AI ethics in business: how to turn principles into everyday decisions
AI ethics only changes something when it is built into the decisions your teams make every week, and this guide shows how.
BY BETTINA MEYER··8 MIN READ
AI ethics only changes something when it is built into the decisions your teams make every week, and this guide shows how.
TL;DR
Plenty of organisations now have a document that mentions ethical AI. The harder question is who checked the last AI-assisted decision about a customer or an employee, and what that person looked for.
This guide explains what AI ethics means inside a business, the principles that matter most, what South African law already expects, and a seven-step checklist for turning principles into decisions your teams make every week.
AI ethics is the set of principles that guide how AI is designed, bought, and used so that it respects people's rights, avoids harm, and stays under human accountability. For a business, that becomes a set of practical questions: is this system fair to the people it affects, can we explain what it did, whose data is it using, and who answers for the result?
The field has broad international agreement. In November 2021, UNESCO's 193 member states adopted the first global standard on AI ethics, built on ten principles that include fairness, transparency, privacy, and human oversight. The OECD AI Principles, first adopted in 2019 and updated in May 2024, have 47 adherents, and South Africa is among the G20 members committed to the G20 AI Principles drawn from them.
The major frameworks share a common core. The wording varies, but the questions a business needs to answer are much the same.
| Principle | What it means at work | The question to ask |
|---|---|---|
| Fairness and non-discrimination | The system does not treat groups of people worse because of patterns in its training data | Have we compared outcomes across groups on our own data? |
| Transparency and explainability | People can find out that AI was involved and how it reached a result | Could we explain this decision to the person it affects? |
| Privacy and data protection | Personal information goes only where it is allowed to go | What data goes into this tool, and on what terms? |
| Accountability | A named person or team answers for what the system does | Who signs off, and who fixes it when it goes wrong? |
| Human oversight | AI informs consequential decisions about people, and a person makes them | Where exactly does a person review the output? |
| Safety and security | The system behaves reliably and resists misuse | What happens when someone tries to break it? |
| Proportionality | AI is used only as far as the purpose needs | Is AI the right tool here at all? |
Fairness is where failures are easiest to see. Amazon scrapped an early AI recruiting tool after it scored women lower than men, because it had been trained on a pool of CVs drawn largely from men. The tool copied the history it was given. That is why fairness has to be tested rather than assumed, and why AI in recruitment needs particular care.
Writing principles down is the easy part, and most companies have now done some version of it. The share of businesses with no responsible AI policies fell from 24% to 11% in 2025, and AI-specific governance roles grew 17%. Over the same year, documented AI incidents rose to 362, up from 233 in 2024. The most cited obstacles to putting responsible AI into practice were gaps in knowledge (59%), budget (48%), and regulatory uncertainty (41%).
It mirrors a pattern we see with AI tools generally: something new arrives, and the work carries on exactly as before. Think of quality control on a production line. A quality manual in the office does not stop defects. Checks built into each station do, because they happen where the work happens and someone owns each one. AI ethics works the same way. It has to live inside the workflows where AI touches people, with an owner at each point.
In South Africa, part of AI ethics is already law. Section 71 of POPIA says a person may not be subject to a decision with legal consequences, or one that affects them to a substantial degree, based solely on automated processing that profiles them, meaning software builds a picture of the person from their data, including their performance at work, creditworthiness, reliability, health, or conduct. There are exceptions, such as decisions linked to a contract, but they come with conditions: the person must be able to make representations, and must be given enough information about the underlying logic to do so.
Read as an ethics brief, section 71 asks for human oversight, explainability, and a route for the affected person to respond, all at once. Hiring, performance management, credit, and claims decisions are where it applies first. National AI policy is a separate question, so check where South Africa's AI policy stands before relying on any draft.
More often than vendors suggest. Proportionality, one of UNESCO's ten principles, says the use of AI should not go beyond what is necessary to achieve a legitimate aim. If a decision is rare, high-stakes, and hard to explain, a well-designed manual process is, in my view, the more ethical choice.
We hold one further line. AI-native work, as we see it, means a more capable team, with repetitive and administrative work moving off people's plates. We have never run an engagement where the goal was headcount reduction, and in my view the way a leadership team answers that question shapes whether its people trust the tools at all.
AI ethics is the set of principles that guide how AI is designed and used so that it respects people's rights, avoids harm, and stays under human accountability. In a business, it comes down to fairness, explainability, privacy, and a named person answering for each AI-assisted decision.
The common core is fairness, transparency and explainability, privacy, accountability, human oversight, safety and security, and proportionality. UNESCO's Recommendation sets out ten principles and the OECD lists five values-based ones, and the two overlap heavily.
Staff know which tools are approved and what data may go into them, a person checks AI output before it reaches a client or informs a decision, and anyone affected by an AI-assisted decision can ask how it was reached. The checklist above turns that into seven steps.
Yes. POPIA governs personal information wherever it is processed, including by AI tools, and section 71 restricts decisions with legal or substantial effects that are based solely on automated profiling. This is not legal advice, so take advice on specific systems.
One accountable executive should own it, with each AI use case owned by the business leader whose decisions it affects. Legal, IT, and HR all contribute, but ownership that sits only with a technical team tends to miss the effects on people.
If you want a scored view of where your organisation stands before you start, the free AI Readiness Diagnostic produces a report you can take to your exec team.
Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.
Take the AI Readiness Diagnostic