Adverse Automated Decision Explanation Letter

How to use this document

• Use it to explain in writing to a person why a decision made wholly or partly by an AI or scoring system went against them (denial of credit, insurance, a job or a service, or an account restriction), what data was used and how to request human review.
• Suggested owners: the system owner (or Human Resources for employment decisions) for sending, Legal for approving the text and the system owner's data team for the reason library.
• This letter is the written T5 response under the "AI Transparency and User Explanation Policy", which sets the circuit (channel, deadline, owner and register of explanations). Any subsequent human review is handled under the "Automated Decision Review Procedure", and the catalog of rights is in the "AI User Rights Notice".
• Before adopting it, decide which decisions trigger the letter automatically and which only on request, who signs, how soon it is sent and through which channel.
• Complete the base letter in section 3, add the block from section 4 that matches the type of decision and the variant for the person's country.
• Reasons are taken from the library in section 5, which must be built from the model's actual variables and validated with the data team.
• Keep only the variants for the countries where you make decisions about people. In the United States, credit letters and letters based on consumer reports have specific elements: use block 4.5 and review it with your advisor.
• Ask internal legal or an outside advisor to review the final text, especially each country's deadlines and mandatory statements.
• Approve the letter in the [AI COMMITTEE] or the body responsible for the decision process, record it with a version number and keep a copy of every letter sent in the register of explanations delivered.

Adverse Automated Decision Explanation Letter

1. Purpose and use

1.1. [ORGANIZATION NAME] uses this letter to explain to people adverse decisions in which the AI system [AI SYSTEM] is involved, so that they can understand the main reasons, correct inaccurate data and ask for a person to review the decision.

1.2. The letter is sent:

• automatically, for the decisions the organization has defined or where the country's law provides for it; and
• on request, for any other decision in which the system was involved.

1.3. Internal deadline for sending: [COMPLETE: number] days from the decision or the request, or the shorter period set by the country's law.

1.4. Each letter is linked to the "Automated Decision Log Standard" by its identifier, so that it is possible to check which model version, data and reasons were used, and is entered in the register of explanations delivered under the "AI Transparency and User Explanation Policy".

2. Drafting rules

1. Specific reasons. State the actual principal reasons for the decision, usually two to four, in order of importance. Do not use generic wording such as "you did not meet our internal standards" or "system decision."
2. Plain language. Short sentences, no technical jargon or variable names. If a technical term is needed, explain it.
3. Role of the system. Say whether the decision was fully automated or reviewed by a person, and how much weight the system's output carried.
4. Data used. State the categories of data and their sources (provided by the person, generated by the relationship, obtained from third parties).
5. What could change the outcome. Where possible and not misleading, state which factors, if different, would have changed the decision.
6. Trade secrets. Protecting trade secrets does not justify refusing any explanation. Describe the logic and factors without revealing code or exact weights, and reserve detailed information for the authority or court if they request it.
7. Consistency. The reasons match those in the decision log. Never invent "more presentable" reasons.
8. Rights and avenues. Explain how to request human review, how to provide information or correct data and how to complain to the authority.
9. Accessibility. Send the letter in the language of the relationship with the person and in an accessible format on request.