Start with the decision, not the model
An AI-assisted decision is one in which a system helps a person or organization choose an outcome. It might sort applications, flag a customer request for review or suggest which submissions meet a set of criteria. The AI may not make the final choice, but its output can still affect what a person sees and how quickly they act. An appeal process should therefore cover the whole decision, not just whether the model ran as intended.
Begin by mapping the path from information received to outcome communicated. Note where AI was used, what it produced, who could override it and which records were available to the decision-maker. Then ask a practical question: if someone says the outcome is wrong, who can examine it and what could they change? If no one has the authority or information to do that, a contact form alone is not an appeal process.
Tell people enough to ask a useful question
A decision notice should state the outcome in plain language and explain whether AI played a role. Describe that role accurately: “AI suggested a category that a staff member reviewed” means something different from “AI automatically assigned the category.” Give the main criteria that mattered where you can, along with any relevant information the person supplied. Avoid presenting a generated explanation as a definitive account of a model’s internal reasoning.
The notice should also say what the person can do next. Provide an accessible route to request review, a way to submit a correction or missing context, and a realistic expectation for when they will hear back. People should not have to know technical terms or guess which department owns the system. If there is a time limit or a required format, make it prominent and provide an alternative for someone who cannot use the default channel.
Make review capable of changing the outcome
Meaningful review is more than having a person confirm the AI output. Give the reviewer the original materials, the criteria for the decision, a record of the AI’s contribution and the person’s explanation of what may be wrong. The reviewer should be able to consider new context, correct inaccurate inputs and reach a different outcome. For decisions with substantial consequences, it may help to assign the review to someone who did not make the initial choice.
Set out a simple sequence for the team: acknowledge the request, identify the disputed point, check the underlying information, reconsider the decision and communicate the result. Distinguish a data error from a disagreement about judgment. For example, a misread document may call for correction and a fresh assessment; a disputed interpretation may call for a second person to apply the criteria. If a decision cannot be reversed, be honest about what remedy is possible rather than promising an appeal that cannot help.
Protect privacy while keeping a useful record
An appeal may involve sensitive information. Ask only for details needed to locate and review the decision. Do not require people to repeat personal information that the team already has unless it is needed to confirm identity or clarify a disputed fact. Limit access to appeal records, use appropriate channels for documents and avoid pasting private submissions into general-purpose AI tools without an approved reason and safeguards.
Keep a record that lets the team explain what happened: the original outcome, the AI’s role, the information reviewed, the appeal request and any change made. Decide who may see that record and how long it needs to be kept. A record supports consistent review and helps identify repeated problems, but storing everything indefinitely creates its own risks.
Use appeals to improve the workflow
Individual appeals can reveal weaknesses that routine checks miss. Look for recurring input errors, unclear criteria, misleading decision notices or situations in which reviewers rarely depart from an AI suggestion despite having that authority. Do not assume that a low number of appeals means decisions are fair; people may not know they can question an outcome, or the process may be too difficult to use.
For a small team, the first version can be modest: a clear notice, one reliable contact route, a named owner, a review checklist and a way to record outcomes. Test it with an ordinary example and a case involving missing or incorrect information. The goal is not to guarantee that every appeal succeeds. It is to give people a genuine opportunity to be heard and give the team a defensible way to find and fix mistakes.
