Consider an illustrative case: a health ministry uses an AI tool to compare bids for medical supplies. The tool extracts specifications, produces a ranked list and recommends a supplier. Under deadline pressure, the procurement officer accepts the recommendation. A losing bidder later challenges the evaluation, but the team cannot reconstruct which documents, criteria and system version produced the ranking.
This is a hypothetical example, not a documented ministry deployment. It exposes a practical question for any institution considering AI-assisted procurement: what evidence would remain if a decision were challenged? A plausible explanation generated afterwards would not establish what actually happened.
AI-assisted procurement must preserve the institution’s ability to explain, review and defend a decision. Faster analysis can be useful, but the organisation making an award retains responsibility for the process. The operating model needs to connect a recommendation to authorised criteria, source evidence, human judgement and a route for correction.
Efficiency is useful only when the evidence survives
Procurement teams can face large document volumes, changing specifications and pressure to evaluate bids promptly. Where records are fragmented or partly paper-based, finding the relevant evidence may take longer than applying the evaluation criteria. The problem differs between institutions: a national purchasing agency with a mature digital platform has a different starting point from a district office reconciling emailed documents and scanned forms.
AI can assist with extracting fields, comparing submitted specifications, identifying missing documents and flagging cases for review. These are candidate uses, not evidence that a particular product can perform them reliably. Procurement leaders should test the task against their own documents, languages and rules before relying on the output.
The distinction matters when moving from document support to supplier scoring. Extracting a delivery date can be checked against a source passage. Recommending an award combines evidence, eligibility rules, technical judgement and financial comparisons. A system may be good at the former without being suitable for the latter. Deterministic calculations and explicit rules should remain deterministic where that makes the result easier to verify.
Accountability cannot be delegated to a ranking
John Karani, writing in The Kenyan Wallstreet, argues that AI changes the skills procurement professionals need and raises questions about delegated authority, accountability and data integrity. His article is professional commentary, not an empirical study of African public procurement deployments. It provides a useful leadership question; it does not establish a continent-wide adoption rate or prove efficiency gains.
An institution should answer that question before choosing a system. Who approves the criteria? Who verifies extracted evidence? Who can override a recommendation? Who investigates a complaint? Those responsibilities cannot be resolved by a vendor’s model-accuracy score. They belong in the procurement workflow and in the contract governing the technology.
A human signature alone is insufficient. A reviewer needs access to the original evidence, enough time to examine it and authority to reject or pause the recommendation. If staff are rewarded only for processing speed, the review step can become a rubber stamp. Monitoring should therefore include correction effort, overrides and unresolved exceptions alongside turnaround time.
An explanation is not the same as an audit trail
A generative system can produce a persuasive account of why it favoured a supplier. That account may not faithfully describe the process that generated the ranking. The stronger requirement is a traceable decision record: the submitted documents, applicable criteria, extracted evidence, scoring method, system configuration and the reviewer’s recorded judgement.
The UK government’s Algorithmic Transparency Recording Standard guidance offers a comparative example. Its template covers ownership, supplier involvement, deployment context, human decisions, risks and review routes. Its mandatory scope is specific to the UK public-sector policy; it is not an obligation imposed on African institutions. The useful design lesson is that transparency concerns the whole operational process, rather than the model alone.
Public transparency also needs boundaries. Bid documents can contain commercial information and personal data. An internal audit record need not become an unrestricted public dataset. Institutions should distinguish what reviewers and auditors need to access from what can be disclosed publicly, applying the relevant procurement, confidentiality and data-protection requirements in their jurisdiction.
Governance has to be built into the workflow
The NIST AI RMF Playbook provides voluntary suggested actions across Govern, Map, Measure and Manage. It is neither a complete checklist nor a procurement-compliance certificate. Its value here is the discipline of connecting risk measurement and management to the system’s intended use.
Discover. Map one decision and its evidence chain. Identify the officers, suppliers, auditors and affected service teams who depend on the outcome. Establish a baseline for time, errors and review effort, and ask whether a simpler search, rules or document-management improvement would solve the problem.
Design. Set the permitted role of AI, the criteria it may apply and the point at which a person must decide. Specify access controls, retention, supplier data-use restrictions and complaint handling. Require an exportable audit record and a fallback workflow; control should not disappear when a vendor contract ends.
Build and test. Evaluate representative documents, difficult scans, ambiguous specifications and incomplete submissions. Historical awards are not automatically correct labels: reviewers should assess whether the underlying decisions and criteria were valid. Test permission boundaries and malicious instructions embedded in documents as well as extraction accuracy.
Enable and monitor. Train officers to verify evidence and recognise limits. Record overrides and corrections, investigate recurring errors and reassess material changes to models, prompts or scoring rules. Include the staffing, support and storage costs of those controls in the business case, not only the software licence.
National strategy does not settle a tender decision
Kenya’s Ministry of ICT announced the launch of its AI Strategy 2025–2030 in March 2025, identifying infrastructure, data and AI governance, and research, innovation and commercialisation as its three pillars. This is evidence of national strategic direction. It is not, by itself, evidence that an individual procuring entity has approval to delegate supplier selection to an AI system.
An implementation in Kenya, Uganda or another market needs to be assessed against that jurisdiction’s actual procurement process, review mechanisms and data-protection obligations. There is no single African permission model. Institutional capability also varies within a country, including between central agencies and local administrations.
Operating conditions should change the design. Intermittent connectivity requires a way to continue reviewing documents without silently losing evidence. Mixed-language submissions require testing in the languages actually used. Small suppliers should not be disadvantaged because their documents differ from the examples on which a tool was evaluated. These are conditions to examine, not assumptions that every institution faces the same limitations.
Begin with support before granting authority
A defensible first use is a bounded document-support task: extract specified information from a bid and link every extracted item to its source location. Officers retain the evaluation decision. An unavailable service, unreadable scan or conflicting record should lead to an explicit exception rather than a confident guess.
Compare the tool with the existing process on a representative, governed test set. Measure extraction errors, missed exceptions, review time and total cost per completed case. Agree unacceptable failure conditions before the trial. A faster average result should not outweigh a recurring error that excludes a supplier or conceals missing evidence.
Expansion should follow evidence. If the system cannot preserve source links or staff cannot correct its output, adding automated rankings increases the organisation’s exposure. A successful document-support pilot may justify that narrow capability without justifying autonomous award recommendations.
The institution must still be able to defend the award
The leadership test is whether the organisation can reconstruct and review a consequential decision months after it was made. That requires records, ownership and intervention routes that survive staff turnover, supplier changes and appeals. Explainability and efficiency are complementary requirements; neither replaces lawful process or reliable evidence.
Xelius supports public-sector teams through AI evaluation, governance design and workflow integration. A useful starting point is one procurement task where the team can define the decision boundary, establish a baseline and test whether assistance improves the process without weakening accountability.
