action: role, accepted or overridden, reason code, timestamp. The registry record additionally holds an explanation view: which of these fields may go to affected persons, and in which wording. The business owner signs it with the record.
The template. The explanation has five fixed parts, which the explanation service fills from the evidence:
Decision: What was decided, when, and which role in the organisation decided.
Role of the AI system: One sentence from the purpose field of the registry record (“prepares the assessment, does not decide”), whether the recommendation was accepted or overridden, and the registry ID as a reference.
Main elements: Which categories of the person’s data were used; which rules from which policy version fired, in words; for the scoring model, the features with the largest contribution and their direction; for the language model, the retrieved documents with version and the criteria marked as not met, with the passages the human verified.
Human review: What the person checked, confirmed or changed, taken from the reviewer action.
Routes: A fresh review by a human, a complaint under Art. 85, the rights under the GDPR.
What must not go in. Other people’s data: no comparison group and no other applicants. No model weights, prompts or system instructions. Exact thresholds where they are trade secrets: then we name category and direction (“debt-to-income ratio above the internal limit”), not the number, and document that in the registry record. The CJEU has clarified for Art. 15 GDPR that trade secrets do not block an explanation wholesale but have to be weighed. And: no free text from the model. The generated rationale stays in the evidence for the reviewer and never goes to the person.
The interpretability gap. I am deliberately sober here. When a language model prepares the decision, the “main elements” we can honestly name are the retrieved documents, the rules applied and the passages the human confirmed. Not why the model rated a criterion the way it did. We write that limit into the explanation itself. For the scoring model, feature contributions are approximations; the method is in the registry record, and the quality gate checks before go-live whether the contributions are stable across similar cases. Claiming more than the evidence supports is, to me, a compliance risk, not a service.
GDPR and response times. For us, paragraph 3 means one process for both worlds. If the decision is solely automated, the request runs as an access request under Art. 15(1)(h) GDPR, with the same evidence and the same template, extended by logic, significance and consequences. If a human decided, it runs under Article 86. Because the AI Act names no deadline, we adopt the month from Art. 12(3) GDPR as the upper bound. The target is a few working days: the explanation is a query with sign-off by the business unit, not research.
The auditor sees the explanation view with version and signature, per request the timestamp, the response and the evidence ID it was generated from. More on how this fits into our AI governance.
Typical mistakes
Asking the model why. The generated rationale becomes the explanation. It sounds good, but it cannot be verified and often is not what drove the output.
“A human decided, so we do not have to explain anything.” That is true for Art. 22 GDPR and is exactly the case Article 86 was written for.
Reconstruction from memory. Months later the loan officer explains what he probably saw, because the log lacks the policy version and the retrieved documents.
Other people’s data in the explanation. The comparison group or the “better candidate” shows up, and the explanation turns into a data breach.
EU AI Act Article 86 checklist
The evidence log holds, per decision, the field list, model version, policy version with the rules that fired, retrieved documents with version, output, signals and reviewer action.
An explanation view per registry record defines which fields go to affected persons and in which wording; the business owner has signed it.
The template names the role of the system, the main elements, the human review and the routes to a fresh review and to a complaint under Art. 85.
Exclusion list checked: no third-party data, no model internals, no generated rationale text, handling of trade secrets documented.
Process agreed with data protection: one intake for Art. 86 and Art. 15 GDPR, deadline at most one month, every request logged as an event in the evidence.
Rehearsed once: an explanation generated from a real evidence record and read by the business unit before the first request arrives.
Open questions
How detailed “main elements” must be is neither defined nor standardised. We orient ourselves on the purpose in Recital 171: is the explanation enough for a reasoned appeal?
Where the human merely waves the recommendation through, the decision may be solely automated after all and Art. 22 GDPR applies. The blind re-checks in our oversight tier are then the only proof that the human really decided.
For language models there is no method that faithfully says which part of the input drove the output. Until there is, retrieved documents and applied rules are the honest ceiling of what we can explain.
Frequently asked questions
Does this apply to us as a deployer if a human always decides?
Yes, especially then. Art. 22 GDPR covers only solely automated decisions. Article 86 only requires that the deployer’s decision rests on the output of an Annex III system, and it does when the human had the model recommendation in front of them.
Do we have to disclose our score or the thresholds?
What is required is the role of the system and the main elements of the decision, not the model parameters. Which features and rules tipped the decision belongs in; exact thresholds and weights can be trade secrets, in which case we name category and direction.
How fast do we have to respond?
Article 86 names no deadline. We adopt the month from Art. 12(3) GDPR as the upper bound and measure ourselves in working days, because the explanation is generated from the evidence and does not have to be researched. The other articles of the series: all articles at a glance.
EU AI Act Article 86: Right to Explanation, Implemented in Practice
EU AI Act Article 86 gives every person about whom a deployer has made a decision on the basis of a high-risk system the right to a clear and meaningful explanation of the role the AI played and of the main elements of the decision. We generate that explanation from the evidence log, not from what the model says about itself.
Alongside the right to complain in Art. 85, Article 86 is the only article of the AI Act that gives an affected person a direct claim against the deployer. It is addressed to the deployer, not the provider. For banks, insurers and HR departments this means: if you prepare a decision with an AI system, you have to be able to explain it afterwards, even when a human made the final click. That exact case, model prepares, human decides, is the gap Article 86 of the AI Regulation closes, because Art. 22 GDPR does not reach it.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. For Article 86, Art. 113 names 2 August 2026 as the date of application; in practice the right only bites once a decision rests on an Annex III system that is operated as a high-risk system. The Omnibus moved deadlines; it removed no obligation. The overview is in EU AI Act in practice.
What Article 86 requires
The technical example
Two cases we see in projects exactly like this. First: a bank rejects a consumer loan. The application passes through the policy gateway, which pseudonymises, checks scope and data zone, and calls a pre-check model, a high-risk system under Annex III point 5(b), creditworthiness assessment of natural persons. The model returns a class, a score and the features with the largest contribution. In parallel, the credit policy runs as a rules engine: debt-to-income ratio, negative markers, minimum income. A loan officer sees the application, the model recommendation and the rules that fired, and rejects with a reason code. The gateway writes the evidence.
Second: a company rejects a job application. A language-model-based assistant, high-risk under Annex III point 4(a), receives the pseudonymised CV, retrieves the job profile in version 3 and the business unit’s criteria list, and produces a structured check: criterion met or not met, with the passage in the CV it relied on, plus a summary. A recruiter reads both, verifies two of the passages and declines.
Where the obligation bites: in both cases a human decided. Art. 22 GDPR therefore does not apply; Article 86 does. The language model delivered a rationale along with its output, fluent and convincing. But that rationale is a second model output, not a look inside, and nobody can verify whether it describes what drove the assessment. A plausible explanation is not the same thing as a faithful one.
How we implement it
At SIFAMO we treat the Article 86 explanation as a query on the evidence log, not as a text somebody writes or a model generates. What is not in the evidence is not in the explanation. The components are described in the governance layer.
The decision record. Per decision, the gateway writes: the input hash and the list of fields used, model ID and version from the registry, policy version and the rules that fired, the retrieved documents with version and passages, the output, the signals (score, confidence, disagreement between model and rules engine, feature contributions for the scoring model) and the reviewer action: role, accepted or overridden, reason code, timestamp. The registry record additionally holds an explanation view: which of these fields may go to affected persons, and in which wording. The business owner signs it with the record.
The template. The explanation has five fixed parts, which the explanation service fills from the evidence:
What must not go in. Other people’s data: no comparison group and no other applicants. No model weights, prompts or system instructions. Exact thresholds where they are trade secrets: then we name category and direction (“debt-to-income ratio above the internal limit”), not the number, and document that in the registry record. The CJEU has clarified for Art. 15 GDPR that trade secrets do not block an explanation wholesale but have to be weighed. And: no free text from the model. The generated rationale stays in the evidence for the reviewer and never goes to the person.
The interpretability gap. I am deliberately sober here. When a language model prepares the decision, the “main elements” we can honestly name are the retrieved documents, the rules applied and the passages the human confirmed. Not why the model rated a criterion the way it did. We write that limit into the explanation itself. For the scoring model, feature contributions are approximations; the method is in the registry record, and the quality gate checks before go-live whether the contributions are stable across similar cases. Claiming more than the evidence supports is, to me, a compliance risk, not a service.
GDPR and response times. For us, paragraph 3 means one process for both worlds. If the decision is solely automated, the request runs as an access request under Art. 15(1)(h) GDPR, with the same evidence and the same template, extended by logic, significance and consequences. If a human decided, it runs under Article 86. Because the AI Act names no deadline, we adopt the month from Art. 12(3) GDPR as the upper bound. The target is a few working days: the explanation is a query with sign-off by the business unit, not research.
The auditor sees the explanation view with version and signature, per request the timestamp, the response and the evidence ID it was generated from. More on how this fits into our AI governance.
Typical mistakes
EU AI Act Article 86 checklist
Open questions
Frequently asked questions
Does this apply to us as a deployer if a human always decides?
Yes, especially then. Art. 22 GDPR covers only solely automated decisions. Article 86 only requires that the deployer’s decision rests on the output of an Annex III system, and it does when the human had the model recommendation in front of them.
Do we have to disclose our score or the thresholds?
What is required is the role of the system and the main elements of the decision, not the model parameters. Which features and rules tipped the decision belongs in; exact thresholds and weights can be trade secrets, in which case we name category and direction.
How fast do we have to respond?
Article 86 names no deadline. We adopt the month from Art. 12(3) GDPR as the upper bound and measure ourselves in working days, because the explanation is generated from the evidence and does not have to be researched. The other articles of the series: all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments