To use AI in HR responsibly, decide what the technology should make better before you build, buy or switch anything on. Then judge every tool against those standards, with evidence. At Peopletree Group we use seven principles to do this: strengthen human trust, challenge bias, elevate talent value, apply contextual intelligence, promote cultural integrity, be transparent and protect data. Each principle comes with a test you can run on your own implementation, and the scoring framework at the end of this article shows you how.
Why set principles before you use AI in HR?
We're seeing an extraordinary number of AI tools being introduced into HR. Tools that can screen thousands of CVs, write performance reviews, recommend development plans, identify successors and predict which employees might leave. And most of them promise to make things faster, easier and cheaper.
But I've been wondering whether we're starting in the wrong place. Whenever we introduce a new technology, surely the first question should be: What principles are we going to use to determine whether this technology is making things better?
Not after we've built it. Not once we've spent the money and convinced everybody to use it. Before we start.
Because if we haven't defined what good looks like, how will we know whether what we've built and deployed is doing the job it should be doing?
What should you check before you use AI in HR?
The question I hear most from HR leaders at the moment is whether AI can predict who's going to resign. It can help, within limits. Predictions about turnover, promotion readiness and succession risk depend on the data underneath them, and in most mid-sized companies that data is patchy and spread across systems that were never designed to talk to each other.
Before you rely on AI-generated people insights, check that these foundations are in place:
- Structured role profiles that describe what success looks like in each role, including the technical and behavioural requirements.
- A standardised skills framework, so the same skill has the same name in every department.
- Performance and behavioural evidence from managers, peers and direct reports, so one person's view can't skew the picture.
If any of these are missing, AI will fill the gaps with assumptions. And if your historical data favours one group or working style, AI will repeat that pattern at scale (principle 2 covers this in detail). Our article on DataHub explains how we bring this data together.
Starting with first principles
At Peopletree Group, we've spent a considerable amount of time thinking about this while developing TAILA, our AI talent assistant. We started by agreeing how we would judge it. The result was a set of seven principles that we wanted to use to guide how we designed and deployed it.
I think these principles are relevant to anybody building, buying or implementing AI within an HR function. And importantly, they're practical. They should help us make decisions about what AI should do, what it shouldn't do, and how we evaluate whether it's working.
Here are the seven principles, with examples of what happens when we get them right, and when we don't.
1. Strengthen human trust: how can AI improve the quality of interactions?
Does the technology improve the quality of human relationships and interactions, helping people trust each other's intentions and decisions more?
Think about a career conversation between a manager and an employee. We have an enormous amount of information that could help make that conversation more useful: the employee's experience, skills, behavioural strengths, aspirations, performance history and potential career opportunities.
But having access to that information doesn't automatically mean the manager will have a better conversation. In fact, an AI tool that tells the employee which role they should pursue, or which competencies they need to improve, could undermine the conversation altogether. The employee might disagree with the recommendation. The manager might not understand how it was reached. And neither feels particularly involved in the decision.
We've approached this differently with TAILA. The intention is to use the available information to help managers ask better questions. What does this person enjoy doing? Where are their strengths most valuable? What might they find difficult in a future role? What experiences could help them develop? What does the employee want?
AI should help the manager explore the right issues and provide more informed guidance. The conversation itself still belongs to the manager and the employee.
The test: Are managers and employees having better conversations because of the technology?
2. Challenge bias: how do you stop AI repeating bias?
Does the technology actively identify, minimise and challenge the biases that influence people decisions?
Recruitment is a particularly useful example because, in many respects, it's the assessment of external talent. We know that recruitment decisions are vulnerable to bias. Similarity bias, confirmation bias, halo effects and assumptions about education, experience or career history can all influence who gets shortlisted.
AI creates an opportunity to improve this. But it also creates the possibility of making existing problems worse, at a much greater scale.
Imagine a recruitment tool trained to recognise patterns in the CVs of previously successful candidates. It might start favouring particular career paths, universities, job titles or ways of describing experience. It could also favour applicants who know how to write CVs that appeal to AI screening systems. The result might be a shortlist that looks very convincing, but excludes people who could perform exceptionally well in the role.
The better approach is to start with a clear understanding of what success in the role requires. What skills, knowledge, behaviours and capabilities matter? Which criteria are essential? Which are merely preferences? And how do we test whether the screening process is excluding suitable candidates?
The same principle applies to internal talent assessment, performance reviews and talent identification. AI should help managers recognise where their judgements might be influenced by bias, introduce relevant evidence and challenge conclusions that aren't adequately supported.
The test: Are we improving the fairness and accuracy of our talent decisions, or simply processing existing biases more efficiently?
3. Elevate talent value: how can AI help you identify and grow the unique value each person brings?
Does the technology help us recognise and grow the unique value each person brings to the business?
One of the problems I've always had with traditional competency-based development is the assumption that everybody should eventually look like the ideal person described in a role profile. We assess somebody against a list of competencies, identify their gaps and then create a development plan to close those gaps.
But people aren't interchangeable. Some are exceptional at building relationships. Others are brilliant at solving complex problems, creating new ideas, organising work or driving execution. And the value of those strengths depends enormously on the context in which they're being applied.
Consider somebody who is highly creative, comfortable with uncertainty and particularly good at challenging conventional thinking. Put that person in a role that requires innovation and they'll potentially create enormous value. Put them in a tightly controlled environment where every decision requires strict procedural compliance, and those same characteristics might become a source of frustration.
A conventional AI development tool might identify their lack of procedural discipline and recommend training to correct it. But is that really the best way to maximise their value? Perhaps the better question is whether they're in the right role, whether their responsibilities could be adjusted, or whether somebody with complementary strengths could support them.
This is how we think about talent assessment and development. AI should help us understand the relationship between the person and the context in which they work. Sometimes the answer is to change the role.
The test: Are we helping people contribute more of their particular strengths?
4. Apply contextual intelligence: how do you make AI recommendations relevant?
Does the technology use relevant research, organisational information and business context to provide recommendations that make sense?
This is one of the areas where I think AI has enormous potential, but also where it's easy to get things wrong. Ask a general-purpose AI assistant to create a leadership development plan and you'll probably get something that looks quite reasonable. It might recommend improving communication, strategic thinking, delegation, coaching and emotional intelligence. All perfectly sensible.
But for whom?
A first-time manager has different development needs from somebody moving from managing managers to managing an entire business function. The work changes. The decisions change. The relationships change. The time horizon changes. And the behaviours that made somebody successful in their previous role might not be the behaviours they need to rely on in the next one.
There are well-established frameworks, such as the Leadership Pipeline, that help us understand these transitions. So if AI is going to recommend development priorities, shouldn't it understand which transition the person is making? Shouldn't it also understand the organisation's strategy, the requirements of the role, the individual's strengths and the resources available to support development?
This is where the quality of the underlying frameworks and data becomes so important. AI is very good at producing plausible recommendations. Whether they're appropriate is a separate question. The better approach is to combine research-based models with information about the individual, the role and the organisation, so that recommendations are relevant to the situation.
The test: Can we explain which frameworks, data and assumptions the AI is using, and why its recommendation is appropriate for this particular person in this particular role?
5. Promote cultural integrity: how can AI strengthen your organisation's culture?
Does the technology make it easier for people to act in ways that strengthen the organisation's culture, and harder to act in ways that damage it?
Most organisations have a reasonably clear idea of the culture they want. They talk about accountability, collaboration, trust, respect and developing people. But culture is what people experience through the decisions and behaviours of their managers. The values on the wall only count when managers act on them.
Take performance management. A manager might avoid addressing poor performance because they don't enjoy difficult conversations. An AI tool could make this worse by allowing them to generate a carefully worded review that says very little and avoids the underlying issue. The process gets completed, but nothing changes.
Alternatively, AI could help the manager prepare for a constructive conversation. What evidence supports the concern? What questions should be asked? How can the manager be clear about expectations while remaining empathetic? What support might the employee need?
The responsibility stays with the manager. The technology helps them exercise it more effectively. The same applies to succession decisions, development opportunities and retention conversations. If we say we value fairness, accountability and development, the technology should make those behaviours easier to practise consistently.
The test: Does the AI reinforce the behaviours we want managers to demonstrate?
6. Be transparent: how should AI explain its recommendations?
Does the technology explain its reasoning so that people understand why a particular recommendation is being made?
We've all experienced the frustration of being told what to do without understanding why. And when it comes to decisions about people's careers, the consequences can be significant.
Imagine an AI succession tool that identifies somebody as a potential successor for a critical role and concludes that they'll be ready in three years. Why three years? Is it because they lack a particular skill? Have they not yet demonstrated the ability to operate at the required level? Do they need experience managing a larger function? Is there insufficient evidence of their capability?
Without understanding the reasoning, the manager has little basis for evaluating the recommendation. And the employee has little basis for acting on it.
A better approach would explain which requirements have been met, which still need to be demonstrated, and what evidence or development experiences could improve readiness. It should also distinguish between what is known, what has been inferred and what remains uncertain. That makes it possible for the manager to challenge the recommendation, introduce additional evidence and make a more informed decision. Transparency is particularly important because AI can sound extremely confident even when its underlying information is incomplete.
The test: Can the people using the technology understand, question and challenge its recommendations?
7. Protect data: how do you keep employee data safe when using AI in HR?
Does the technology protect people's information and ensure that it is used appropriately?
Talent management involves some of the most sensitive information an organisation holds. Assessment results, performance reviews, career aspirations, succession plans, behavioural feedback and retention risks can all have a significant impact on somebody's career. The fact that AI can use this information to produce useful insights doesn't automatically mean that every application should have unrestricted access to it.
Consider a manager who wants help preparing for a difficult performance conversation. They might be tempted to paste an employee's entire performance history, assessment report and personal circumstances into a general-purpose AI tool. The output might be useful.
But what information has been shared? Where has it gone? Who can access it? How long will it be retained? And was all that personal information necessary to produce the recommendation?
Responsible AI design starts by determining what information is required for a particular purpose, and how that information can be protected. That includes appropriate access controls, minimising personal identifiers, clear data-use rules and ensuring that sensitive information isn't shared with systems that haven't been approved for that purpose. It also means being transparent with employees about how their information is used.
The test: Are we protecting people's information by design, so it stays safe whichever user is interacting with the AI?
How to evaluate your own AI implementation
These seven principles are useful as design guidelines, but I think they're even more useful when they're used to evaluate technology. Whether you're considering a new recruitment platform, introducing AI into performance management or building an internal talent assistant, you should be able to assess the technology against each principle.
And importantly, you should be able to provide evidence to support your assessment. Here's a simple framework that HR teams can use.
The seven-principle AI assessment
For each principle, answer the question and give your implementation a score from 1 to 5.
| Principle | Evaluation question | What evidence would demonstrate success? |
|---|---|---|
| Human Trust | Is AI improving the quality of interactions and decisions between people? | Feedback from employees and managers, quality of career or performance conversations, adoption and trust measures |
| Challenge Bias | Does AI identify and reduce the influence of bias? | Bias testing, review of selection and assessment outcomes, documented challenges to recommendations |
| Talent Value | Does AI help recognise and maximise individual strengths and contributions? | Evidence of better role fit, targeted development, internal mobility and effective use of strengths |
| Contextual Intelligence | Are recommendations grounded in appropriate frameworks, data and organisational context? | Documented frameworks, data quality checks, expert review of recommendations and relevance to roles |
| Cultural Integrity | Does AI encourage behaviours consistent with the organisation's values? | Manager behaviour, quality of feedback, accountability, employee experience and adoption of desired practices |
| Transparency | Can users understand and challenge recommendations? | Clear explanations, visible assumptions, uncertainty indicators and processes for human review |
| Data Protection | Is personal information protected and used only for appropriate purposes? | Privacy assessments, access controls, data minimisation, approved processing arrangements and security testing |
How to score your AI implementation
Use the same five-point scale for each principle.
| Score | Description |
|---|---|
| 1: Not considered | The principle wasn't part of the design or selection process. |
| 2: Intended | The principle is recognised, but there is little evidence that it has been implemented. |
| 3: Implemented | The principle has been incorporated into the technology or process, but its effectiveness hasn't been properly validated. |
| 4: Validated | There is evidence that the implementation meets the principle in its intended use. |
| 5: Continuously improved | The principle is regularly measured, tested and used to improve the technology and the way people use it. |
You could calculate a total score out of 35, but I wouldn't use that as the only measure of success. A technology that scores exceptionally well on six principles but fails to protect employee data shouldn't receive a pass because its average score looks good. The same applies to a recruitment system that is efficient, transparent and easy to use but consistently excludes qualified candidates from particular groups.
The individual scores matter more than the total.
And there's another question I'd add before completing the assessment: what specific outcome are we trying to improve, and how will we know whether we've improved it?
- For recruitment, that might be the quality and fairness of shortlisting, measured alongside time-to-shortlist.
- For performance management, it might be the quality of goals, feedback and performance improvement.
- For succession, it might be the reliability of successor readiness and the strength of the leadership pipeline.
- For development, it might be improved capability and contribution.
The principles tell us how AI should operate. The outcomes tell us whether it's creating value. We need both.
The real opportunity for AI in HR
I've been involved in developing technology for talent management since we built our first web-based application back in 1997. One of the things that experience has taught me is that technology doesn't automatically improve a process. If the underlying process is poor, technology can help you execute it faster and more consistently.
AI makes this even more important because it can also interpret information, generate recommendations and influence decisions. That creates enormous opportunities. But it also means we need to be much clearer about the standards we're using to judge whether the technology is effective.
For me, the objective is to improve the quality of the human judgement involved in talent management. That means making better information, relevant research and practical guidance available when decisions are being made.
Before asking what AI can do for your HR function, ask what you want it to make better. Better people decisions? More productive career conversations? Fairer assessments? More credible successors? More effective development? Then establish the principles that will guide how you build, buy and deploy the technology.
Because an AI tool that saves time but produces worse decisions isn't an improvement. It's just a faster way of getting things wrong.
Frequently asked questions about AI in HR
What are some examples of AI in HR?
Common uses include CV screening and shortlisting, drafting performance reviews, recommending development plans, identifying successors, flagging retention risk and helping managers prepare for career conversations. Each of the seven principles above includes an example of how one of these uses can go well or badly.
What are the benefits of AI in HR?
Used well, AI gives managers better information at the moment they make a people decision. It can bring together assessment results, skills data and performance history, flag where a judgement might be biased, and suggest questions for a career or performance conversation. The benefit shows up in the quality of decisions, so measure outcomes such as fairer shortlists and stronger successors alongside time saved.
Will AI replace HR?
AI will take over some administrative and analytical work in HR. People decisions still need human judgement and someone accountable for them. The aim is to improve that judgement with better information. A manager should always be able to question a recommendation, add context and explain the final decision, especially when it affects someone's career.
What laws apply to using AI in HR?
It depends on where you operate. In South Africa, the Protection of Personal Information Act (POPIA) governs how you process employee data. In the European Union, the AI Act classes many AI systems used in recruitment and employment decisions as high-risk, which brings obligations around risk management, human oversight and transparency. Check with your legal adviser before you deploy a tool that informs decisions about people.
