The Funda Five

The Framework

Five questions. Any AI situation.

The Funda Five is introduced in Module 1 and applied in every module after that. By Module 6, it’s the way you think about every AI encounter at work. Select any step below to see what it means in practice.

Step 1 of 5

Task

Is this the right kind of job for AI? What exactly am I trying to accomplish?

Not every professional task benefits from AI assistance. Some tasks are well-suited — drafting, summarising, generating options. Others carry risks that outweigh the convenience — anything involving verification of facts, high-stakes decisions with individual consequences, or tasks where the process of thinking is itself the value.

Ask yourself: If the AI gets this wrong, what happens? Is the task low-stakes enough that a quick human scan is sufficient — or does it require rigorous human judgment at every step?

Step 2 of 5

Data

What information is involved? Is any of it personal, confidential, client data, or POPIA-protected?

When you paste text into a public AI tool, that text may be used to improve the model, stored, or accessible to third parties. In South Africa, the Protection of Personal Information Act (POPIA) creates specific obligations around how personal information is processed — including by AI tools. Many AI tools fall outside POPIA-compliant infrastructure.

Ask yourself: Does this involve client names, employee records, patient data, financial information, or confidential business information? If yes, is the AI tool I’m using approved for this data type?

Step 3 of 5

Tool

What type of AI fits this task? Is it approved by my organisation? Do I understand its limitations?

Different AI tools are built for different purposes. A language model is strong at generating and summarising text, weak at arithmetic and factual citation. An image model is strong at visuals, not at reasoning. Matching the tool to the task is a basic professional judgment — the same way you wouldn’t use a spreadsheet to write a contract.

Ask yourself: Is this tool approved by my organisation? Do I understand what it’s actually doing — not just what it produces? What does it specifically struggle with that is relevant to this task?

Step 4 of 5

Trust

How reliable is AI for this specific task? What are its failure modes here? How would I detect an error?

Trust in AI output should be calibrated — higher for tasks where errors are easy to spot, lower for tasks where errors are invisible until they cause harm. The key question isn’t “is this AI reliable?” — it’s “how would I detect an error in this specific output?” If you can’t answer that question, the verification step is incomplete.

Ask yourself: Can I independently verify the most important claims in this output? What’s the specific check I would run? What would an error look like — and would I be able to see it?

Step 5 of 5

Human

Who checks the output before it matters? Who is accountable if it is wrong?

When AI output affects a real person — their job application, their loan, their medical treatment, their legal standing — someone at the institution that deployed that AI is accountable. “The algorithm decided” is not a legal defence under South African law. The Employment Equity Act, POPIA, and consumer protection legislation all locate accountability in the organisation and its decision-makers.

Ask yourself: Who signs off on this before it reaches someone who could be affected by it? Is there a clear review step in place — or is the AI output going straight into action? If something goes wrong, who answers for it?

Walkthrough Example

The Funda Five on the DCDT case

Here's how the Funda Five would have caught the problem — before the policy was published.

Scenario: A policy analyst has used ChatGPT to help draft a section of a national AI policy. The AI produced text with three research citations supporting key recommendations. The analyst is about to submit it to the Director-General.

Task

Appropriate for AI? Yes, drafting policy language is a legitimate AI task. But citing academic research is not — AI cannot reliably retrieve and verify real sources.

Data

What information? Policy data is not POPIA-sensitive, but it is politically sensitive. This step doesn't flag an issue here — but it reminds the analyst to check what the AI was trained on.

Tool

Right tool? ChatGPT is strong at generating persuasive-sounding text. It is specifically weak at sourcing — it generates plausible-looking citations that may not exist. This is a known limitation of the tool type being used for citation-heavy work.

Trust

How to verify? Search each institution and journal cited. Check whether the specific study exists. Search for the authors. A 10-minute check would have found that none of the three sources in the DCDT document were real.

Human

Who is accountable? The policy analyst submitting the document. The Director-General signing off. The Department publishing it. Not ChatGPT. There needed to be a source-verification step in the review process — not just a content quality check.

Five questions. Any AI situation.

The Funda Five on the DCDT case

See the Funda Five in action.