AI Assurance in the Public Sector: The Evidence a Government AI System Has to Produce

AdwumaTech defines AI assurance as the continuous production, by design, of the evidence that an AI system behaved as the state said it would.

In the public sector AdwumaTech structures that evidence as six artifacts, the Evidence Chain: an impact assessment made before deployment, a data lineage record, a disaggregated evaluation record, a decision log, an oversight record, and a change record. A system that cannot produce all six on demand is not assured. It is trusted, which is a different condition and a weaker one.

The six are not a proprietary invention. Each corresponds to an obligation already written into the EU AI Act, a control objective in ISO/IEC 42001, and a function of the NIST AI Risk Management Framework. What the Evidence Chain does is state them as artifacts a system emits at runtime instead of documents an organization produces at audit time.

Assurance in government is not tested by a procurement questionnaire. It is tested when a citizen is denied a benefit, a payment reaches the wrong beneficiary, or a court asks the state to explain a decision a model made eighteen months earlier. At that moment the state either holds the record or it does not.

What is the accountability lag?

The accountability lag is the interval between the moment an AI system begins making decisions and the moment the state becomes able to prove how it made them. AdwumaTech uses the term to describe the structural gap that assurance work exists to close.

In most deployments the lag is created at the same point: assurance is scoped as a documentation exercise scheduled after go-live. The system ships. The audit trail is planned for the next phase. Then the first contested decision arrives and the evidence the state needs was never written, because the system was not built to write it.

The lag cannot be closed retroactively. A decision log that begins in month nine has nothing to say about month three. A disaggregated evaluation run after a fairness complaint cannot establish what performance was at the time of the decision under challenge. Evidence is either generated at the moment of the decision or it does not exist.

This is why AdwumaTech treats assurance as an architectural requirement and not a compliance deliverable. It is designed in at the same stage as the database schema.

Where AI regulation stands in August 2026

Public sector AI programmes across Africa, the Caribbean, and Europe are being designed against a regulatory picture that changed in July 2026. Teams working from guidance published before that point are working from the wrong calendar.

The EU AI Act entered into force on 1 August 2024 with staggered obligations. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.

The deferral covers conformity obligations. It does not cover transparency obligations, and it does not reduce the underlying work. The stated reason for the deferral was that harmonised technical standards were not available in time for providers to demonstrate conformity, which is a statement about the maturity of the assurance toolchain and not about the necessity of assurance.

Two implications matter for governments outside the EU. First, the Act reaches providers whose AI outputs are used inside the Union. Second, the Annex III risk categories are becoming the reference vocabulary national AI policies borrow from. A public sector system built to produce Annex III grade evidence is portable across regimes. One built to satisfy a single national checklist is portable nowhere.

The Evidence Chain, crosswalked

1. The impact assessment, completed before deployment

Article 27 of the EU AI Act requires deployers that are bodies governed by public law, and private entities providing public services, to carry out a fundamental rights impact assessment before putting a high-risk system into use. ISO/IEC 42005:2025 is the guidance standard for AI system impact assessment.

AdwumaTech runs this assessment before design, against the sovereignty and governance constraints the state names, because it sets the depth of everything below it.

2. The data lineage record

Article 10 requires training, validation, and testing data sets to be relevant, sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose.

Satisfying that requires a record, not an assertion. The record states where the data came from, the legal basis on which it was collected, who annotated it and to what written specification, the inter-annotator agreement achieved, and the demographic and linguistic composition of the corpus.

For national systems this record answers the question that arrives first in any fairness challenge: was this system evaluated on people like the complainant.

3. The disaggregated evaluation record

Aggregate accuracy conceals the failure that matters. A verification system reporting 98 percent accuracy nationally can be failing at 12 percent for a specific age band, skin tone, document type, or regional accent, and the aggregate will not show it.

For biometric systems the metrics are standardised. Recognition performance is expressed as false match rate and false non-match rate, with failure to enrol and failure to acquire rates carried alongside them. Demographic variation is quantified under ISO/IEC 19795-10:2024.

This is the discipline behind NOKORE AI, AdwumaTech's identity integrity system, where liveness and presentation attack detection aligned to ISO 30107 is calibrated against the demographics and fraud patterns of the populations it defends.

4. The decision log

Article 12 requires high-risk systems to record events automatically over their lifetime, with logging capabilities appropriate to the intended purpose.

AdwumaTech implements decision logging as a schema: every automated decision, with its inputs, the model version that produced it, the confidence attached to it, the threshold in force at the time, the reference data consulted, and whether a human intervened.

The objection raised against decision-granular logging is cost, and the objection does not survive arithmetic. A national system running 20 million authentications a month at 2 kilobytes per structured log record produces roughly 40 gigabytes a month and under half a terabyte a year.

5. The oversight record

Article 14 requires high-risk systems to be designed so that natural persons can effectively oversee them. Article 26(2) requires deployers to assign oversight to persons with the necessary competence, training, and authority.

Human oversight is a claim until it is evidenced. The record shows which decisions were reviewed, by whom, under what authority to override, how often overrides occurred, and what happened to the cases that were overridden.

6. The change record

Article 72 requires providers of high-risk systems to establish and document a post-market monitoring system. Article 73 requires reporting of serious incidents. Article 43(4) treats substantial modification as triggering a fresh conformity assessment.

The record therefore covers every model version, every retrain, every threshold adjustment, and the re-evaluation that followed each one, timestamped.

Why the oversight record carries more weight than it appears

Human oversight is the control that every AI governance framework leans on, and it is the control with the weakest evidence base behind it.

Automation bias is the tendency to over-rely on automated output as a substitute for vigilant independent judgement. The research on how humans behave when supervising automated advice shows that erroneous decision support can materially increase the risk of an incorrect decision, and that bias can intensify as system reliability rises.

This inverts how the oversight record should be read. An oversight function reporting a zero percent override rate across a year is not evidence of a system performing perfectly. It is evidence that must be investigated for automation bias.

The record has to support four questions:

1. What is the override rate, and how has it moved over time relative to measured system performance?
2. What is the review dwell time, and has it fallen?
3. When the system was wrong, did the reviewer catch it?
4. What happened to the overridden cases?

AdwumaTech uses seeded error injection where appropriate: a low, known rate of deliberately incorrect outputs introduced into the review queue, with the reviewer catch rate measured against them.

The arithmetic of exclusion

Assurance arguments become concrete at national scale.

Take a national identity system with 20 million enrolled residents, supporting one authentication per person per month for a government-to-person payment. At a failure to enrol rate of 1.2 percent, 240,000 people cannot be enrolled at all. They do not appear in any downstream accuracy metric, because every downstream metric is computed over the register they are absent from.

At an aggregate false non-match rate of 1.0 percent across 20 million monthly authentications, the system produces 200,000 failed authentications a month. Suppose one demographic group comprising 25 percent of the population records an FNMR of 1.9 percent, while the remaining 75 percent record 0.7 percent. The aggregate is unchanged at 1.0 percent, but the smaller group accounts for 95,000 of the 200,000 monthly failures.

Biometric failures are not independently distributed across the population. Some users fail repeatedly instead of randomly. If the exception path costs a citizen 40 minutes including travel and waiting, 95,000 monthly failures represent roughly 63,000 hours of citizen time a month, borne disproportionately by one group, on a system reporting 99 percent success.

None of this requires a new measurement technology. It requires the evaluation record to be disaggregated, failure to enrol to be carried as a headline metric, and repeat failures to be tracked at the individual level.

Scoring the Evidence Chain

Governments rarely lack assurance entirely. What they usually have is assurance that exists on paper and cannot be produced under time pressure.

Score each of the six artifacts from 0 to 4. The total is the Evidence Chain Score, on a scale of 0 to 24. Below 12, the system is not defensible. Between 12 and 17, it is defensible to an internal auditor and not to a court or regulator. At 18 and above, with no single artifact below 2, the state can answer for the system.

The critical reading is not the total. It is level 2, where most public sector AI programmes sit and where the accountability lag lives. The move from level 2 to level 3 is an architecture change.

The assurance drill

A score is a claim about the Evidence Chain. AdwumaTech validates it the way disaster recovery is validated: by attempting the thing under timed conditions before it is needed.

Pick a decision at random from a defined period in the past, typically 6 to 18 months. Reconstruct it end to end within a fixed time budget: the input the system received, the model version, threshold, confidence, reference data, human review, reviewer authority, measured performance for the relevant demographic group, and what the impact assessment said about the decision class before deployment.

The drill ends at the first point where the chain breaks. That point is the finding.

Supporting practices include log completeness sampling, appeals path testing, retention verification, and adversarial evaluation.

What disqualifies a public sector AI deployment

AdwumaTech treats four conditions as disqualifying regardless of measured performance.

There is no appeals path a person affected by the decision can actually reach.

The decision class cannot produce an Evidence Chain.

The vendor restricts evaluation or will not describe the training data.

No named official holds authority to halt the system.

How ISO/IEC 42001 structures the work

ISO/IEC 42001:2023 is the international standard for AI management systems. It combines management system requirements in clauses 4 through 10 with 38 controls in Annex A organised under nine control objectives numbered A.2 to A.10.

Annex A is a catalogue drawn from through a Statement of Applicability, justified by the organization's risk and impact assessments. The standard does not sit alone. ISO/IEC 23894:2023 provides AI risk management guidance, ISO/IEC 42005:2025 provides impact assessment guidance, ISO/IEC 38507 addresses governance implications for governing bodies, and ISO/IEC 42006:2025 sets competence requirements for bodies that audit and certify AI management systems.

AdwumaTech is ISO 27001 certified for information security management across all engagements. AdwumaTech government deployments are additionally aligned to ISO/IEC 42001 for AI management systems, and data handling is GDPR aligned.

Assurance inside a national system

The four capabilities a digital state runs on each generate a different assurance obligation.

Digital identity carries the deduplication and false match question. Payments infrastructure carries the targeting question. Citizen services carries the grounding and language question. AI assurance is the AdwumaTech function that holds the other three to account: risk assessment, model evaluation, bias auditing, human oversight, and evidence prepared for audit aligned to ISO/IEC 42001.

Sovereignty is an assurance property

A state cannot assure a system it does not control.

Assurance requires access to the model, the weights, the evaluation harness, the logs, and the data at the granularity the Evidence Chain demands. A system operated on infrastructure the state cannot inspect, under a licence restricting evaluation, or with training data the vendor will not describe, cannot reach level 3 on any artifact that depends on those things.

Sovereignty and assurance are therefore the same requirement approached from two directions. Citizen data stays within the country or a jurisdiction the state names. The system, the data, and the cryptographic keys belong to the government. AdwumaTech delivers from inside the regions it serves and trains and certifies the teams who operate and extend the system for the long term.

Related reading: Which AI Use Case Should You Build First sets out how AdwumaTech scores oversight cost and failure tolerance before a system is commissioned. The ExploitGym Incident covers containment architecture for agentic systems.

Bring the mandate. Keep the system. Talk to us.

References

• Regulation (EU) 2024/1689, the Artificial Intelligence Act.
• ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system.
• ISO/IEC 42005:2025, Information technology, Artificial intelligence, AI system impact assessment.
• ISO/IEC 23894:2023, Information technology, Artificial intelligence, Guidance on risk management.
• ISO/IEC 19795-1, Biometric performance testing and reporting, Principles and framework.
• ISO/IEC 19795-10:2024, Quantifying biometric system performance variation across demographic groups.
• ISO/IEC 30107-3, Biometric presentation attack detection, Testing and reporting.
• ISO/IEC 42006:2025, Requirements for bodies providing audit and certification of AI management systems.
• NIST, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)," NIST AI 100-1, January 2023.
• K. Goddard, A. Roudsari, J. C. Wyatt, "Automation bias: a systematic review of frequency, effect mediators, and mitigators," Journal of the American Medical Informatics Association, 2012.
• R. Parasuraman, D. H. Manzey, "Complacency and Bias in Human Use of Automation," Human Factors, 2010.
• T. Gebru et al., "Datasheets for Datasets," Communications of the ACM, 2021.
• M. Mitchell et al., "Model Cards for Model Reporting," ACM FAT*, 2019.

About AdwumaTech

AdwumaTech AI engineers applied AI systems for enterprises and sovereign digital infrastructure for governments, building the full chain from data acquisition through post-training to the deployed system. The government practice covers digital identity, payments infrastructure, citizen services, and AI assurance, deployed in country and under state control. The enterprise practice covers AI consulting, productized AI, data operations, and NOKORE AI identity integrity. AdwumaTech is ISO 27001 certified and aligned to ISO/IEC 42001, operates from Accra, Ghana, and serves governments across African and Caribbean markets. The state retains total ownership of the system, the data, and the cryptographic keys.