A refusal can now be produced in the time it takes a server to compare a record with a rule. Contesting that refusal may still require a person to decipher a letter, find the responsible office, collect documents the administration may already hold, and wait without knowing whether anyone has opened the case. The authority saves minutes at scale; the individual pays in hours, uncertainty and sometimes lost income.
That imbalance is not an accidental side effect of automation. It is a design choice. Before an institution automates a consequential decision, it should build the route by which a person can understand, challenge and correct it. The appeal does not need an algorithm to decide who is right. It needs joined-up machinery for notification, explanation, evidence, tracking and genuinely empowered human review.
Automate the appeal first, and the institution has to confront what its decision system actually knows, how errors travel and who can put them right. Automate only the decision, and those questions are exported to the person least equipped to answer them.
The strongest case for automating decisions
Public bodies make many decisions under pressure: whether an application is complete, a payment is due, a permit condition has been met or a case should be routed for further examination. Where the law supplies clear criteria and reliable data, software can apply routine checks quickly and consistently. It can reduce queues, remove repetitive clerical work and leave trained staff more time for cases that demand interpretation.
That is a serious argument. A tired official working through a backlog is not an automatic guarantee of fairness, and a slow service can itself deny people practical access to a right. A well-bounded rules engine may also leave a more useful record than an undocumented human shortcut.
But the case is strongest on the happy path: the right person has been matched to the right record, the data are current, the rule is correctly encoded, the evidence fits the expected format and the case contains no exceptional circumstance. Human cost is determined by what happens when one of those assumptions fails.
A fast decision can create a slow error
An adverse result can have several different causes. A source register may hold an old address. A document may have arrived after a data extract was taken. Two people may have been incorrectly matched. A policy rule may have changed while one service still runs an earlier version. A model may assign a score that is reasonable in aggregate but wrong for this case. Or the facts may be accurate while the rule simply does not capture an unusual but legitimate situation.
Those failures require different remedies. A missing document can be added to the case. An inaccurate source record must be corrected where it originates and the correction propagated. A coding fault needs a system-wide fix and a search for other affected decisions. An exceptional circumstance needs judgement, not another pass through the same rule.
This is why a generic button marked “contact us” is not an appeal system. A usable decision notice should arrive through an accessible channel and identify the responsible authority, the decisive facts, the source of those facts, the rule or model version used, the date of the decision, the deadline for challenge and a case reference that survives across departments. It should explain enough for the person to see whether the disagreement concerns data, evidence, interpretation or law.
European law illustrates parts of this principle without supplying a universal blueprint. The General Data Protection Regulation gives a qualified right not to be subject to certain solely automated decisions and, in specified circumstances, safeguards including human intervention, the opportunity to express a view and the ability to contest a decision. Its access provisions also address meaningful information about the logic, significance and envisaged consequences of covered automated processing. The scope and exceptions matter; these are not blanket rights against every use of software.
A February 2025 judgment of the Court of Justice of the European Union sharpened what “meaningful information” can require in a credit-profile case. The explanation should make it possible to understand which personal data were used and how, rather than hiding behind the complexity of the operation. The judgment concerns Article 15 of the GDPR, not every administrative decision, but its practical lesson travels: an explanation should help someone identify a contestable input or inference.
The appeal is one journey
Once notified, a person should be able to submit evidence without starting a second, disconnected case. The system should acknowledge exactly what was received, preserve the original decision and its provenance, show the case’s current state, and say what will happen next. If another register owns the disputed fact, the authority should route the correction rather than instruct the person to diagnose the government’s architecture.
Tracking is more than a convenience. It changes whether someone can act. A person facing a suspended payment needs to know whether enforcement is paused, whether interim support is available and when a reviewer will respond. Silence forces repeated calls, duplicate submissions and defensive escalation. A status trail also lets the institution see where cases stall.
Human review must be real. A reviewer needs authority to disregard, override or reverse a machine output, enough context to understand the system’s limits, and time to consider new evidence. Simply showing a score to an official and asking for confirmation can convert automation bias into a rubber stamp. The EU AI Act’s rules for covered high-risk systems expressly treat human oversight as a capacity to understand limitations, avoid over-reliance and override outputs; its Article 86 also establishes a right, within its defined scope, to clear and meaningful explanations of AI’s role and the main elements of a significantly adverse decision.
Measure recovery, not just accuracy
Accuracy remains important, but a single percentage can conceal the experience of failure. An institution should measure how long it takes an adverse decision to reach the person in an understandable form, how long a challenge waits for acknowledgement and human review, how often evidence has to be supplied twice, and how quickly a corrected source record changes downstream decisions.
It should also publish or oversee patterns: which reasons produce appeals, how often decisions are changed, whether the same error recurs, and whether waiting times or outcomes differ by channel, language, disability or location. A rising overturn rate could reveal a defective rule; a low appeal rate could indicate accuracy, or an unusable appeal route. Both require investigation.
The burden is asymmetric. An authority spreads the cost of automation across thousands of cases. Each affected person encounters a single unfamiliar system at a moment that may already be financially or emotionally difficult. Digital access, literacy, available time and confidence dealing with institutions are unevenly distributed. A redress process that is technically open but practically exhausting selects for those best able to persist.
Guidance in one narrower domain makes the point plainly. The OECD’s March 2026 financial consumer protection compendium says complaint and redress mechanisms should not impose unreasonable cost, delay or burden, should account for vulnerability, and should expose aggregate complaint and resolution information. Financial services are not public administration, but the operational test is useful: can the least advantaged person use the route before the harm compounds?
The counterargument is about delay
Requiring a complete appeal system before automating decisions could postpone useful services. Easy challenges might invite speculative complaints, overload reviewers or reveal controls to fraudsters. These are real risks. A correction path still needs identity checks, evidence controls and proportionate disclosure; not every case should halt enforcement, and not every technical detail can be made public.
Yet “automate the appeal” does not mean automating the legal judgement or building an elaborate tribunal before testing a simple administrative rule. It means creating reusable infrastructure for intake, receipts, provenance, routing, status and escalation. It means deciding in advance which harms require a pause, which cases need independent review and which staff have authority to correct data or reverse outcomes. Those capabilities improve human decisions too.
If this argument is useful to your work, you can subscribe to Alkemata for more essays on technology designed around human capability rather than administrative convenience.
Continue exploring
A citizen-facing appeal cannot work if the institution cannot trace data back to its source. What Would a Citizen Control Panel for Government Data Actually Require? examines the registers, logs, correction routes and delegated access behind that promise.
The decisive procurement question is therefore not only how many cases a system can process or how accurately it performs in testing. It is how quickly, intelligibly and completely the institution can recover when a person says the system is wrong. Until that recovery standard is specified and tested with the people most likely to bear its burden, the automated decision is not ready.