A cobalt-blue pivot opens a passage through terracotta walls. Original editorial painting illustrating the possibility of reconsidering a consequential decision.
A consequential decision should come with an understandable reason, a reachable reviewer and a practical route to correction.
By Arbab Naseebullah Kasi, Chief Executive Officer - CEO, Feel Worldwide Foundation Inc.
Consider this fictional illustration. A worker applies for an evening training course. An automated eligibility check rejects the application because an old record places the worker outside the service area. The worker has moved, meets the published criteria and has the necessary evidence. The rejection arrives immediately. Finding someone who can correct it takes much longer.
The important question is what happens between those two moments. Can the applicant see the address that was used? Is there a clear way to submit the current one? Can a reviewer change the decision before enrollment closes? Will the correction prevent the same refusal next time?
A system can process applications quickly while making the consequences of a mistake painfully slow. For the person affected, the experience is measured in a missed opportunity, an interrupted service or hours spent trying to reach someone who has authority to help.
Organizations using automated decisions should plan that encounter as carefully as they plan the original decision. Contestability means giving people a usable way to question an outcome and have relevant evidence considered. It requires an operating process, not simply a promise that someone can contact support.
Begin with the decision a person actually receives
The word algorithm can conceal several different arrangements. A fixed rule might reject every incomplete form. A statistical model might assign a risk score. A staff member might combine that score with other evidence. A vendor might supply a recommendation that becomes a decision only after an organization acts on it.
Map the whole chain. Identify who sets the eligibility policy, who supplies the information, what the system calculates and who decides what happens to the person. A good review process needs access to each relevant part. Rechecking a score cannot repair a policy applied to the wrong situation.
Three questions help distinguish the problem. Was the information wrong? Was a valid rule applied incorrectly? Or does the rule itself produce an unfair, unlawful or inappropriate result in this context? These questions may lead to different remedies and different people with authority to provide them.
Someone correcting a misspelled name should not need to make the same argument as someone challenging a discriminatory criterion. Nor should a frontline employee be expected to resolve a legal or policy issue beyond their role. The organization needs a route from the first contact to the right decision-maker, with responsibility remaining clear throughout.
Explain enough for someone to act
An explanation should help the recipient locate the disagreement. “You did not meet our requirements” leaves almost everything hidden. A useful notice identifies the decision, the main reasons that actually affected it, the relevant information used and the available next step, subject to lawful limits on disclosure.
In the United States, credit provides a specific legal example. Regulation B requires applicable adverse-action notices to provide specific reasons or disclose the right to obtain them. Its text says that references to internal standards or a failure to achieve a qualifying credit score are insufficient. The official commentary requires reasons to accurately describe factors actually considered or scored. These are credit-specific obligations, with procedural variations and exceptions, rather than a general appeal right covering every automated service. Regulation B, section 1002.9.
A notice should also distinguish an observed fact from an inference. An overdue payment appearing in a record is different from an estimate of future repayment risk. A person may dispute the record, the connection between the record and the decision, or both. Presenting an estimate as an established fact makes that conversation harder.
Explanations must remain faithful to the actual decision. A reassuring paragraph generated after the event is of little use if it describes reasons the system did not use. Organizations should be able to connect the explanation to the relevant application, information and decision process, without exposing other people's private information.
Know which rights apply, and where
Legal protection depends on jurisdiction, sector, the kind of decision and how it was made. An organization should obtain qualified advice on its own obligations. People seeking help should check the applicable notice and official complaint or appeal route, especially where deadlines are short.
Under the European Union's GDPR, Article 22 addresses decisions based solely on automated processing that produce legal effects or similarly significant effects. It includes exceptions for contractual necessity, authorization by Union or Member State law, and explicit consent. For the contractual and consent exceptions, Article 22(3) requires safeguards including human intervention, the opportunity to express a viewpoint and the ability to contest the decision. The law-authorized exception must itself provide suitable safeguards. This is a defined legal framework, not a promise that every digital recommendation everywhere carries an identical right of appeal. GDPR, Article 22.
In February 2025, the Court of Justice of the European Union clarified meaningful information in the Dun & Bradstreet Austria case. In the covered automated-decision setting, the explanation must make the procedure and principles actually applied understandable. Merely supplying a complex formula or an exhaustive technical description is insufficient. The judgment also requires contested claims about trade secrets or third-party data to be considered by the competent supervisory authority or court, balancing the rights involved. It does not require indiscriminate public release of a company's software. Judgment in Case C-203/22.
These examples offer a practical lesson even where their legal provisions do not apply: information has value when it enables a person to understand and challenge what happened. An organization can choose to provide a fair review process beyond the legal minimum, while being precise about what that process can and cannot do.
Give the reviewer authority, time and evidence
Adding a person to the process does not automatically make it fair. A reviewer who sees only the original recommendation, has seconds to respond and cannot change the outcome may simply repeat the system's conclusion.
Meaningful review requires access to the relevant evidence, enough training to recognize the system's limitations and authority to take an appropriate action. That action might be correcting an input, reconsidering eligibility, requesting specialist advice, suspending an adverse step or escalating a policy concern. The reviewer should record the reasoning rather than merely checking a box.
The Federal Trade Commission's Rite Aid case shows why the actions surrounding an automated output matter. Its December 2023 complaint alleged that facial-recognition alerts led employees to wrongly accuse, search or exclude consumers. One enrollment generated more than 900 match alerts across more than 130 stores during five days. That is an illustration of implausible matching, not an overall accuracy estimate: the complaint did not supply a complete denominator from which to calculate a system-wide error rate. FTC complaint, particularly paragraphs 4 and 31.
The 2024 stipulated order includes, for covered biometric systems where their use is permitted under the order, consumer notices, several complaint channels, investigation of both inaccurate outputs and operators' responses, and written findings. Its requirements are specific to the case. They show how accountability can extend beyond receiving a complaint to examining the conduct that followed a signal. Stipulated order, Part IV.
A fair process should allow staff to disagree with the machine without being punished for slowing throughput. It should also check human decisions. Reviewers can misunderstand evidence, favor familiar applicants or apply inconsistent standards. Independence, supervision and reasoned records matter on both sides of the screen.
Look at denials, including the ones nobody contests
Appeals alone cannot tell an organization how well its decisions work. People who know the rules, have time to gather documents and can persist through a difficult process are more likely to reach review. Silence from everyone else is ambiguous.
An official healthcare review provides a useful warning about denominator choice. In an April 2022 report, the U.S. Department of Health and Human Services Office of Inspector General examined a stratified random sample of 250 prior-authorization denials and 250 payment denials from 15 large Medicare Advantage organizations during June 1–7, 2019. It estimated that 13% of the denied prior-authorization requests met Medicare coverage rules. The relevant population was denied requests in that review's scope, not all requests or all patients. HHS OIG report OEI-09-18-00260.
This was a review of coverage decisions, not an experiment isolating artificial intelligence. The figure must not be presented as an algorithmic error rate. Its relevance here is methodological: examining case files can reveal problems that a total count of decisions or complaints would miss. GAO subsequently summarized this finding with its sample and period in its own oversight report. GAO's May 2025 report.
Organizations should therefore examine a justified sample of unchallenged denials as well as appeals, using suitable expertise and privacy safeguards. Ask what the available evidence supports, whether the explanation matches the decision and whether the affected person had a realistic opportunity to respond. Look for repeated causes rather than treating every correction as an isolated exception.
Build a route that works under pressure
Return to the fictional training applicant. A well-designed process could identify the address and geographic eligibility criterion used, offer an accessible way to submit a correction and give a named team responsibility for reviewing it. Where lawful and feasible, the organization could preserve the application while the correction is considered. It should explain whether a place can actually be held, rather than allowing the applicant to assume that review guarantees admission.
The person should receive a case reference, the expected response time and a way to flag an urgent deadline. Provide a route that does not depend entirely on the same digital account or interface that failed. Depending on the service, this may include telephone, post, in-person assistance, translation or accessible formats.
Ask for evidence proportionate to the dispute. Requiring an entire personal history to correct one address adds burden and privacy risk. Where representatives are allowed, make the authorization process clear. Where identity checks are needed, explain what is required and offer supported alternatives when a standard document is unavailable.
Distinguish internal review from external remedies. A courtesy reconsideration does not necessarily pause a statutory filing deadline or replace an appeal to an independent body. Notices should say so when relevant and direct people to accurate information. Do not send someone around an internal loop until an external option expires.
Urgency should change the workflow. A disagreement about next year's membership renewal and a decision interrupting an essential service cannot always wait in the same queue. Establish lawful interim measures and escalation rules before the urgent case arrives. Be candid about limits, especially where an organization lacks authority to suspend a decision.
Make correction reach the next decision
A corrected letter is only part of the remedy. If an inaccurate record remains in the operational system, the person may face the same refusal again. Identify which information must be corrected, which dependent decisions need reconsideration and which recipients should be notified under the applicable rules and authorized arrangements.
The GDPR offers a specific example: Article 16 addresses correction of inaccurate personal data, while Article 19 generally requires communicating qualifying corrections to recipients, subject to stated exceptions. These provisions have their own scope and do not mean that any disputed prediction must automatically be replaced by the person's preferred result. GDPR, Articles 16 and 19.
Operationally, someone should verify that the correction has reached the places where it matters. An appeal marked successful while a denial remains active is an unfinished task. Where harm has already occurred, identify any further remedy available, rather than promising compensation or restoration that the organization cannot provide.
NIST's voluntary AI Risk Management Framework 1.0 treats these responsibilities as part of ongoing management. Its Manage 4.1 outcome includes mechanisms for feedback, appeal and override, incident response and recovery after deployment. It is guidance for managing risk, not a law creating individual appeal rights or a certification that a system is safe. NIST's framework core.
Use that perspective to close the loop. If several people encounter the same faulty criterion, fix the source and identify others who may have been affected. The person who persisted should not be the only beneficiary of discovering the mistake.
Measure access to remedy, not just speed
Consider a deliberately hypothetical monthly review, not a Foundation result or a real organization's data. An organization issues 1,000 denials. It receives 100 appeals, completes 80 by the reporting date and changes 20 of those completed decisions.
The appeal rate is 100 divided by 1,000, or 10%. The change rate among completed appeals is 20 divided by 80, or 25%. Another 20 appeals remain unresolved. None of these figures establishes that 25% of all denials were wrong. The appealed cases may differ from the others, and some outcomes may change because new circumstances arose rather than because the original decision was incorrect.
Report those distinctions. Track time to first meaningful response, time to final action, unresolved urgent cases and whether a successful review produced an operational correction. Record why decisions changed: inaccurate information, procedural failure, inappropriate application of a rule, new evidence or another documented reason.
Pair the counts with people's experiences. Could they understand the notice? Did they know what evidence would help? Could they reach a person without repeatedly retelling the story? Include those who abandoned the process where ethical and feasible. Review differences between groups only with appropriate legal authority, safeguards and enough context to avoid misleading conclusions from small numbers.
Low appeal volume may reflect good decisions. It may also reflect an inaccessible process. A high reversal rate may reveal a problem in the first decision or a review route successfully reaching those who need it. Interpretation requires evidence beyond the headline percentage.
The standard worth pursuing is practical. A person should be able to find out what happened, bring relevant evidence to someone who can act and receive a reasoned outcome while it still matters. Automation can support that work. Responsibility for making it possible remains with the organization using the decision.
If your organization wants to explore a practical contribution to accessible digital services or responsible technology education, submit a partnership inquiry to Feel Worldwide Foundation Inc.. Describe the proposed contribution, intended participants and scope so its fit and feasibility can be reviewed. To offer financial support, visit the Foundation's donation page and review the available giving options and terms.
