When the Algorithm Recommends, Who Decides?

AI is increasingly influencing recruitment, performance assessment, workforce planning and other employment decisions. For Ghanaian employers, this can sometimes seem like a future problem: something being worked out by regulators in Europe while businesses here experiment with new technology.

It is not.

Ghana's existing data-protection law already recognises that individuals should not simply be subjected to significant decisions based solely on automated processing. At the same time, the Data Protection Commission has signalled a more enforcement-focused approach, while government is developing legislation that would address artificial intelligence, automated decision-making and emerging technologies more directly.

The important question is therefore no longer what AI can do. It is this: when an algorithm reaches a conclusion about an individual, who actually makes the decision?

The direction of travel

The regulatory direction is becoming clearer internationally.

Italy provides a particularly useful example. On 23 July 2026, Italy's Council of Ministers gave preliminary approval to a draft legislative decree implementing the EU Platform Work Directive. The proposed rules place the individual at the centre of algorithmic management: the algorithm would not be able to dismiss a platform worker, while workers would have to be informed about the parameters affecting tasks, working time and remuneration. The proposal remained subject to further parliamentary and social-partner processes.

It would be easy for conventional employers to dismiss this as regulation of the platform economy. That would miss its wider significance.

Platforms simply provide regulators with a particularly visible example of algorithmic management. Similar questions arise whenever software ranks job candidates, identifies an employee as an attrition risk, evaluates productivity, flags poor performance or recommends someone for promotion, investigation or redundancy.

The issue is not confined to Europe.

On 21 August 2026, the Dutch Data Protection Authority imposed a fine of €824.99 million on Uber after finding that drivers' accounts had been blocked through automated decision-making. The blocking directly affected drivers' ability to earn income through the platform. Uber has appealed the decision.

The case is significant not simply because of the size of the fine. It illustrates the potential consequences when an organisation allows an automated system to move from making a recommendation to effectively making a consequential decision about a person's livelihood.

And Uber operates in Ghana.

Ghana is not starting from zero

Ghana already has a legal foundation for addressing this problem.

The Data Protection Act, 2012 (Act 843), contains specific provisions dealing with automated decision-making. Section 41 addresses decisions that significantly affect an individual and are based solely on automated processing, and provides a mechanism through which an affected individual can require reconsideration. The provision also contains exceptions, including in relation to certain decisions connected with contracts or authorised by law.

In an employment context, therefore, the provision can become relevant to consequential decisions involving recruitment, performance, discipline, promotion or dismissal, although its application will depend on the circumstances and the statutory exceptions.

The regulatory environment is also changing.

On 31 January 2026, the Data Protection Commission declared that 2026 would be a "year of enforcement", indicating a tougher approach to compliance with Ghana's data-protection requirements.

Then, in March 2026, the Government announced that it was developing a new Data Protection Bill to address, among other issues, artificial intelligence, automated decision-making and cross-border data transfers. The Government also said that an Emerging Technologies Bill was being developed to provide structured oversight of AI systems, advanced analytics, digital assets and new digital platforms.

Ghana's National Artificial Intelligence Strategy (2025-2035), officially launched on 24 April 2026, adds a broader policy framework. The strategy aims to build a responsible, human-centred AI ecosystem and to develop the legal and ethical frameworks needed to govern AI. It is a policy strategy, not itself an enforceable employment law.

None of this yet amounts to a Ghanaian regulatory regime equivalent to the European framework.

But that is not the point.

The direction of travel is what employers should be watching: an increasingly active regulator, legislation being developed specifically with AI and automated decision-making in mind, and growing recognition that automated systems can have real consequences for individuals.

Employers that wait for a major Ghanaian enforcement action before taking algorithmic accountability seriously risk confusing today's level of enforcement with tomorrow's level of regulatory expectation.

The European developments are therefore not a preview of a debate Ghana will eventually have. The legal foundation for that debate already exists here.

What is missing, for many employers, is the operational discipline to act as though it does.

A human in the loop may not be enough

The important distinction for HR is not simply between an automated decision and a human decision. It is between real human judgement and the appearance of human judgement.

Suppose an AI system identifies an employee as a serious performance concern. A manager receives a dashboard containing a score, a recommendation and an explanation. The manager knows that the system has processed vastly more information than any individual could examine and accepts its recommendation.

Technically, a human made the final decision. But did the human exercise judgement?

The danger is that apparent human oversight becomes little more than administrative confirmation. The more authoritative an algorithm appears, the easier it becomes for managers to defer to it, particularly where challenging the recommendation requires additional explanation, approval or personal responsibility.

Meaningful human involvement must therefore mean more than inserting a person at the end of an automated process.

A genuine decision-maker must be able to understand the basis of the recommendation, question its assumptions, introduce circumstances the system may not know about and reach a different conclusion without treating disagreement with the technology as evidence of poor judgement.

There is a useful parallel outside HR.

In August 2026, the Solicitors Regulation Authority in England and Wales published a warning notice on the responsible use of AI. It emphasised that using AI does not change the professional standards expected of solicitors and that individuals remain responsible for the work and advice they produce, regardless of whether AI is used.

The same principle applies to employment decisions.

An HR professional or manager cannot discharge responsibility merely by pointing to a dashboard and saying: "The system recommended it".

Technology can supply evidence. It cannot accept responsibility for what an organisation decides to do with that evidence.

Four tests for AI-assisted HR decisions

A general policy requiring a "human in the loop" is therefore inadequate. HR needs to determine what that human is actually required and empowered to do.

Four questions provide a practical starting point.

How consequential is the decision?

There is an obvious difference between software recommending a training course and software contributing to a decision to dismiss someone. As the consequences increase, so should independent human scrutiny.

Can the recommendation be understood and challenged?

An impressive accuracy rate across thousands of employees does not explain why a particular conclusion is right for one individual. The decision-maker must understand enough about the basis of the recommendation to question it intelligently.

Can the human genuinely override it?

An override mechanism is meaningless if managers are discouraged from using it, must obtain exceptional approval when they do, or are subsequently judged against the algorithm's recommendations. The ability to disagree must exist in practice, not merely in the system specification.

Who owns the decision?

There should be an identifiable person able to explain why the decision was made and prepared to accept responsibility for it. "The system recommended it" cannot become an organisational defence.

These tests do not imply retreating from AI.

AI can reveal relationships, inconsistencies and patterns that human decision-makers might otherwise miss. Used properly, it can substantially improve the evidence available to HR and help organisations make faster and potentially better-informed decisions.

But better evidence and better judgement are not the same thing.

Judgement is not a defect

There is a tendency to regard human judgement as the unreliable component of decision-making that better technology will progressively eliminate.

Human decisions can certainly be inconsistent, biased and poorly informed. But many employment decisions contain matters that cannot be resolved simply by adding more data.

Context matters. Proportionality matters. An employee's explanation matters. So do changing circumstances, previous behaviour, management failures and information that may never have entered the system.

Human judgement is therefore not merely an imperfect safeguard that should remain until AI becomes sufficiently sophisticated to replace it.

In consequential employment decisions, judgement is part of the process through which evidence acquires meaning.

This distinction will become increasingly important as multinational employers manage workforces that cross borders. An employee may be recruited in Ghana, managed from another country, assessed by software developed elsewhere and subject to an employment decision made using data processed across several jurisdictions.

The technology may be global.

The consequences for the individual are local.

That makes accountability harder, not less important.

For HR leaders in Ghana, the Data Protection Commission's enforcement posture may still be developing, but the principles in Act 843 are not hypothetical. They form part of the existing legal framework.

Employers should therefore begin asking not merely whether they have a "human in the loop", but what that human actually does. Can they question the system? Can they depart from its recommendation? Do they have sufficient information to make an independent assessment? And, ultimately, are they prepared to own the decision?

The emerging regulatory direction should prompt employers to examine their AI governance now, rather than after a disputed decision exposes its weaknesses.

AI can recommend, but someone still has to decide.

Author: Zhu Li

Zhu Li is a Director at the Federation of International Employers (FedEE), where she works closely with multinational employers on international HR, workforce and employment issues. Her work spans cross-border HR policy, labour-market developments and emerging workforce risks, with a particular interest in how organisations adapt job design and people practices to changing labour-market conditions. She has contributed analysis to HR publications on subjects including poly-working, global HR compliance and the challenges of managing one workforce across multiple legal jurisdictions.