Purpose
- Sparrows makes extensive use of AI tools — for drafting policy and correspondence, for research and analysis, and for building the software Sparrows itself runs on. This policy sets out the conditions under which that use happens, so that AI is a tool applied under human judgement and Board oversight, not a source of authority in its own right.
- This policy is the binding statement on AI use for everyone at Sparrows, including the Board's own responsibility for overseeing AI use across the organisation. It governs staff, volunteer and contractor conduct directly, and sets the standard the Board holds itself to in exercising that oversight.
Scope
- This policy applies to the Board, the Executive Director, all employees, and any volunteer, contractor or consultant using an AI tool on Sparrows' behalf or with Sparrows' information, whatever the tool and whatever the task — drafting, research, communications, data analysis, or building or maintaining Sparrows' own software.
Core Principle: AI Drafts, a Human Is Accountable
- Nothing produced with AI assistance — a policy, a report, a piece of correspondence, or a line of software — is treated as authoritative or relied upon until a named human with the relevant knowledge has reviewed and approved it. An AI tool's output is a draft, however confident or fluent it reads, until that review happens.
- Accountability for AI-assisted work sits with the person who approved it for use, the same as it would for work delegated to any other party. Using an AI tool to do something does not change who is answerable for the result.
- A policy or plan drafted with AI assistance still only takes effect once approved by the Board under Sparrows' own policy approval process; AI assistance in drafting is not a shortcut around that approval.
- The person who approves AI-assisted work under §3.1 checks every fact, figure, quotation, source and legal reference in it against the original, and removes any they cannot find. An AI tool can state a source, a figure or a legal provision that does not exist, and state it with confidence. The check applies in full to a return or submission to a government body, to the South African Revenue Service, to the auditor or to a funder.
- An AI tool does not decide anything about a person. A person decides, having read the information the decision rests on and not only the AI tool's output, and can explain the decision in their own words. This applies to a decision about an employee, a volunteer, a job applicant or a donor, and to a decision about a child, including admission, allocation, discipline and discharge. Section 71 of POPIA bars a decision that has legal consequences for a person, or affects them to a substantial degree, when it rests only on automated processing that profiles them. Reading the information behind the decision keeps the decision with a person.
- An AI tool is not used to assess, rank or give an opinion on a concern about a child's safety. The concern is reported under the Child Protection Policy §4.2, because the duty to report arises the moment a person concludes on reasonable grounds that a child has been abused, and putting the concern to a tool first can delay it.
Data Handling
- Before any of Sparrows' information is given to an AI tool — in a prompt, an uploaded document, or otherwise — the person doing so is responsible for knowing what that tool does with it: whether it is retained and whether it is used to train the tool's own models. The tool is assumed to send the information out of South Africa, which makes the transfer rules in the POPIA Data Protection Policy §7 apply, unless it runs on equipment in South Africa that Sparrows controls, or its provider confirms in writing that it processes information only in South Africa. Where a tool's data handling is unclear, the safer assumption is that the information is not kept private, and it is disclosed accordingly. The AI Provider Data-Handling Reference sets out what is currently known about the major providers and tiers in use or under consideration at Sparrows, to help apply this test in practice; it is a working aid kept current by the Executive Director, not itself part of this policy's binding text.
- A child's personal information is Sparrows' highest-risk category of data, and POPIA's own stricter protection for it (sections 34–35) applies in full to an AI tool exactly as it would to any other third party. Giving a child's personal information to an AI tool — in a prompt, an uploaded document, or a training or fine-tuning process — must fit one of the five closed grounds at the POPIA Data Protection Policy §5.1, the same test that governs disclosing that information to anyone else. Routine AI-assisted drafting of a policy, plan, or general correspondence that does not name or otherwise identify a specific child does not raise this question at all; it only arises once a specific child's identifying information is actually being disclosed to the tool.
- Whether a specific disclosure of a child's personal information to an AI tool meets one of those grounds is not left to individual judgement below the Executive Director. Where the answer is genuinely unclear, or where the disclosure is anything beyond a single, narrow, task-specific use, it is brought to the Board rather than decided alone. Where a task can be done just as well without disclosing a child's personal information at all, that is the version used.
- Financial, donor, and staff personal information carry real but lesser sensitivity than a child's information, and ordinary POPIA processing conditions apply to them (POPIA Data Protection Policy §4). The same "know what the tool does with it" test at §4.1 applies before any of this information is disclosed to an AI tool.
- A record created or substantially shaped by AI assistance is a record like any other for the purposes of the Records and Retention Policy — its retention period is set by what the record is (a policy, a financial record, an HR file), not by how it was produced.
- A person uses an AI tool for Sparrows' work only if the Executive Director has approved it, and then only through an account that belongs to Sparrows and not through a personal account. A personal account can keep what it is given and train the tool on it by default, and Sparrows has no contract with the provider for it.
- The Executive Director records each approved AI tool in the Operator Register, with the kinds of information it may be given and the account it runs on. A tool that is not recorded is given only information that is already public, so that nothing reaches a provider that nobody has checked.
- An AI feature inside software that Sparrows already uses, for example in email, in a design tool or in a document editor, is an AI tool. The Executive Director decides, feature by feature, whether it is switched on, and records the decision in the Operator Register against that software, because the feature can send what a person is working on to its provider.
- The website privacy notice says that Sparrows uses AI tools, what kinds of information they may be given, and that a person and not an AI tool decides about a person (§3.5). The Executive Director updates it when what is recorded under §4.7 changes, so that a donor or supporter can learn how their information is used. The Executive Director tells a member of staff or of the Board the same before giving any of their personal information to an AI tool.
- Before the Board decides under §4.3 or §6.2, the Executive Director gives it a written assessment of the use: what the AI tool or AI agent is for, what information it receives, where its provider processes it, what could go wrong for a child and what limits that, and how the use is stopped. The Board decides on the written assessment, so that the decision rests on written facts and can be explained later.
Use in Building and Maintaining Sparrows' Software
- Sparrows' internal software is built and maintained directly by the Executive Director using an AI tool, rather than commissioned from an external developer. This heading governs that practice specifically, in addition to §§3–4 above.
- Software that will hold personal information — particularly a child's information, or financial or payroll data — is reviewed for who can access it and what it stores before it goes into real use, not assumed safe because it works. Access to any such system is limited to the people who need it for their role, consistent with the access restrictions the POPIA Data Protection Policy already sets for the underlying information itself (for example, its restriction on CCTV access).
- A defect in AI-built software that risks exposing personal information, miscalculating a statutory obligation (such as payroll or a filing deadline), or otherwise causing a compliance failure is treated as seriously as any other safeguarding or compliance risk Sparrows carries, not as an ordinary bug to fix at convenience.
- Sparrows does not give an AI tool access to a live system holding a child's personal information as part of its ordinary build-and-maintain workflow. Development and testing use dummy or de-identified data; real data is only handled by the finished, reviewed system.
AI Assistants Used by Staff
- This heading applies to every staff AI assistant, in addition to §§3–4 above.
- A staff AI assistant is switched on only once the Board has approved it. The Board approves what it may do, and approves any later extension of that before the extension starts. It runs only on accounts and numbers that belong to Sparrows. Every member of staff and every member of the Board can use it. The Executive Director keeps the list of the people who can, and a person is removed from it on their last day as a member of staff or when they leave the Board.
- The Executive Director tells staff before they first use the assistant that it is an AI and not a person, what it records, and who can read what they send it. It says so itself if asked.
- Staff refer to a child by the first name the child is known by at Sparrows, because a child is called by name and not by a number. Staff do not send the assistant a child's surname, identity number, home or school address, photograph, medical or court reference number, or any other detail that identifies the child beyond that first name. The assistant does not ask for these. If a message contains one, it still deals with the message, does not repeat the detail in its reply, and reminds the staff member that only a child's first name may be used in the messaging app, because of the Protection of Personal Information Act. Giving a child's first name to the assistant is a disclosure of a child's personal information to an AI tool, which the Board decides under §4.3 when it approves the assistant under §6.2.
- The first time a staff member sends the assistant a detail barred by §6.4, the assistant reminds them of the rule, lists what is not allowed, and notifies them that a repeat will be reported to the Executive Director. The assistant records every such breach by category only, such as surname or identity number, with the sender and the time, and does not record the detail itself in the breach log. A breach counts for 90 days from its own date. A second or later breach by the same staff member within that time is reported to the Executive Director, who is notified by a summary once a day. The Executive Director decides whether a breach is assessed under the POPIA Data Protection Policy §11.
- Sending a message to the assistant is not reporting. A concern that must be reported under the Children's Act, or under Sparrows' own reporting procedures, is reported to a person as those procedures require, whatever else is said to the assistant. The assistant answers only from Sparrows' approved documents and says when it cannot answer. It does not decide, and it does not give legal, medical or disciplinary advice. It directs the staff member to the person or procedure that does.
- What the assistant records is stored in Sparrows' own records system, and not only in the messaging app or the AI tool. Each entry keeps the exact words sent, who sent them, and when. A summary or report the assistant produces from entries is a draft under §3.1 until a named person approves it, and it states which entries it draws on. A transcript of a voice note is machine-generated and is marked as such.
- The assistant does not carry what one staff member tells it into a conversation with another. Sparrows does not use what staff send it to train or tune any AI tool.
- The provider of the AI tool, and the messaging service that carries the messages, are operators under the POPIA Data Protection Policy §8. Each is entered in the Operator Register, with the §7.1 ground relied on, before the assistant is used.
- The assistant answers in the language the staff member writes in, including isiXhosa. Where it is unsure of a term, it says so. For a question about policy or a child's safety it quotes the approved source text, and a staff member can always ask for the English text.
- §§6.3 to 6.10 apply to a member of the Board who uses the assistant in the same way as they apply to a member of staff. The Executive Director tells a member of the Board what §6.3 requires before they first use it, and a breach of §6.4 by a member of the Board is logged and reported to the Executive Director under §6.5 in the same way as a breach by a member of staff.
AI Agents
- This heading applies to every AI agent, in addition to §§3–4 above. A staff AI assistant is governed by §6 and not by this heading.
- An AI agent is used only once the Executive Director has approved it under §4.6 and has recorded in the Operator Register each account and system it can reach and what it may do there. It does nothing outside what is recorded, so that a person reading the register can tell what it is able to touch.
- An AI agent has its own account or access key, so that what it does can be told apart from what a person does (IT Security and Acceptable Use Policy §3.4). Where it has to act through a person's account, that person agrees to it, and it acts only on that person's own mail and files. Its access is limited to what its task needs (IT Security and Acceptable Use Policy §3.1) and is removed when the task ends, so that access does not outlast the need for it.
- A person who may take the action for Sparrows approves each of the following before an AI agent takes it: sending a message to anyone other than staff or the Board; publishing anything; making a payment; deleting anything that cannot be restored; changing a credential, a permission or a security setting; and changing a record about a child. An instruction from that person to take a named action is that approval. The Executive Director may approve a named action that recurs, once, and records the approval in the Operator Register. These actions speak for Sparrows, move its money or cannot be undone, so a person must own each one before it happens.
- Giving an AI agent access to a mailbox, a drive, a system or a record that holds a child's personal information is a disclosure of that information to an AI tool, and is decided under §4.3.
- What an AI agent does on a system is recorded, in its own log or in the system's, so that the Executive Director can see what it did and when. How long the agent's own log is kept is set in the Records and Retention Policy §7.5.
- The Executive Director may stop any AI agent at once. A person who sees an AI agent act outside what is recorded under §7.2, or take an action in §7.4 without approval, reports it under §8.
- An AI agent that replies to a person does so only from a Sparrows policy or from a document about Sparrows that it holds in its repository, such as the Transfer Payment Agreement, unless a person directs it to reply otherwise. It says when it cannot answer from those. It tells the person in its first message that it is an AI, because the provider's usage policy requires an external-facing or interactive AI agent to say so. A message an AI agent sends is treated as Sparrows' own, and Sparrows is presumed bound by an agreement an AI agent forms, whether or not anyone reviewed it (Electronic Communications and Transactions Act, 2002, sections 20 and 25), so the agent is kept to what Sparrows has already decided.
AI Incidents
- A person who learns of any of the following reports it to the Executive Director immediately, because the harm grows for as long as the information or the output stays where it is:
- information given to an AI tool that §4 does not allow;
- an AI tool used for Sparrows' work that the Executive Director has not approved under §4.6;
- an AI agent acting outside what is recorded under §7.2, or taking an action in §7.4 without approval; or
- AI output that was wrong or harmful being relied on or published.
- The Executive Director stops the use that caused it, finds out what information or output was involved, takes down anything that was published, and decides whether it is a breach to be handled under the POPIA Data Protection Policy §11.
- The Board is told of an incident at its next meeting, or sooner where it is serious.
- A breach of §6.4 through a staff AI assistant is logged and reported under §6.5, and not under this heading.
Risk Register
- "AI-assisted work not adequately reviewed before reliance," "AI tool data-handling practices," "a defect in AI-built software," "an AI tool or AI agent used without approval, or acting beyond what was approved," and "a staff AI assistant receiving more of a child's identifying information than a first name or giving unsafe guidance" are standing entries on the risk register the Risk Management Policy establishes, reviewed on the same cycle as every other risk category, not treated as one-off items that close out once addressed.
Relationship to Other Policies
- This policy works alongside the POPIA Data Protection Policy (what may be disclosed, and on what basis), the Records and Retention Policy (how long an AI-assisted record is kept), and the Risk Management Policy (how the risks at §9 are tracked). Where this policy and one of those differ on a shared point, the more specific policy on that point governs.
Review
- This policy is reviewed annually by the Board. Given how quickly AI tools and their data-handling practices change, the Executive Director brings a proposed amendment to the Board sooner than the annual cycle if a real change in practice or risk makes that necessary.
- Before each annual review, the Executive Director gives the Board a written report of every AI tool and AI agent approved under §4.6, each change since the last report, every incident under §8 and every breach logged under §6.5, so that the Board oversees AI use on facts and not on assurance.