Why AI Shapes HR Decisions Before Anyone Approves Them
Quick answer:
AI already shapes many HR decisions before a person approves them. Scrutiny should rise with the effect on the individual, human review must be able to reach a different conclusion, and someone must be able to explain the outcome.
AI is already influencing decisions about people.
It ranks candidates, summarises employee feedback, identifies skills gaps, predicts who might leave and suggests what managers should do next. Yet we still tend to describe it as something that simply supports a human decision.
Sometimes that is true. Sometimes the technology has already shaped most of what the person sees by the time they become involved.
We have spent a lot of time asking what AI can do for HR. We now need to ask what we are comfortable allowing it to influence.
Is AI already making decisions about people?
In practice, yes. It often shapes the decision before anyone approves it.
If a system ranks candidates before a recruiter reviews them, it has influenced who receives attention. If it describes an employee as a flight risk, that may change the way their manager interprets what they say and do. If it turns thousands of survey comments into a short summary, senior leaders will probably discuss what appeared in that summary. They may never know what was left out.
A person might still approve the final outcome, but that does not tell us where the decision really started.
What problem is AI in HR actually solving?
Often, a problem that was never really about the system.
HR has bought plenty of technology over the years, hoping it would solve problems that were never really about the system. A process takes too long, so we automate it without asking why it is so complicated. Information is inconsistent, so we move it to a new platform without agreeing on what it means. Managers lack confidence, so we give them another workflow and hope the prompts will make up the difference.
Sometimes it works. Sometimes we end up doing the same confused work more quickly. AI will not somehow avoid this.
If an organisation has never agreed what good performance looks like, putting AI into the performance process will not resolve that disagreement. If job titles and skills data are inconsistent, a new system will not make the information reliable. If managers avoid difficult conversations, generated guidance may help them prepare. It may also give them words that sound right but which they do not really understand.
The awkward part is that the output can still look impressive. A well-written summary seems complete. A score looks precise. A ranked list gives the impression that a careful comparison has already happened. We can easily give the presentation more confidence than the underlying reasoning deserves.
Why does AI in recruitment favour familiar candidates?
Because a system trained on previous appointments finds the people you already hire easier to recognise.
Imagine two people applying for a front office role. One has worked for recognised hotel brands, held familiar job titles and followed the sort of career path we expect to see. The other has built similar capability through independent properties, international experience or a career that has moved in a less obvious direction.
An experienced recruiter might look at both and see potential. A system trained on previous appointments may find the first person easier to recognise. That does not necessarily mean the technology has failed. It may have done exactly what it was asked to do. The organisation may simply have asked it to find people who look like those it chose before.
Human recruiters do this too. We tend to feel more comfortable with experience we recognise, and we sometimes treat a conventional background as evidence that an appointment will be less risky.
AI may help reduce some of those habits. It could also repeat them across far more applications while giving the result an appearance of objectivity. We need to know which of those things is happening. A supplier saying the system is accurate does not answer that question.
Which uses of AI in HR need the most scrutiny?
The uses with the greatest possible effect on a person. Not every use needs the same level of care.
I want HR teams to use AI well. Employees should not have to wait several days for an answer to a routine question. Managers should not need to search through several systems to find basic guidance. Experienced HR professionals should not spend hours arranging information that technology can organise in seconds.
Using AI to improve the wording of a routine email does not require a committee or a lengthy approval process. Somebody needs to check the result and apply normal judgement. That will usually be enough.
More care is needed as the possible effect on another person grows. An AI assistant helping an employee find the annual leave policy mainly needs to give an accurate, up-to-date answer. A tool summarising employee feedback needs closer attention. It may capture the broad themes while losing a smaller but important group of comments.
A recruitment system ranking candidates needs more scrutiny again. Its recommendation may influence who receives access to a job, even if a recruiter technically makes the final decision. It makes little sense to apply the same controls to all three uses.
Some organisations will respond by making every use difficult to approve. Employees will then use publicly available tools without telling anyone. Others will take a relaxed view of almost everything until something goes wrong. Neither response feels particularly sensible. The organisation needs to understand each use well enough to decide what level of control is reasonable.
How does AI become part of normal HR work?
A little at a time. The work changes first, and the formal process catches up later.
Someone tries a tool and finds that it saves an hour. Their colleagues begin using it. A supplier adds a new feature to an HR platform the organisation already owns. A recommendation starts appearing in a manager’s workflow. After a while, people stop thinking of it as new.
The scale is what makes AI different. An assumption built into one process can quickly affect hundreds or thousands of decisions. The output can also be hard to question during a normal working day.
A hotel manager might be dealing with a guest complaint, staff absence, agency costs and an employee issue before lunch. If the HR system suggests what to do next, the manager is likely to want a straightforward answer about whether they can rely on it. They will not have the time or access needed to investigate how the system reached its recommendation. We can say that the manager remains accountable, but that starts to feel unfair if they cannot see the information behind the answer or understand the system’s limitations.
The way the process is designed will also influence what they do. If accepting the recommendation takes two minutes but choosing a different course requires another form, a written explanation and additional approval, most people will follow the recommendation. They have not been told that they must. They have simply been shown which option the organisation expects to be easier.
What should human review of AI decisions involve?
The reviewer needs to understand the recommendation, have time to consider it, and know that they can disagree.
We often hear that a human will remain “in the loop”. I am not convinced that the phrase tells us very much. The person may be reviewing the evidence, adding context and reaching their own conclusion. They may understand what the technology has considered and what it could have missed. They may simply be approving what appears on the screen.
For human review to mean something, the reviewer needs to understand the subject and see enough information to make sense of the recommendation. They need time to consider it. They also need to know that they can disagree. That last point will depend on what the organisation does, not what the policy says.
A manager may officially be allowed to override a recommendation. If every override is questioned, highlighted in reports or treated as a failure to follow the process, people will quickly become reluctant to do it. The review then becomes a formality.
For decisions about recruitment, performance, pay, progression or redundancy, somebody should be able to explain the outcome in ordinary language. They should know what information mattered, how the system contributed and why the final decision was reasonable. If nobody can do that, attaching a person’s name to the outcome does not make the process meaningfully human.
Why does HR need to be involved in AI governance?
Because HR understands what happens when organisational processes meet real working lives.
AI governance cannot sit with HR alone. Technology teams understand systems and security. Legal and risk colleagues bring different expertise. Procurement should know what suppliers have promised and what happens to the information. Operational leaders know whether the process will work outside the project meeting.
HR sees the gap between what a policy expects and what a manager can reasonably do. We know that an efficient employment process can still produce a poor or confusing experience. We also know how easily accountability becomes blurred when several functions each own one part of the process. That gives HR a clear role.
Candidates should know when AI has materially influenced the way their application was assessed. Employees should be able to correct inaccurate information. Managers should understand the limitations of recommendations they are expected to use. Somebody should be able to explain a significant decision without hiding behind the technology or the supplier.
None of this is excessive. Some organisations may find that greater transparency makes a particular use of AI harder to justify. That is worth knowing before the system goes live, not after an employee or candidate challenges the outcome.
People are also entitled to feel differently about different uses of AI. Someone may welcome a tool that removes a tedious task and dislike a system assessing their performance. That is not inconsistency. The first use helps them. The second has some power over them. HR should understand why that matters.
How can HR make a conscious choice about AI?
By answering six questions before AI goes into any significant HR process.
Understand what it is doing. Which part of the decision does the system shape, and at what point does a person become involved?
Know what information it relies on. If the underlying data is inconsistent, the output will be too, however precise it looks.
Ask what it might fail to recognise. A system trained on previous appointments will find familiar profiles easier to see.
Know how much weight managers will give its output. If accepting the recommendation is easy and departing from it is hard, the process has already decided.
Know what happens if the result is wrong. Who notices, who can correct it, and what the person affected is told.
Get a clear answer about who remains responsible. That responsibility needs to exist in practice, not on paper.
Responsibility cannot be spread across so many functions, committees and suppliers that nobody can explain the outcome. AI should help HR use its time and information better. It should not make responsibility harder to find.
The important question is no longer whether HR will use AI. It is what we allow it to influence and whether we make that choice with our eyes open.
How AI in HR connects to judgement, systems and blind spots
AI sharpens a question this series has kept returning to: where does the decision really start?
A recommendation on a screen is not a judgement. Someone still has to make one, and judgement, not policy, is the capability that shapes every organisation. When overriding a system takes a form and an explanation, the organisation is quietly teaching people to comply rather than decide.
The way AI arrives, a feature at a time, is how the systems organisations learn to defend are built. After a while nobody thinks of it as new, which is exactly how organisational blind spots form. What has become familiar stops being noticed, including who is really making the decision.
In a nutshell
AI can influence an outcome before anybody makes the final decision.
A system may work as designed while still reflecting assumptions the organisation has not properly examined.
The level of scrutiny should increase as the possible effect on candidates or employees grows.
Human review only matters when the reviewer understands the output and can reach a different conclusion.
HR should help ensure that significant decisions remain understandable, challengeable and clearly owned.
Frequently asked questions
Is AI already making decisions about people in HR?
In practice, often yes. AI ranks candidates, summarises feedback, predicts who might leave and suggests what managers should do next. A person may approve the final outcome, but the system has frequently shaped most of what that person sees. Approval does not tell us where the decision really started.
What does “human in the loop” mean for HR decisions?
On its own, very little. The reviewer may be weighing the evidence and adding context, or simply approving what appears on the screen. Meaningful review needs a person who understands the subject, sees enough information to make sense of the recommendation, has time to consider it, and knows in practice that they can disagree.
How can AI in recruitment repeat existing hiring bias?
A system trained on previous appointments learns to recognise the profiles the organisation already hires, such as familiar hotel brands and conventional career paths. It may then repeat that preference across far more applications while giving the result an appearance of objectivity. It has not necessarily failed. It has found people who look like those chosen before.
Which HR uses of AI need the least and the most scrutiny?
Improving the wording of a routine email needs a quick check and normal judgement. An assistant answering policy questions mainly needs to be accurate. A tool summarising employee feedback needs closer attention, and a system ranking candidates needs more scrutiny again, because it influences who receives access to a job. The level of control should rise with the possible effect on a person.
Why should HR be involved in AI governance?
Because HR understands what happens when organisational processes meet real working lives. HR sees the gap between what a policy expects and what a manager can reasonably do, knows that an efficient process can still produce a confusing experience, and knows how easily accountability blurs when several functions each own one part of the process.
What should candidates and employees be told about AI in HR decisions?
Candidates should know when AI has materially influenced the way their application was assessed. Employees should be able to correct inaccurate information. Managers should understand the limitations of the recommendations they are expected to use. And somebody should be able to explain a significant decision without hiding behind the technology or the supplier.
