AI in the workplace raises a practical question for South African employers: what should staff be told when technology records their conversations, measures their performance or helps decide who gets hired or promoted?
The answer depends on what the system does. A tool that helps draft a general presentation raises different concerns from software that analyses employees’ behaviour or ranks candidates for a vacancy.
Employers should identify the privacy, employment and contractual obligations attached to each use. Calling a system “AI” does not create an exemption from ordinary law. Equally, it does not mean that every use requires the same announcement, consent form or consultation process.
The important distinction is between disclosure required by law and additional transparency that helps a workplace use technology responsibly.
Start with what the AI tool actually does
Before preparing a staff notice, establish how the system operates within your business.
Does it produce a draft that someone checks? Does it record conversations? Does it combine information from several systems to create a performance score? Can its output affect an employee’s pay, working hours or continued employment?
Those differences should shape the assessment.
A useful starting point is to record:
- What information enters the system.
- Where that information comes from.
- What the system produces.
- Who receives or uses the output.
- Which decisions it influences.
- Who checks its accuracy.
- How errors can be corrected.
A supplier’s product description may not answer all these questions. Employers need to understand how the tool is configured and used in their own workplace.
Separate administrative assistance from assessment
Consider an AI meeting assistant used to prepare minutes. Even where its purpose is administrative, it may record names, opinions and confidential discussions.
Now consider software that uses those recordings to assess participation or leadership potential. That is a materially different use.
A system introduced for one purpose should not gradually become an employee assessment tool without a fresh review of its purpose, disclosures and safeguards.
Tell staff how their personal information is collected and used
POPIA applies to personal information processed through workplace systems. Section 18 sets out collection disclosures, subject to exceptions. These include the purpose, responsible party, relevant recipients and applicable rights.
The practical task is to give employees a notice that matches the actual processing. A general statement that the business “uses technology” tells them very little.
For a tool recording meetings, explain when recording occurs, what records are created and who can access them. For productivity software, explain what activity is captured and how management intends to use it.
Put important information where staff will see it
A privacy policy can provide detail, but significant new practices deserve clear communication.
A short introduction supported by a fuller notice is often more useful than sending a long document with no explanation of what has changed.
For example, a staff communication might explain that meetings will be transcribed to prepare minutes, that participants can report transcription errors and that access is restricted to particular roles. It should also accurately describe retention arrangements and any sharing with service providers.
This is a recommended way to communicate clearly, rather than a prescribed format for every workplace.
Disclosure does not make every use lawful
Telling staff about a system is only part of the assessment.
An employer should still examine whether the proposed collection is necessary, whether the purpose is legitimate and whether the system gathers more information than the task requires.
Nor should a consent form be treated as a universal solution. The appropriate legal basis needs to be identified for the particular activity.
An employee’s acknowledgement that they received a notice is also different from giving consent. Documents should make that distinction clear.
Explain how AI influences employment decisions
Staff are likely to be concerned when a system contributes to decisions about performance, promotion, scheduling or discipline.
As a practical transparency measure, employers should explain the system’s role. Is its score one source of information, or does it effectively determine the result?
A statement that “management makes the final decision” is unhelpful if managers routinely accept the output without investigating it.
Understand the limits on solely automated decisions
Section 71 of POPIA restricts certain decisions based solely on automated profiling that have legal or substantial effects. It contains exceptions, with safeguards including an opportunity to make representations and sufficient information about the underlying logic in relevant cases.
This is not a blanket prohibition on all workplace algorithms. The particular process and its consequences need assessment.
Employers should avoid assuming that a human signature resolves the issue. Meaningful review requires someone who can understand the result, question the supporting information and change the outcome where justified.
A score needs context
Suppose software rates an employee poorly because fewer tasks were completed during a particular week.
The employee may have been training a colleague, handling unusually complex work or dealing with a system outage. Those circumstances may not appear in the data used to calculate the score.
A manager should therefore ask what the result measures and what it leaves out.
Before relying on AI output in an employment process, establish whether the information is accurate, relevant and sufficiently complete. Keep a record of the reasoning behind the decision.
Check for unfair discrimination
An automated system can reproduce patterns in the information used to develop it. It can also produce unfair results because its assessment criteria do not suit the workplace.
For example, a recruitment tool might favour career histories resembling those of previous successful applicants. That may overlook candidates with different but relevant experience.
The Department of Employment and Labour’s guidance explains that employers must address unfair discrimination in employment policies and practices. Using software does not remove that responsibility.
Test the results before trusting them
Employers should examine whether the tool assesses qualities that are relevant to the job.
A system that works adequately for one role may be unsuitable for another. Differences in language, working arrangements or accessibility needs may also affect its output.
Testing should include realistic examples and a review of questionable results. Ask the supplier how the system has been evaluated, but also check how it performs in your own setting.
An explanation that “the algorithm selected them” does not tell anyone why the selection was appropriate.
Give employees clear rules for their own AI use
Transparency should work in both directions. Staff need to understand what they may do with AI tools in the course of their work.
A policy should identify approved tools, permitted activities and situations requiring permission. It should explain the treatment of personal information, confidential material and client records.
Practical rules can address:
- Whether public AI tools may be used.
- What information must never be entered.
- Which tasks require an approved business account.
- How factual claims and references must be checked.
- Who approves important output.
- When AI assistance must be disclosed internally or externally.
These are policy choices that should reflect the employer’s legal obligations and business needs. They should not be presented as identical statutory requirements for every role.
Make the rules workable
A policy saying “use AI responsibly” leaves too much unanswered.
An employee preparing a general outline needs different guidance from someone reviewing confidential personnel records. Examples should reflect the work people actually perform.
Training also needs to explain what to do when something goes wrong. Staff should know whom to contact if they enter information into an unapproved tool or discover an inaccurate result.
Prompt reporting gives the business a better opportunity to respond.
Consult where workplace changes trigger consultation duties
Introducing new software does not automatically require the same formal consultation process in every business.
The position depends on what changes, existing agreements and the applicable legal requirements. Employers should check these before implementing changes to employment arrangements.
Technology does not bypass retrenchment procedures
Where an employer contemplates dismissals for operational requirements, section 189 of the Labour Relations Act requires consultation with the appropriate parties. Technological needs fall within the Act’s definition of operational requirements. Larger retrenchments may engage section 189A.
An announcement that AI has made certain roles unnecessary is not a substitute for the required process.
Consultation should address the relevant proposals and alternatives. Depending on the circumstances, those alternatives might include training, redeployment or a different allocation of work.
Engage before assumptions become fixed
Even where a particular formal consultation duty is not triggered, early discussion can improve implementation.
Employees may identify work that the proposed system does not capture or explain why a suggested performance measure is misleading.
That information is valuable before management builds targets, staffing decisions or disciplinary expectations around the tool.
Provide a clear route to question the output
Employees should know how to report inaccurate information or challenge a result.
POPIA provides qualified rights of access and correction, while employment disputes follow the applicable labour processes. An internal AI review procedure should support those routes rather than replace them.
As a practical measure, identify who receives concerns, what supporting information is useful and who can reconsider the matter.
Make review capable of changing the outcome
A complaint process is of limited value if the reviewer cannot investigate the data or depart from the system’s recommendation.
Record what was challenged, what was checked and why the original result was confirmed or changed. Where an error affects other employees, consider whether a wider review is needed.
An employee should not have to discover the existence of a scoring system only after being told that it has counted against them.
Do not wait for a dedicated AI policy
South Africa’s national AI policy process has been developing. Cabinet’s June 2026 statement confirmed the withdrawal of the draft policy for reworking. A draft policy should not be confused with legislation already in force.
Employers still need to assess their systems against existing privacy, equality, labour and contractual obligations.
The useful starting point is to explain meaningful use, limit unnecessary collection and establish who is responsible for decisions. Then test whether those arrangements work in practice.
AI can assist a workplace. Employees should be able to understand when it affects them, and employers should be able to explain why relying on it was justified.
