An AI use policy for companies in South Africa, with a template

An AI use policy tells staff which AI tools they may use for work, what they may never put into them, who checks the output, and what to do when something goes wrong. In South Africa most of its hard rules already exist in the Protection of Personal Information Act (POPIA): security safeguards, written contracts with providers that process personal information for you, limits on sending it abroad, and limits on decisions made by a machine alone. The rest is your company's choice, and the template below marks which is which.

Write it before staff settle into habits with free tools.

None that we found. When we checked on 12 October 2026 we found no statute written for AI, and no adopted national AI policy.

The Department of Communications and Digital Technologies has published two documents on the subject:

• A national AI policy framework. The South Africa National Artificial Intelligence Policy Framework is dated October 2024 (the draft policy calls it the framework of August 2024), and the draft policy describes itself as based on it and on the submissions received on it. A framework is a starting point for policy, not policy.
• A draft national AI policy. The Minister published the Draft South Africa National Artificial Intelligence (AI) Policy for public comment in the Government Gazette on 10 April 2026, after Cabinet approved it for comment. Its withdrawal was announced on 26 April 2026, and by a further notice in the Gazette of 12 June 2026, the Minister withdrew it "in its entirety".

So the status is a framework and a withdrawn draft. Until a policy is adopted, and any law follows it, your AI use is governed by POPIA, by your contracts and by any rules for your sector. Put the national policy on your review list, not in your policy.

In our reading, typing customer details into an AI tool is processing personal information, so the Act's conditions apply. The duties that matter most for an AI policy:

• Safeguards. Section 19 requires appropriate, reasonable technical and organisational measures against loss and unlawful access. You must identify "all reasonably foreseeable internal and external risks", keep safeguards against them, check that they work and update them.
• Providers. Under section 20 an operator processes personal information "only with the knowledge or authorisation of the responsible party" and keeps it confidential. Section 21 requires a written contract binding the operator to the security measures in section 19, and the operator "must notify the responsible party immediately" of a suspected breach.
• Breaches. When there are reasonable grounds to believe someone unauthorised has reached personal information, you notify the Information Regulator and the people affected "as soon as reasonably possible" under section 22.
• Decisions by machine. Section 71 bars a decision with legal consequences, or substantial effect, based solely on automated processing that profiles a person. There are exceptions, and where a decision tied to a contract relies on safeguards instead of meeting the person's request, those safeguards must let the person make representations and give them enough information about the underlying logic.
• Transfers abroad. Section 72 allows personal information to go to a third party in a foreign country only on listed grounds, such as a recipient bound by a law, "binding corporate rules or binding agreement" giving adequate protection, or the person's consent.
• Records. Section 17 requires you to keep documentation of all processing operations under your responsibility.

The regulations add two duties for your information officer that fit an AI rollout: a personal information impact assessment, and internal awareness sessions on the Act.

Copy the sections below, replace the words in square brackets, and add your approved tools as an appendix. Where a section rests on the Act, it says so; everything else is a choice we recommend, and you may change it.

1. Purpose and scope. This policy sets how [Company] uses artificial intelligence tools. It applies to every employee, contractor and temporary worker who uses an AI tool for company work, on any device. An AI tool is any software that produces text, images, audio, code, summaries or decisions from a prompt or from data, including chatbots, copilots, meeting notetakers, transcription and AI features inside software the company already uses.
2. Owner. [Role] owns this policy and the approved tools list. The information officer, whose duties under section 55 of POPIA include encouraging compliance with the conditions for lawful processing, approves every change that touches personal information.
3. Approved tools. Staff use only the tools on the approved list, through accounts in the company's name. Personal accounts and free consumer versions are not approved for company information. A tool joins the list only once [Role] has read its business terms on training, retention and where data is processed, and recorded them; made sure the impact assessment in section six is done where personal information is involved; signed the provider contract in section seven where the Act requires one; and recorded the ground for any transfer abroad in section eight.
4. Information that may never be entered. No one enters the following into any AI tool unless the tool is approved for that purpose in writing: special personal information as listed in section 26 of POPIA, such as health, biometric, religious or criminal information; personal information about a child, which section 34 restricts; identity numbers, bank and card details; passwords, keys and access tokens; client information covered by a confidentiality agreement; and unreleased financial results, legal advice, source code and trade secrets. In any tool, enter only the minimum personal information the task needs.
5. Security. The company treats each approved tool as part of its information systems under section 19 of POPIA: it lists the foreseeable risks, including wrong answers, over-shared files, and instructions hidden in documents the tool reads; keeps safeguards against them; checks them at each review; and updates them. Access follows each person's role. Before an assistant that searches company files is switched on, file sharing is checked, because such tools show users what their permissions already reach: Microsoft says Copilot surfaces organizational data to which individual users have at least view permissions.
6. Impact assessment. Regulation 4 of the POPIA regulations requires the information officer to ensure that a personal information impact assessment is done. As a company rule, it is done for each tool before the tool processes personal information, and records what information the tool sees, why, where it goes, and who can see the output.
7. Providers. Where a provider processes personal information for the company, a written contract under section 21 of POPIA binds it to the security measures of section 19 and to notify the company immediately of any suspected unauthorised access. The contract also names its own subcontractors, including the model provider, and confirms deletion at the end.
8. Transfers abroad. Before personal information is sent to a tool that processes it outside South Africa, [Role] records which ground in section 72 of POPIA applies, such as a binding agreement giving adequate protection, or the person's consent.
9. Telling people. Where AI processes personal information the company collects, such as a recorded and transcribed call, the company's notices tell people what is collected and why, in line with the duty to make them aware of "the purpose" under section 18.
10. Human review. AI output is a draft. A named person checks it before it is sent to anyone outside the company, published, filed, or used to decide anything about a person. Except as section 71(2) of POPIA allows, no decision with legal consequences for a person, or that affects them substantially, is based solely on automated processing that profiles them. Where the company relies on the Act's exception for safeguards in a contract, the person can make representations to a human and is told enough about the logic to do so.
11. Records. [Role] keeps a register of approved tools and what each is used for, each impact assessment, provider contract and transfer ground, and who has been trained. This supports the documentation the company must keep under section 17 of POPIA.
12. Incidents. Anyone who enters prohibited information, sees output that exposes personal information, or hears of a breach at a provider reports it to [Role] at once, without fear of blame for reporting. The information officer decides whether the Information Regulator and the people affected must be notified "as soon as reasonably possible" under section 22.
13. Training. No one uses an approved tool before completing training on this policy, and everyone repeats it every [year]. The sessions also serve the internal awareness duty in regulation 4 of the POPIA regulations.
14. Breaches of this policy. A breach is handled under the company's disciplinary code. Reporting your own mistake promptly counts in your favour.
15. Review. [Role] reviews this policy every [six months], and sooner when a tool is added, a provider changes its terms, the law changes or a national AI policy is adopted. Approved by [name and role] on [date]. Next review: [date].

This template is not legal advice. It was checked against the text of POPIA, its regulations and the Copyright Act on 12 October 2026. Have an attorney check it against your business and sector rules before you adopt it.

Ownership of AI output. The Copyright Act names an author for a computer-generated work: "the person by whom the arrangements necessary for the creation of the work were undertaken". The Act's text does not say when a prompt to an AI tool counts as those arrangements, or when AI output is original enough to be protected. So treat ownership of AI output as a matter for your contracts and your policy: state that output made for the company belongs to the company, and check that a tool's terms do not claim it.

Pending changes. The Copyright Amendment Bill would amend the Act, but it is not law: when the Constitutional Court ruled on the President's referral of it on 26 June 2026, the Bill had not yet been enacted, and we did not find it enacted when we checked.

Using other people's work. This guide does not answer whether putting someone else's material into an AI tool infringes their copyright. Our recommendation, as a policy choice: do not upload material the company has no right to copy, and do not publish output that closely reproduces an identifiable work.

Confidentiality. POPIA's confidentiality duty covers personal information. Your trade secrets, and what clients share with you under confidentiality agreements, need a rule of their own, and section four of the template is that rule: a choice, not a quotation of the law. What happens to what staff type is then decided by the provider's terms. Google's notice for its consumer Gemini apps says "a subset of chats are reviewed by human reviewers", asks users not to enter confidential information they would not want a reviewer to see, and says work accounts may have different terms. Business offerings often say the opposite by default: Microsoft says Copilot prompts and responses "aren't used to train foundation LLMs", Google treats Workspace prompts as customer data, and OpenAI says data sent to its API is not used for training unless the customer opts in. That gap is why section three of the template approves company accounts only.

1. List the AI tools staff already use, including free ones and AI features switched on inside existing software. Ask; do not assume.
2. Decide which to approve, and move that work onto company accounts.
3. Do the impact assessment and sign provider contracts for each approved tool that touches personal information.
4. Train everyone before access, and keep the attendance record.
5. Name the person who checks output for each use, and write the handover to a person into anything customer facing.
6. Diarise the review date.

Expect wrong answers. Anthropic's documentation says even the most advanced language models can sometimes generate text that is factually incorrect. The withdrawn national draft is the warning: as SAnews reported, the Minister said it "contains various fictitious sources in its reference list". The human review section of the template exists for exactly that.

IAMX delivers custom AI agents and copilots through a vetted network of specialist partners: agents that plug into your own tools, workflows and data, run on accounts in your company's name, with operator terms in writing, records kept and a handover to a person, taken from pilot into daily use, with one accountable lead who owns the outcome.

• Does a South African company have to have an AI use policy?

  No law we found names one. But POPIA requires security safeguards, written contracts with operators and records of processing, and its regulations require an impact assessment and internal awareness sessions. If staff use AI on personal information, a written policy is the simplest way to show you meet those duties.

• Does South Africa have a national AI policy?

  Not an adopted one. The Department of Communications and Digital Technologies published a policy framework, dated October 2024 (the draft policy calls it the framework of August 2024), and a draft policy for comment on 10 April 2026. The draft was announced withdrawn on 26 April 2026 and withdrawn in its entirety by a Gazette notice of 12 June 2026.

• Can staff use free ChatGPT or Gemini for work?

  That is your choice, not the law's, but we recommend against it for company information. Consumer terms differ from business terms: Google's consumer Gemini notice says some chats are reviewed by human reviewers. Approve company accounts whose terms you have read.

• Is an AI tool provider an operator under POPIA?

  Often, where it processes personal information for you under a contract and outside your direct authority. Then section 21 requires a written contract binding it to the Act's security measures. Whether a given provider is your operator depends on the arrangement, so ask an attorney about yours.

• Who owns content made with AI?

  The Copyright Act gives the authorship of a computer-generated work to the person who made the arrangements necessary for its creation, but it does not settle how that applies to AI prompts. Say in your policy and contracts that work made for the company belongs to the company.

• Can an AI tool decide on a customer's application by itself?

  Not where the decision has legal consequences or affects the person substantially and rests solely on automated profiling, unless an exception in section 71 applies; the exception that rests on safeguards requires a way for the person to make representations. Keep a person in the decision.

• How often should the policy be reviewed?

  The Act sets no interval. We suggest every six months, and whenever you add a tool, a provider changes its terms, or a national AI policy is adopted.

• South African Government (gov.za): The guide's legal points were checked on 12 October 2026 against the text of POPIA (sections 17 to 22, 26, 34, 55, 71 and 72), regulation 4 of the POPIA Regulations, 2018, section 1 of the Copyright Act, the Constitutional Court's summary of its judgment on the Copyright Amendment Bill, and the Department of Communications and Digital Technologies' Gazette notices 3880 and 3978 of 2026; no statute written for AI and no adopted national AI policy were found. Checked 12 October 2026.
• Department of Communications and Digital Technologies (copy held by OECD.AI): The South Africa National Artificial Intelligence Policy Framework of the Department of Communications and Digital Technologies is dated October 2024 and describes itself as a first step in developing the National AI Policy; it invited public comment. Checked 12 October 2026.
• Department of Communications and Digital Technologies, Government Gazette 54477: The Department of Communications and Digital Technologies' Draft South Africa National AI Policy says it is based on the South Africa National Artificial Intelligence Policy Framework of August 2024, the 32 submissions received on that Framework, and consultations through the Cabinet Cluster process. Checked 12 October 2026.
• Department of Communications and Digital Technologies, Government Gazette 54477: By Notice 3880 of 2026 in Government Gazette 54477 of 10 April 2026, the Minister of Communications and Digital Technologies published the Draft South Africa National Artificial Intelligence (AI) Policy, approved by Cabinet on 25 March 2026, for public comment within sixty days, ending 10 June 2026. Checked 12 October 2026.
• SAnews, Government Communication and Information System: On 26 April 2026 the Minister of Communications and Digital Technologies announced the withdrawal of the draft National AI Policy because it contains various fictitious sources in its reference list. Checked 9 October 2026.
• Department of Communications and Digital Technologies, Government Gazette 54840: By Notice 3978 of 2026 in Government Gazette 54840 of 12 June 2026, signed on 4 June 2026, the Minister of Communications and Digital Technologies withdrew Notice 3880, and the Draft South Africa National Artificial Intelligence (AI) Policy is withdrawn in its entirety from the date of publication. Checked 12 October 2026.
• Information Regulator, Government Gazette 42110 (copy published by National Treasury): Regulation 4(1)(b) of the Regulations relating to the Protection of Personal Information, 2018 (Government Notice R. 1383, Gazette 42110, 14 December 2018), as amended (the Information Regulator's amendment signed on 21 January 2025 deleted regulation 4(1)(c) and 4(2) and left 4(1)(b) unchanged), requires an information officer to ensure that a personal information impact assessment is done to ensure that adequate measures and standards exist to comply with the conditions for lawful processing. Checked 12 October 2026.
• Information Regulator, Government Gazette 42110 (copy published by National Treasury): Regulation 4(1)(e) of the Regulations relating to the Protection of Personal Information, 2018, as amended (the Information Regulator's amendment signed on 21 January 2025 left 4(1)(e) unchanged), requires an information officer to ensure that internal awareness sessions are conducted regarding the provisions of the Act, regulations made under it, codes of conduct, or information obtained from the Regulator. Checked 12 October 2026.
• Government of South Africa, Government Gazette: Section 19 of the Protection of Personal Information Act requires a responsible party to take appropriate, reasonable technical and organisational measures to secure personal information against loss, damage and unlawful access. Checked 9 October 2026.
• South African Government (gov.za): Section 19 of POPIA requires the responsible party to identify all reasonably foreseeable internal and external risks to personal information and to establish, verify and update safeguards against them. Checked 9 October 2026.
• South African Government (gov.za): Section 20 of POPIA says an operator must process personal information only with the knowledge or authorisation of the responsible party and must treat it as confidential. Checked 9 October 2026.
• South African Government (gov.za): Section 21(1) of POPIA requires a written contract between the responsible party and the operator that ensures the operator establishes and maintains the security measures in section 19. Checked 9 October 2026.
• South African Government (gov.za): Section 21(2) of POPIA says the operator must notify the responsible party immediately where there are reasonable grounds to believe personal information has been accessed or acquired by an unauthorised person. Checked 9 October 2026.
• South African Government (gov.za): Section 22 of POPIA requires the responsible party to notify the Information Regulator and the data subject of a security compromise as soon as reasonably possible after it is discovered. Checked 9 October 2026.
• South African Government (gov.za): Section 71 of POPIA bars decisions with legal consequences or substantial effect based solely on automated processing that profiles a person, with exceptions that must let the person make representations and receive enough information about the underlying logic. Checked 9 October 2026.
• South African Government (gov.za): Section 72 of POPIA bars transferring personal information to a third party in a foreign country unless a listed ground applies, such as a recipient bound by a law, binding corporate rules or binding agreement giving adequate protection, or the data subject's consent. Checked 9 October 2026.
• South African Government (gov.za): Section 17 of POPIA requires a responsible party to maintain the documentation of all processing operations under its responsibility as referred to in section 14 or 51 of the Promotion of Access to Information Act. Checked 12 October 2026.
• South African Government (gov.za): Section 55(1) of POPIA says an information officer's responsibilities include the encouragement of compliance by the body with the conditions for the lawful processing of personal information, and otherwise ensuring compliance with the Act. Checked 12 October 2026.
• South African Government (gov.za): Section 26 of POPIA prohibits, subject to section 27, processing personal information concerning religious or philosophical beliefs, race or ethnic origin, trade union membership, political persuasion, health or sex life, biometric information or criminal behaviour. Checked 12 October 2026.
• South African Government (gov.za): Section 34 of POPIA provides that a responsible party may, subject to section 35, not process personal information concerning a child. Checked 12 October 2026.
• Microsoft: Microsoft says Copilot only surfaces organizational data to which individual users have at least view permissions. Checked 9 October 2026.
• Department of Justice and Constitutional Development (Act 4 of 2013): POPIA section 18(1): when personal information is collected, the responsible party must take reasonably practicable steps to make the data subject aware of the information collected, the purpose, and any intended transfer to a third country, among other things. Checked 9 October 2026.
• Department of Trade, Industry and Competition: Section 1(1) of the Copyright Act 98 of 1978, paragraph (h) of the definition of author (added by Act 125 of 1992), says the author of a literary, dramatic, musical or artistic work or computer program which is computer-generated means the person by whom the arrangements necessary for the creation of the work were undertaken. Checked 12 October 2026.
• Constitutional Court of South Africa: The Constitutional Court's post-judgment media summary for CCT 306/24, [2026] ZACC 26, says judgment on the President's referral of the Copyright Amendment Bill and the Performers' Protection Amendment Bill was handed down on 26 June 2026, that the majority found sections 12D(1) to (5) of the Copyright Amendment Bill unconstitutional, and that the Bill has not yet been enacted. Checked 12 October 2026.
• Google: Google's Gemini Apps privacy notice says a subset of chats are reviewed by human reviewers, asks users not to enter confidential information they would not want a reviewer to see, and says use with a work or school account may be subject to different data handling terms. Checked 12 October 2026.
• Microsoft: Microsoft says prompts, responses and data accessed through Microsoft Graph are not used to train foundation LLMs, including those used by Microsoft Copilot. Checked 9 October 2026.
• Google: Google says Workspace user prompts are customer data under the Cloud Data Processing Addendum and that Workspace does not use customer data to train models without the customer's prior permission or instruction. Checked 9 October 2026.
• OpenAI: OpenAI says that as of 1 March 2023, data sent to the OpenAI API is not used to train or improve OpenAI models, unless the customer explicitly opts in to share data. Checked 9 October 2026.
• Anthropic: Anthropic's documentation says even the most advanced language models can sometimes generate text that is factually incorrect, and that its techniques to reduce this do not eliminate it entirely. Checked 9 October 2026.