Automated Decision Contestability Governance

2. Summary

Automated Decision Contestability Governance requires that every AI agent decision that produces legal effects or similarly significant effects on individuals is subject to a structured contestability process — enabling the affected individual to receive a meaningful explanation, express their point of view, challenge the decision's accuracy, and obtain human review that can override the automated outcome. The dimension addresses a fundamental asymmetry: an AI agent can make thousands of consequential decisions per hour, each based on opaque pattern matching across data the individual cannot see, while the affected individual has no way to understand why the decision was made or how to challenge it. Without AG-062, automated decisions become unchallengeable facts — a credit application denied, an insurance claim rejected, a benefits determination made, an employment screening failed — with no effective mechanism for the individual to contest the outcome. AG-062 requires that contestability is built into the decision architecture from the start, not offered as an afterthought when a complaint arrives. The right to contest is not meaningful if exercising it requires understanding the inner workings of a neural network; AG-062 requires that explanations are provided in terms the data subject can understand and act upon.

3. Example

Scenario A — Credit Application Denied Without Meaningful Explanation: An AI agent for a consumer lending platform processes 12,000 credit applications per day. The agent uses a gradient-boosted decision tree model trained on 4.2 million historical applications. A 34-year-old applicant with stable employment and no adverse credit history is declined. The automated explanation provided is: "Your application was assessed against our lending criteria and unfortunately was not successful." The applicant calls the lender's customer service line and asks why they were declined. The customer service representative can see only the agent's final decision (decline) and a risk score (0.73 out of 1.00) — they cannot see which factors influenced the score. The representative tells the applicant: "The decision was made by our automated system. I can't see the specific reasons." The applicant complains to the Financial Ombudsman Service, which asks the lender to produce the reasoning behind the decision. The lender discovers that the decline was driven primarily by the applicant's postcode (a known proxy for ethnicity in the UK) and the frequency of address changes in the past 5 years (the applicant relocated twice for work). Neither factor was disclosed to the applicant.

What went wrong: The AI agent made a consequential decision (credit denial) with no meaningful explanation. The customer service pathway could not access decision-level explanations. The applicant had no effective mechanism to contest the decision or present additional information. The factors driving the decision included a potential proxy for protected characteristics (postcode), which could not be challenged because the applicant never knew it was a factor. Consequence: Financial Ombudsman Service complaint upheld, £350 compensation awarded to the applicant, FCA supervisory enquiry into the lender's AI decision-making processes, £1.4 million remediation programme to implement decision explanations and review the model for proxy discrimination, and 2,300 historical decisions flagged for manual re-review.

Scenario B — Benefits Determination With No Appeal Route: A public sector AI agent automates initial eligibility determinations for disability benefits. The agent assesses medical evidence, employment history, and functional capacity scores to produce an eligibility recommendation. In 87% of cases, the recommendation is accepted by the human decision-maker without independent review (the human clicks "approve" on the agent's recommendation within an average of 12 seconds per case). An applicant with a complex chronic condition receives a "not eligible" determination. The determination letter states: "Based on the evidence provided, you do not meet the eligibility criteria for this benefit." The applicant requests an appeal. The appeal process requires the applicant to identify specific errors in the determination — but no explanation of the determination logic is provided. The applicant's representative requests a breakdown of how the functional capacity score was calculated; the department cannot provide one because the scoring is performed by the AI agent and the human decision-maker did not independently assess the evidence.

What went wrong: The AI agent's recommendation was treated as the decision in practice, despite a nominal human review step. No meaningful explanation was provided to the applicant. The appeal process required the applicant to identify errors they could not see. The human review was perfunctory (12-second average review time), meaning the automated decision was effectively final. Consequence: Judicial review of the determination process, court finding that the appeal process was not effective because the applicant could not meaningfully contest a decision they could not understand, order to implement a genuine contestability mechanism, 4,200 historical determinations reviewed manually at a cost of £3.8 million.

Scenario C — Insurance Claim Rejected Based on Uncontestable Fraud Score: An AI agent for a motor insurance company processes claims and assigns a fraud risk score. Claims scoring above 0.85 are automatically referred to the Special Investigation Unit (SIU), which delays payment by an average of 4 months. A policyholder submits a legitimate claim for a vehicle collision. The agent assigns a fraud score of 0.91, triggering SIU referral. The fraud score is driven by three factors: the claim was filed on a Monday (statistically correlated with staged collisions in the training data), the policyholder has held the policy for less than 6 months, and the repair estimate exceeds £8,000. None of these factors indicate actual fraud — the collision occurred on a Monday because that is when the policyholder commutes, the policy is recent because the policyholder recently purchased the vehicle, and the repair cost reflects genuine damage. The policyholder is not informed of the fraud referral, the fraud score, or the factors driving it. The SIU investigation takes 5 months and finds no evidence of fraud. The claim is paid, but the policyholder has spent 5 months without their vehicle, incurred £2,400 in alternative transport costs, and experienced significant distress.

What went wrong: The AI agent made a consequential determination (fraud referral) with significant effects on the policyholder (5-month payment delay). No explanation was provided. No mechanism existed for the policyholder to contest the fraud score or provide information that would have reduced it. The factors driving the score were individually innocuous and would have been easily explained if the policyholder had been given the opportunity. Consequence: FCA Consumer Duty complaint (failure to deliver good outcomes), Financial Ombudsman Service award of £4,800 (alternative transport costs plus distress compensation), requirement to implement fraud score explanations and a pre-referral contestability mechanism, and review of 8,400 historical fraud referrals for similar patterns.

4. Requirement Statement

Scope: This dimension applies to all AI agents that make, recommend, or materially influence decisions that produce legal effects or similarly significant effects on individuals. Legal effects include: credit decisions, insurance underwriting or claims decisions, employment screening or termination recommendations, benefits eligibility determinations, criminal risk assessments, immigration assessments, and regulatory enforcement decisions. Similarly significant effects include: decisions that significantly affect the individual's financial position, access to services, health treatment, educational opportunities, or personal autonomy. The scope covers decisions that are nominally reviewed by a human but where the human review is substantively rubber-stamping the agent's recommendation — the test is whether the human reviewer independently evaluates the underlying evidence, not merely whether a human clicks "approve." The scope also covers scoring and profiling activities where the score or profile is used as input to a consequential decision, even if the final decision is made by a different system or human.

4.1. A conforming system MUST identify and classify all agent decisions that produce legal effects or similarly significant effects on individuals, and MUST apply the contestability requirements of this dimension to each classified decision.

4.2. A conforming system MUST provide the affected individual with a meaningful explanation of any automated decision, including: the principal factors that influenced the outcome, the data used as input, and the general logic of the decision process — expressed in terms that a non-technical person can understand and act upon.

4.3. A conforming system MUST provide the affected individual with an accessible mechanism to contest the decision — express their point of view, challenge the accuracy or relevance of the input data, and provide additional information that was not considered.

4.4. A conforming system MUST ensure that contested decisions are reviewed by a competent human who independently evaluates the underlying evidence and the individual's representations, and who has the authority and practical ability to override the automated decision.

4.5. A conforming system MUST ensure that the human reviewer is not presented with the automated decision in a way that creates anchoring bias — the reviewer must evaluate the evidence independently, not merely confirm or reject the agent's recommendation.

4.6. A conforming system MUST retain the decision record, explanation, contestation, and human review outcome for each contested decision, creating an auditable trail from initial decision through final resolution.

4.7. A conforming system SHOULD provide the explanation proactively at the time of the decision, rather than requiring the individual to request it.

4.8. A conforming system SHOULD implement the contestability mechanism as an integrated part of the decision notification process — the notification of the decision includes the explanation and the pathway to contest.

4.9. A conforming system SHOULD monitor human review patterns and flag reviewers who consistently confirm automated decisions without substantive analysis (e.g., average review time under 30 seconds, override rate below 2%).

4.10. A conforming system MAY implement a pre-decision contestability stage where the individual can review and correct the data inputs before the automated decision is made.

5. Rationale

Automated Decision Contestability Governance implements the fundamental right of individuals not to be subject to decisions made solely by automated processing that produce legal or similarly significant effects — and, where such decisions are permitted, the right to obtain meaningful information about the logic involved and to contest the decision (GDPR Article 22, EU AI Act Article 86, UK GDPR Article 22).

The rationale for contestability is not merely legal compliance — it is a matter of fundamental fairness and effective governance. An automated decision that cannot be contested is an exercise of unchallengeable power. Democratic societies have long recognised that consequential decisions about individuals must be subject to review: court judgements can be appealed, administrative decisions can be challenged by judicial review, employment decisions are subject to tribunal oversight. AI agents making consequential decisions at scale create the risk of a "black box bureaucracy" where decisions are made by systems that neither the decision-maker nor the affected individual can explain or challenge.

Contestability also serves as an error correction mechanism. AI models make mistakes — they reflect biases in training data, they fail on edge cases, and they can be influenced by factors that are individually reasonable but collectively discriminatory (as in the postcode-as-proxy example). Without a contestability mechanism, these errors accumulate uncorrected because no feedback loop exists. An individual who contests a decision and provides additional information creates a feedback signal that can improve not only their own outcome but the quality of future decisions. Contestability is therefore not only a rights mechanism but a quality mechanism.

The requirement for meaningful explanation is critical. A formal explanation that says "the decision was made based on your application data" provides no basis for contestation. The individual needs to know which factors mattered, what data was used, and how the logic works — expressed in terms they can understand and challenge. "Your application was declined primarily because your income-to-debt ratio of 42% exceeds our threshold of 35%, and your residential history shows 3 address changes in 24 months which our model associates with higher default risk" is meaningful. The individual can respond: "My income has increased by 30% since the data was collected" or "My address changes were due to work relocation, not financial instability." This dialogue is what contestability requires.

6. Implementation Guidance

AG-062 establishes the decision record as the central governance artefact. The decision record captures, for each consequential automated decision: the input data, the decision logic (at a level that can be explained to the individual), the principal factors, the output decision, and the contestability pathway. The decision record enables both explanation and review.

Recommended patterns:

• Decision explanation engine. Implement a dedicated explanation service that generates human-readable explanations for each automated decision. The service takes the decision model's output and the input features, applies an explanation method (SHAP values, LIME, counterfactual explanations, or rule extraction for tree-based models), and produces a structured explanation identifying the top 3-5 factors, their direction of influence (positive/negative), and the threshold or comparison point. For a credit decision: "Factor 1: Income-to-debt ratio (42%) — above the threshold of 35%, negative influence. Factor 2: Length of current employment (8 months) — below the preferred minimum of 12 months, negative influence. Factor 3: Number of address changes in 24 months (3) — above the typical range of 0-1, negative influence." The explanation is stored as part of the decision record and presented to the individual.
• Integrated contestation workflow. Build the contestation mechanism into the decision notification. When an individual receives a decision notification (e.g., credit decline letter, claim decision, eligibility determination), the notification includes: the explanation, a link to the contestation portal, and clear instructions on how to contest. The contestation portal allows the individual to: view the explanation, identify specific factors they wish to challenge, upload supporting evidence, and provide a written statement. Contested decisions are automatically routed to the human review queue with all materials attached. Target: the individual should be able to initiate a contestation within 3 clicks or interactions from receiving the decision notification.
• Independent human review with anti-anchoring design. Present the human reviewer with the individual's case materials, the contestation statement, and the input data — but not the agent's decision or recommendation. The reviewer makes an independent assessment before seeing the automated outcome. Only after recording their independent assessment is the reviewer shown the agent's decision, allowing them to confirm or override with full awareness of any discrepancy. This design prevents anchoring bias, where the reviewer unconsciously defers to the automated decision. Monitor review patterns: flag reviewers with average review times under 2 minutes for credit decisions, or override rates below 5%, for quality assurance review.
• Feedback loop from contestation to model improvement. Aggregate contestation outcomes to identify systematic patterns: if 30% of contestations for a specific decision type result in overrides, the model for that decision type has a systematic issue. Route aggregate contestation data to the model governance team for investigation. Track override rates by factor: if most overrides involve the same factor (e.g., "number of address changes"), that factor may be unreliable or discriminatory.

Anti-patterns to avoid:

• Explanation theatre. Providing an explanation that is technically accurate but practically useless — for example, listing 200 features and their weights without identifying which ones actually drove the decision. Explanations must be concise, highlighting the 3-5 principal factors, expressed in terms the individual can understand and challenge.
• Inaccessible contestation mechanisms. Requiring the individual to submit a formal written complaint, engage a solicitor, or navigate a complex bureaucratic process to contest a decision. The contestation mechanism must be proportionate to the decision: a consumer credit decision should be contestable through a simple online portal; it should not require a legal representative.
• Rubber-stamp human review. Implementing a nominal human review step where the reviewer sees the agent's recommendation prominently displayed and confirms it in seconds. This does not constitute genuine human review and does not satisfy the contestability requirement. The test is whether the reviewer independently evaluated the evidence, not whether a human was technically in the loop.
• Retrospective explanations only. Providing explanations only when the individual specifically requests one, rather than proactively at the time of the decision. Many individuals do not know they have the right to an explanation. Proactive explanation empowers contestation; reactive explanation limits it to the most informed and persistent individuals.
• Conflating contestability with complaint handling. Routing contested automated decisions through the general complaint process, where they are handled by staff without authority to override the automated decision and without access to decision-level explanations. Contested automated decisions require a specialist pathway with appropriately trained and authorised reviewers.

Industry Considerations

Financial Services. The FCA's Consumer Duty requires firms to deliver good outcomes for customers, including in automated decision-making. For credit decisions, Section 157 of the Equality Act 2010 and Consumer Credit Act requirements for adverse action notices create specific explanation obligations. AI credit scoring agents must provide explanations that satisfy both data protection requirements (GDPR Article 22) and financial regulation requirements (CRA adverse action notices in the US). The explanation must be specific enough for the consumer to understand and challenge — "your application did not meet our criteria" is insufficient under both GDPR and financial regulation. Firms should expect that contested credit decisions will be scrutinised by the Financial Ombudsman Service and should design the contestation process to produce records that demonstrate fair treatment.

Healthcare. AI agents making clinical recommendations (treatment suggestions, triage classifications, diagnostic assessments) create contestability requirements under both data protection law and clinical governance. Patients have the right to understand how their treatment was determined and to seek a second opinion. When an AI agent influences clinical decisions, the explanation must be in clinically meaningful terms — not model feature importance scores. "The AI system assessed your symptoms as consistent with a lower-priority presentation based on: reported pain level 4/10, no fever, and stable vital signs" allows the patient to contest: "My pain is actually 8/10 but I understated it." Clinical contestability must route to a qualified clinician, not a customer service representative.

Public Sector. Automated decisions by public authorities are subject to heightened scrutiny under administrative law, including requirements for procedural fairness and the right to reasons. In the UK, the common law duty to give reasons applies to significant administrative decisions, and the Equality Act 2010 public sector equality duty requires authorities to consider the impact on protected characteristics. AI agents making or influencing public sector decisions (benefits, housing, education, law enforcement) must provide explanations that satisfy these additional requirements. The Council of Europe's recommendation on AI in the public sector specifically calls for effective contestability mechanisms for automated administrative decisions.

Maturity Model

Basic Implementation — The organisation has identified automated decisions that produce legal or significant effects on individuals. Explanations are available upon request but are not provided proactively. The explanation is a generic description of the decision process rather than a specific account of the factors that influenced the individual's outcome. Contestation is handled through the general complaint process. Human review of contested decisions is performed by staff who see the agent's recommendation and typically confirm it. Average review time: under 60 seconds. Override rate: under 3%. This level meets the minimum legal requirement for a review mechanism but does not deliver meaningful contestability.

Intermediate Implementation — Explanations are provided proactively with each decision notification, identifying the top 3-5 principal factors in non-technical language. An integrated contestation workflow allows the individual to challenge specific factors, provide additional evidence, and submit a statement. Human reviewers are trained in independent assessment and are presented with case materials separately from the agent's recommendation. Override rates are monitored; reviewers with consistently low override rates or review times are flagged for quality review. Decision records, explanations, contestations, and outcomes are retained as auditable artefacts. Average review time: 8-15 minutes for credit decisions. Override rate: 8-15% of contested decisions.

Advanced Implementation — All intermediate capabilities plus: counterfactual explanations that tell the individual what would need to change for a different outcome ("if your income-to-debt ratio were below 35%, the decision would have been approve"); pre-decision contestability allowing individuals to review and correct input data before the decision is made; aggregate contestation analysis feeding back into model improvement; anti-anchoring review design verified through reviewer outcome tracking; independent annual audit of the contestability process including mystery-shopper testing of the contestation pathway; and monitoring of demographic patterns in contestation outcomes to detect potential discrimination. The organisation can demonstrate that its contestability mechanism delivers genuine review, not theatre.

7. Evidence Requirements

Required artefacts:

• Decision classification register. Inventory of all automated decisions that produce legal or significant effects on individuals, including: the decision type, the agent responsible, the data inputs, the decision criteria, and the contestability pathway. Updated whenever new decision types are deployed.
• Decision records with explanations. For each consequential automated decision: the input data, the principal factors, the explanation provided to the individual, the decision outcome, and the contestability pathway offered. Format: structured data with the explanation in both machine-readable and human-readable form.
• Contestation and review records. For each contested decision: the contestation statement, the evidence provided, the human reviewer's independent assessment, the final outcome (confirmed, modified, or overturned), and the reasoning for the final outcome. Minimum retention: 7 years for financial decisions, 5 years for other decision types.
• Human review quality metrics. Monthly statistics showing: number of decisions contested, average review time, override rate (percentage of contested decisions where the human reviewer changed the outcome), and distribution of review times. Reviewers with anomalous patterns flagged and investigated.

Retention requirements:

• Decision records and explanations: minimum 7 years for regulated financial services; minimum 6 years for public sector decisions; minimum 5 years for other sectors.
• Contestation and review records: same as decision records.

Access requirements:

• Individual decision records and explanations producible to the affected individual within 30 days of request. Aggregate statistics producible to regulators within 72 hours of request.

8. Test Specification

Testing AG-062 compliance requires verifying that explanations are meaningful, contestation pathways are accessible, and human review is genuine.

Test 8.1: Explanation Specificity

• Stimulus: Generate automated decisions for 10 test cases with known input data and known principal factors. Examine the explanations produced for each decision.
• Expected behaviour: Each explanation identifies the top 3-5 factors that influenced the decision, expressed in non-technical language. The factors identified match the known principal factors for each test case.
• Pass criteria: Explanations are specific to the individual case (not generic), identify the correct principal factors, and are expressed in terms a non-technical person can understand. An independent reviewer (non-technical) can read the explanation and articulate why the decision was made.
• Fail criteria: Explanations are generic (the same text for different decisions), fail to identify the principal factors, or require technical expertise to understand.

Test 8.2: Contestation Pathway Accessibility

• Stimulus: From the decision notification, attempt to initiate a contestation as a non-technical user. Measure the number of steps, the time required, and the information needed.
• Expected behaviour: The contestation can be initiated within 3 clicks or interactions from the decision notification. No technical expertise is required. The pathway is accessible on mobile devices and meets WCAG 2.1 AA accessibility standards.
• Pass criteria: Contestation initiation within 3 interactions. No specialist knowledge required. Pathway is accessible and available 99.5% of the time.
• Fail criteria: Contestation requires more than 5 interactions, requires technical or legal knowledge, or is not accessible to users with disabilities.

Test 8.3: Human Review Independence

• Stimulus: Submit 20 contested decisions to human reviewers: 10 where the automated decision is correct, and 10 where the automated decision is incorrect (determined by pre-defined ground truth). Verify whether reviewers are presented with the automated decision during their initial assessment.
• Expected behaviour: Reviewers are presented with case materials and the contestation statement, but not the automated decision, during their initial assessment. Reviewers override the incorrect decisions at a rate significantly higher than the baseline override rate.
• Pass criteria: The review interface does not show the automated decision during initial assessment. Reviewers identify at least 60% of the incorrect decisions (threshold accounts for genuine ambiguity in some cases). Average review time exceeds 5 minutes for complex decisions.
• Fail criteria: Reviewers see the automated decision during initial assessment (anchoring risk), override rate does not differ between correct and incorrect cases, or average review time is under 60 seconds.

Test 8.4: Decision Record Completeness

• Stimulus: Generate an automated decision, contest it, and complete the human review process. Examine the complete decision record chain.
• Expected behaviour: The decision record contains: input data, principal factors, explanation, decision outcome, contestation statement, evidence provided, reviewer's independent assessment, final outcome, and reasoning.
• Pass criteria: All required fields are populated. The record is sufficient for an auditor to reconstruct the full decision and contestation history.
• Fail criteria: Any required field is missing, or the record does not support full reconstruction of the decision and contestation process.

Test 8.5: Proactive Explanation Delivery

• Stimulus: Generate an automated decision and examine the decision notification sent to the individual.
• Expected behaviour: The notification includes the explanation and the contestation pathway without the individual needing to request them.
• Pass criteria: Explanation and contestation instructions are included in the initial notification. No separate request is needed.
• Fail criteria: The notification contains only the decision outcome without explanation, or the individual must request the explanation separately.

Test 8.6: Contestability Under Load

• Stimulus: Submit 500 contestation requests within a 1-hour period. Verify that the system processes all requests and routes them to human review within defined timescales.
• Expected behaviour: All contestation requests are acknowledged, logged, and queued for human review. No requests are lost or delayed beyond the defined acknowledgement window (24 hours).
• Pass criteria: 100% of requests acknowledged within 24 hours. All requests appear in the human review queue. System performance does not degrade under load.
• Fail criteria: Any request is lost, acknowledgement takes longer than 24 hours, or system performance degrades to the point where the contestation pathway is inaccessible.

Test 8.7: Review Pattern Monitoring

• Stimulus: Simulate a reviewer who consistently confirms automated decisions with an average review time of 15 seconds and an override rate of 0.5%. Verify that the monitoring system flags this pattern.
• Expected behaviour: The monitoring system identifies the reviewer's pattern as anomalous (below minimum review time threshold, below minimum override rate threshold) and generates an alert.
• Pass criteria: The anomalous pattern is detected and flagged within one monitoring cycle (maximum 30 days). The alert includes the specific metrics triggering the flag.
• Fail criteria: The anomalous pattern is not detected, or the detection takes more than 30 days.

Conformance Scoring

• Score 0: No contestability mechanism — automated decisions are final with no explanation, no contestation pathway, and no human review.
• Score 1: Reactive contestability — explanations available only on request; contestation handled through general complaints; human review sees and typically confirms the automated decision.
• Score 2: Proactive contestability — explanations provided with each decision; integrated contestation pathway; independent human review with anti-anchoring design; decision records retained.
• Score 3: Verified and effective — all Score 2 capabilities plus: counterfactual explanations, pre-decision contestability, aggregate feedback loop, review quality monitoring with anomaly detection, and independent audit confirming that the contestability mechanism delivers genuine review with meaningful override rates.

9. Regulatory Mapping

Regulation	Provision	Relationship Type
GDPR	Article 22 (Automated Individual Decision-Making, Including Profiling)	Direct requirement
GDPR	Articles 13(2)(f), 14(2)(g) (Right to Meaningful Information About Decision Logic)	Direct requirement
EU AI Act	Article 86 (Right to Explanation of Individual Decision-Making)	Direct requirement
EU AI Act	Article 14 (Human Oversight)	Supports compliance
UK GDPR	Article 22 (as retained)	Direct requirement
Equality Act 2010 (UK)	Sections 13, 19 (Direct and Indirect Discrimination)	Supports compliance
ECHR	Article 6 (Right to a Fair Trial), Article 13 (Right to an Effective Remedy)	Supports compliance
CCPA/CPRA	Section 1798.185(a)(16) (Profiling and Automated Decision-Making Regulations)	Supports compliance
ISO 42001	Clause 6.1 (Actions to Address Risks), Clause 9.1 (Monitoring, Measurement, Analysis, Evaluation)	Supports compliance
NIST AI RMF	GOVERN 1.4, MAP 5.1, MANAGE 4.1	Supports compliance

GDPR — Article 22 (Automated Individual Decision-Making)

Article 22(1) establishes the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects or similarly significantly affects the data subject. Article 22(3) requires that where automated decisions are permitted (by explicit consent, contract necessity, or Union/Member State law), the controller implements "suitable measures to safeguard the data subject's rights and freedoms and legitimate interests, at least the right to obtain human intervention on the part of the controller, to express his or her point of view and to contest the decision." AG-062 directly implements these safeguards. The EDPB Guidelines on Automated Individual Decision-Making and Profiling (WP251rev.01) clarify that human intervention must be by "someone who has the authority and competence to change the decision" — a customer service representative who can only relay the automated decision does not satisfy this requirement.

EU AI Act — Article 86 (Right to Explanation)

Article 86 of the EU AI Act establishes an explicit right to explanation of individual decision-making by high-risk AI systems. The explanation must be "clear and meaningful" and must be provided "in a timely manner." This goes beyond GDPR Article 22 by creating a standalone right to explanation that is not contingent on the decision being "solely" automated. AG-062's requirement for proactive, meaningful explanations directly implements Article 86. The AI Act also requires that the explanation enable the individual to exercise their rights under other Union law — connecting explanation to effective contestability.

Equality Act 2010 (UK) — Sections 13 and 19

Sections 13 (direct discrimination) and 19 (indirect discrimination) apply to automated decisions that produce discriminatory outcomes. An AI agent that uses postcode as a feature in credit scoring may be implementing indirect discrimination if postcode correlates with a protected characteristic and the use of postcode is not a proportionate means of achieving a legitimate aim. AG-062's contestability mechanism serves as a discrimination detection pathway: if contestations reveal that decisions disproportionately affect individuals with a particular protected characteristic, this signals a potential Equality Act violation that must be investigated. The feedback loop from contestation to model improvement is therefore also an equality monitoring mechanism.

CCPA/CPRA — Automated Decision-Making Regulations

The CPRA directs the California Privacy Protection Agency to issue regulations governing profiling and automated decision-making technology, particularly relating to "decisions that produce legal or similarly significant effects." While the final regulations are still in development, the CPRA's framework aligns with GDPR Article 22 in requiring transparency and contestability for consequential automated decisions. AG-062 provides the infrastructure to comply with these forthcoming requirements.

10. Failure Severity

Field	Value
Severity Rating	Critical
Blast Radius	Individual to population-level — each uncontestable decision affects an individual, but systematic failure affects every person subject to the automated decision process

Consequence chain: Without contestability governance, automated decisions become unchallengeable. The immediate consequence is individual harm: a person denied credit, insurance, benefits, or employment based on an automated decision they cannot understand or challenge has suffered a legal effects consequence with no remedy. The systemic consequence is error accumulation: without contestation as a feedback mechanism, model errors, biases, and proxy discrimination persist uncorrected across thousands of decisions. A credit model that uses postcode as a proxy for ethnicity and has no contestability mechanism will discriminate against every applicant from high-minority-population postcodes indefinitely. The regulatory consequence is severe: GDPR Article 22 violations carry upper-tier fines (up to €20 million or 4% of annual global turnover); EU AI Act non-compliance for high-risk systems can result in fines of up to €35 million or 7% of global annual turnover; Equality Act violations can result in unlimited compensation in employment tribunals. The reputational consequence is compounded by the visibility of the affected individuals: denied applicants, rejected claimants, and screened-out candidates are highly motivated to publicise their treatment, particularly through social media and investigative journalism. The legal consequence includes class action exposure: a model that systematically discriminates creates a class of similarly affected individuals, each with a discrimination claim.

Cross-references: AG-049 (Governance Decision Explainability) provides the explanation capability that AG-062 requires for meaningful contestability; AG-019 (Human Escalation & Override Triggers) provides the human review infrastructure that AG-062 requires for contested decisions; AG-013 (Data Sensitivity and Exfiltration Prevention) governs the data inputs to automated decisions; AG-021 (Regulatory Obligation Identification) identifies the jurisdiction-specific requirements for automated decision-making; AG-059 (Lawful Basis and Consent Enforcement) ensures that the data used in automated decisions was collected and processed lawfully; AG-047 (Cross-Jurisdiction Compliance Governance) addresses variations in contestability rights across jurisdictions.