AI in Hiring in South Africa: Where It Helps, Where the Law Bites

AI in hiring in South Africa has moved from experiment to routine, and most small and mid-sized employers crossed that line without a policy. A post goes live on a Monday, applications arrive in the hundreds, and software has already ranked most of them before a person opens the first CV. The question for an owner is which parts of the decision you hand over and which you keep.

Shortlisting is where AI helps most and where the Employment Equity risk concentrates.

Start with the volume. Stats SA's Quarterly Labour Force Survey put the official unemployment rate at 33.6% in the second quarter of 2026, up 0.9 of a percentage point from 32.7%, with 8.5 million people out of work and employment 16 000 lower at 16.7 million. You can also read the underlying QLFS release. When that many people chase work, every advertised post becomes a magnet.

The recruitment data shows what that does to one advert. CareerJunction's second-quarter insights found the average applications per advertised position up 70% year on year, and 32% above two years earlier, even as recruiters cut their own CV database searches by 18.1%. The graduate market is the extreme case: a record 138 applications per vacancy in 2026, with only 12% reaching a first-round interview, and 88% of candidates using AI somewhere in their job search, up from 63% in 2025.

Building a shortlist in a week from 300 applications is not a hiring process. It is triage.

Why is AI in hiring in South Africa suddenly unavoidable?

Because the manual alternative stopped scaling. A two-person team in Sandton reviewing 250 CVs at four minutes each faces more than 16 hours of reading before a single interview is booked. Three forces pushed the tools in at once.

• Application volume no small team can read by hand in a week.

• Applicants using AI themselves, which makes every CV smoother and more alike.

• Screening features bundled into the applicant tracking systems employers already pay for.

Adoption followed fast. PwC's 2026 Global AI Jobs Barometer, covering more than a billion job ads, found roles requiring AI skills growing at 69% against 9% for the wider market, and the share of South African postings demanding AI skills nearly doubled to 3% by 2025. On the employer side, Sage's HR and Payroll Leaders' Report found 57% of South African HR professionals use AI in recruitment, and 53% name compliance and data security as their main worry.

Where do screening tools genuinely help a South African employer?

The wins that survive a real hiring round are unglamorous. They remove drudgery and enforce consistency.

• Drafting job specs so the advert says what the role needs.

• Parsing CVs into a consistent structure, so an applicant from Soweto is read against the same fields as one from Rosebank.

• Ranking applicants against the published, job-related criteria you set.

• Scheduling interviews and chasing confirmations without a person on the phone all morning.

• Answering routine candidate questions about the closing date and the documents required.

• Drafting interview guides tied to the role's competencies and summarising structured notes.

None of those steps decides anything. They prepare the material a person then decides on, and they do it the same way every time, which is exactly what an unfair discrimination claim attacks. Pnet reports an average of 161 applications per advertised post, which is why a structured first pass beats an exhausted skim.

The useful test is simple: does the tool make the process more consistent, or does it make the decision on your behalf?

Where does AI in hiring go wrong?

The failures cluster in one place: systems that infer something about a person rather than measure something about the job. Score a candidate's words and your model is reading their schooling, home language and postcode. Score a face or a voice and you are guessing at accent, age, disability and composure, none of which predicts whether someone can do the work.

• Video and voice analysis sold as a measure of employability or culture fit.

• Personality or writing-style inference used to filter applicants out.

• Models trained on your existing staff, which reproduce the people you already hired.

• Criteria that proxy for race, gender, age or disability, such as a prestigious-school list.

• Fully automated rejection with no human review and no recorded reason.

If a model cannot explain, in plain language, why it dropped a candidate, that candidate cannot be given a fair reason, and neither can you.

The ILO's research on AI in human resource management makes the same point. In ILO Working Paper 154, Janine Berg and Hannah Johnston argue that much of the optimism around AI in HR is misplaced, because these systems automate flawed processes at scale rather than repair them. A scoring tool inherits every assumption in its criteria and data, then applies it to 400 people instead of four. And the market is watching: a screening bot that goes silent on 200 people after a Germiston warehouse advert costs more in employer brand than it saves in admin.

Hiring use case	Risk level	What good looks like
Drafting job specs and adverts	Use freely	A human signs off the wording
Parsing and structuring CVs	Use freely	The same fields for every applicant
Scheduling and status emails	Use freely	The candidate can still reach a person
Ranking against job-related criteria	Use with controls	Criteria published, scores logged, shortlist reviewed by a person
Answering routine candidate questions	Use with controls	An escalation path to a human and no personal data in prompts
Video, voice or writing-style scoring	Do not use	No defensible link to performance in the job
Automated rejection without review	Do not use	Runs straight into POPIA section 71

Table 1: hiring use cases ranked by risk

What does South African law require when you screen candidates?

There is no single AI hiring law here. There is a stack of ordinary employment and data laws that applied the moment you advertised. Five instruments do most of the work.

• Employment Equity Act 55 of 1998, which bans unfair discrimination in recruitment and selection.

• Employment Equity Amendment Act 4 of 2022, which reset who counts as a designated employer and introduced sectoral targets.

• Labour Relations Act 66 of 1995, which governs fair dismissal and unfair labour practices.

• Basic Conditions of Employment Act 75 of 1997, which controls the particulars you must put in writing.

• Protection of Personal Information Act 4 of 2013 (POPIA), which governs every CV and interview note you hold.

The Employment Equity amendment matters more than most owners realise. The turnover threshold is gone, so from 50 employees upwards you are a designated employer on headcount alone, and the sectoral targets published in April 2025 cover 18 national economic sectors. The first assessment window opened on 1 September 2026 and runs to 15 January 2027, with penalties reaching R1.5 million or 2% of annual turnover. A model that filters on proxies for race or gender creates more than a fairness problem: it works against the very targets you must report on.

The duty underneath it all is simpler and older. If a selection criterion is challenged, the employer must show it is job related and applied consistently. A rule you can write down and defend survives that test; a weight buried in a vendor's algorithm does not. The Department of Employment and Labour publishes the current forms.

You do not have to prove the tool was clever. You have to prove the criterion was job related and the process was the same for everyone.

Hiring stage	Main duty	Instrument
Advert and job spec	No discriminatory requirements, and criteria that are job related	Employment Equity Act
CV collection	Collect only what you need, secure it, say why	POPIA
Shortlisting	Consistent criteria, with a record of how the shortlist was built	Employment Equity Act
Automated rejection	No decision resting solely on automated processing	POPIA section 71
Interview and assessment	The same structured process for comparable candidates	Employment Equity Act
Offer and appointment	Accurate written particulars of employment	Basic Conditions of Employment Act
Retention afterwards	Keep only as long as necessary, then delete safely	POPIA

Table 2: obligations mapped to the hiring stage

What are your duties to candidates under POPIA?

POPIA treats a CV as personal information, whoever sent it. You need a lawful basis, a purpose you can state, a retention rule and reasonable security. The section that catches hiring workflows most often is section 71, which covers decisions based solely on automated processing that affect someone substantially. The Information Regulator enforces it.

A hiring decision is about as substantial as a decision gets, and the word doing the work is solely. If software scores and ranks and a person then reads and decides, you are on firmer ground than if the system rejects anyone below a threshold unseen.

• Tell candidates, in the advert or application flow, what you collect and why.

• Collect only what the role needs, not every field the form offers.

• Keep a record of who reviewed what, so any decision can be explained.

• Confirm a named Information Officer exists and can be reached.

• Set a retention period for unsuccessful applications and delete at the end of it.

• Secure the data. No CVs in a shared WhatsApp group, no exports to a personal drive.

The awkward question a candidate can ask is not whether you used AI. It is what you decided about me, and who decided it.

Our piece on POPIA compliance for small business covers the wider data duties.

How do you keep a human in the decision?

Design the process so the automated step always ends in a handover. The tool produces a ranked, annotated shortlist; a named person accepts, rejects or reorders it; and the reason for the final call is written in one line. That removes the solely automated problem and creates the paper trail a dispute will demand. Write them into a one-page policy.

• One named decision maker per vacancy, recorded in the file.

• Criteria published in the advert and used to score every applicant.

• No scoring of faces, voices or writing style, at all.

• Every shortlist reviewed by a person before any rejection is sent.

• A one-line reason for each candidate who is not progressed.

• A deletion date attached to every application.

• No candidate personal data pasted into a public AI tool.

• An annual check that the criteria and the tool still match the role.

A defensible hiring decision fits on one page: the criteria, the scores, the shortlist and the name of the person who chose it.

What should a Gauteng hiring team do this quarter?

Three cheap moves. List every tool that touches a candidate, from your applicant tracking system to the assistant a manager uses to summarise CVs. Cut anything that scores a face, voice or style, because nothing in your process will ever justify it. Write your selection criteria in job-related language, publish them in the advert and keep them fixed for the round.

Then do the dull work: confirm your Information Officer, decide how long you keep unsuccessful applications, and make one person accountable for each shortlist. If you are standing the process up from scratch, an AI readiness assessment plus a workshop for hiring managers gets you further than another software subscription, because the risk lives in the process, not the tool. Our AI training work with Johannesburg teams covers this step.

This article is general information for business owners. It is not legal advice, and employment and data protection law turns on your own facts, so take advice before you change your process.

Candidates remember how you treated them whether or not they got the job. In a market this tight, an applicant at a Midrand business today is a customer and a referral tomorrow, so buying time by treating people badly is the most expensive saving there is. For more on how work is changing, see our notes on the future of work in South Africa.

AI in hiring in South Africa: Quick Answers

Is AI screening legal in South Africa?

Yes. Screening software is lawful, but the law regulates the decision rather than the tool. The Employment Equity Act requires selection criteria to be job related and applied fairly, and POPIA section 71 stops a decision resting solely on automated processing. Keep a genuine human review step and you are within the rules.

Does POPIA require consent to screen a CV?

Not always. Consent is one lawful basis among several, and a CV sent in response to your advert is already processed for a purpose the candidate understands. POPIA requires a lawful basis, collection limited to what the role needs, and deletion once the purpose ends.

Can I reject a candidate based on an AI score alone?

You should not. POPIA section 71 protects a person from a decision that affects them substantially and rests solely on automated processing, and a rejection qualifies. Have a person review the file and record a short reason before any rejection goes out.

How long may I keep CVs from unsuccessful candidates?

Only as long as you need them, and you must be able to say what that period is. A defensible approach is to keep applications for the recruitment round plus a short window for disputes, then delete or archive them. Keeping every CV forever is the habit POPIA was written to end.

Must I tell candidates that AI was used?

You must be transparent about how their personal information is processed, which is easier if you say plainly that screening tools support a human shortlisting process. There is no requirement to describe your software in detail, but there is one not to hide how a decision about them was made when they ask.

What happens if a rejected candidate complains?

You will need to explain the criteria, show how they were applied and produce the record of who decided. A candidate can also approach the Information Regulator about how their information was handled. One page of notes per round answers those questions in an afternoon; without it you start reconstructing months later.

If you want help mapping your hiring process before the next round opens, talk to us at Social Horde Agency. We work with South African businesses on the strategy, process and people side of this, and the reading behind it sits in our blog and our work with clients. You can read more about the consultancy and what an engagement looks like.