A new question has entered the South African panel room — and most candidates are unprepared.
In 2026, South African interview panels — from graduate programmes at the big banks to mid-sized professional services and tech-adjacent firms — are increasingly asking some version of the same question: “how would you use AI in this role?” This is not hypothetical. This is the question. And most candidates are walking into the panel without having thought about it properly.
The question rarely arrives in the same words twice. But it tends to land in one of three patterns. Recognising which pattern you are being asked changes how you answer.
“Show me a prompt you’d use.” Or: “Walk us through a recent time you used AI for a piece of work.” The panel wants craft evidence — they want to see that you have actually used the tools and can describe the workflow concretely.
“Where would AI help in this kind of work?” Or: “If you joined the team, what’s the first AI workflow you’d propose?” The panel wants strategic thinking — can you reason about where AI fits in the actual work of the role, not just in general terms.
“Where would you NOT use AI?” Or: “What are the risks of leaning on these tools too much?” The panel wants judgement, not enthusiasm. This is often the question that separates the strong candidate from the merely keen one.
Employers are anxious about AI literacy across their workforce. They are not just hiring you for the role you applied for — they are hiring against an organisational concern about whether their teams are ready for the next three years. The candidate who can speak fluently about AI in their own work context lands as a small relief: one less person to retrain, one more person to learn from.
What they are testing is not whether you can write a perfect prompt. It is whether you can reason about AI in your work context — with specificity, with judgement, and with the verification habits of a professional rather than the credulity of an enthusiast.
This is not a trick question. It is an invitation. The candidates who handle it well treat it like any other competency question — with structure, specifics, and a sense of their own judgement underneath the tools.
A three-part structure for any answer to “how would you use AI in this role?”
A strong answer is not a list of tools and a tone of excitement. It is a three-part structure that demonstrates you have thought about the work, the workflow, and the safeguards. Most candidates skip the third part. Including it is what signals professional maturity.
Not generic. The weakest answers begin with “AI helps with productivity” or “AI saves time on admin”. Both are true and both are useless — they could have been said by anyone applying for any role. A strong use case is anchored in the actual work of the job description in front of you.
Compare: “I’d use AI for productivity” versus “I’d use AI to summarise client meeting transcripts before drafting follow-up emails.” The second tells the panel you have read the role, thought about a real workflow, and chosen one specific moment where the tool fits.
Input → prompt → output → what you do with it. Walk the panel through the sequence the way you would walk a colleague through it. Specifics signal craft. Vagueness signals that you have heard about AI more than you have used it.
The workflow does not need to be elaborate. A two-minute description of how you would feed the AI the right context, what you would ask it for, and how you would use the output is more impressive than a sweeping claim about transformation. Concrete beats grand every time.
This is the differentiator. Most candidates skip this step entirely. They describe a use case, they describe a workflow, and then they stop — as if the AI’s output were the finished work. The panel notices the gap immediately, even if they do not name it.
Including verification signals three things at once: that you understand AI is a draft and not a verdict, that you take ownership of the work, and that you have the professional habit of checking your sources. These are the same instincts that make a good junior in any field. Naming them in your answer separates you from the candidate who sounds like they are reading from a vendor brochure.
Specific. Anchored to the actual work of this role, not AI in general.
Owned. You did the thinking; the AI helped. The judgement is yours.
Verified. You named how you would check the output before using it.
Too generic. “AI helps with productivity” could be said in any room.
No verification step. Treats AI output as finished work.
Treats AI as magic. All enthusiasm, no judgement, no critical edge.
The AI question rewards craft over enthusiasm. Vague excitement loses to specific competence. Develop the craft of articulating a workflow you actually understand, and the question stops being threatening — it becomes one of the easier ones on the panel.
Accountant, marketer, engineer. The same question. What separates a thin answer from a strong one.
Reading the framework is the easy part. Hearing it in real role contexts is what makes it stick. Three sample interview moments — bad answer, why it fails, strong answer, why it works.
A panel for a graduate trainee accountant role at a mid-sized SA audit and advisory firm.
“I love AI. I use ChatGPT all the time for everything. I think AI is going to revolutionise accounting and the firms that don’t adopt it quickly are going to fall behind.”
Why it fails: no specific use case, no workflow, no verification, no judgement. Treats AI as a slogan rather than a tool. Could have been said by a school leaver. The panel learns nothing about whether this candidate can actually do the work of an accountant with AI in the mix.
“One concrete thing I’d use AI for is reviewing variance analyses. I’d ask Claude or ChatGPT to look at the trial balance changes month-over-month and flag movements that don’t have an obvious operational explanation. I wouldn’t trust the output without checking — I’d cross-reference the flagged items against the general ledger entries myself before raising anything with my senior. The AI saves me time on the scanning pass; the judgement about what matters is still mine.”
Why it works: specific use case (variance analysis), specific workflow (trial balance → AI scan → ledger cross-check), explicit verification (the candidate checks before escalating), and clear ownership of the judgement call. The panel hears a junior who would actually be useful on Monday morning.
A panel for a junior brand or campaign role at an SA agency or in-house marketing team.
“I’d ask ChatGPT to write me five campaign concepts based on the brief and then pick the strongest one to present back.”
Why it fails: no context input (the AI gets the brief and nothing else), no audience grounding, no critical evaluation, no iteration. Sounds like outsourcing the thinking. The panel hears someone who would hand the AI the brief and hand the AI’s output back to the client — with the candidate as a transparent intermediary rather than a thinker.
“I’d start by giving the AI the brief plus everything I know about the target audience — three or four paragraphs of context, including any prior campaign performance data the team has shared. Then I’d ask for three concept directions, each with the strategic rationale attached. Then I’d pressure-test each one — I’d ask the AI to argue against its own concept, identify what could go wrong, and name the assumption it’s making about the audience. By the time I’d done that, I’d have enough thinking to take into a real brainstorm with the team. The AI helps me think faster. It doesn’t replace the team conversation.”
Why it works: structured input (context first, not just the brief), multiple iterations, explicit pressure-testing, and an honest sense of what the AI is and is not for. The panel hears someone who treats AI as a thinking partner, not an answer machine — and who understands that creative work still happens between people.
A panel for a graduate developer or junior engineering role at an SA software or tech-adjacent firm.
“Yes, I use Copilot for pretty much everything. It’s much faster than writing code myself and the suggestions are usually good.”
Why it fails: no judgement about when AI helps versus hurts, no awareness of code-quality or security risk, no verification practice. Sounds like the candidate is one production incident away from learning a hard lesson. A senior engineer on the panel will hear this and worry about review burden.
“For boilerplate and well-understood patterns, yes — Copilot or Claude saves me time and the risk is low because I can read the output and catch issues immediately. For anything novel — new architecture decisions, security-sensitive code, or business logic I don’t fully understand — I’d write it myself, then maybe use AI to review my own work. The reason is simple: I’m responsible for the code. If it breaks in production, ‘the AI wrote it’ isn’t an answer. So I use AI where my ability to verify the output is strong, and I do the thinking myself where verification is hard.”
Why it works: explicit judgement about when AI helps versus hurts, clear ownership of the work, and a verification heuristic that any senior engineer would respect. The candidate sounds like someone who would be safe to put on the on-call rotation in twelve months, not someone who would be a liability.
The pattern across all three: specific, owned, verified. The role changes. The structure does not.
Three copy-pasteable prompts to use before your next interview. Then the brief, the waitlist, and a way to share this.
Open Claude, ChatGPT, or Gemini in a tab. Use these tonight if your interview is tomorrow. The point is not to memorise an answer — it is to develop the way you think about the question.
When to use: The night before your interview. Run it once with your honest first draft, then again after you’ve tightened the use case. Each pass usually sharpens the workflow.
The AI does not know your sector as well as you will after a few weeks in the role. Treat its suggestions for verification steps as a starting point — the panel will know more about the real risks than the model does.
When to use: After Prompt 1, once you have a draft. Verification is the differentiator — this prompt forces you to name the actual checks rather than gesture at them.
If the AI gives you a generic verification step (“double-check the output”), push back. Ask it what specifically you would check, in what order, against what source. Generic verification reads as no verification.
When to use: Last, after the answer feels solid. The panel may not ask follow-ups — but if they do, you want to have heard the toughest version of the question already.
The AI will sometimes be too generous. If the pushback feels soft, ask it to be tougher: “Argue against this the way a sceptical senior in this field would. Don’t soften it.”
Use these tonight. The point is not to memorise an answer — it is to develop the way you think about the question. The craft is in the answer, not the enthusiasm.
The full brief as a PDF you can read on your phone or print before the interview. Sent to your inbox.
No spam. Unsubscribe anytime. We’ll also send you occasional new material in this style — no more than once a month.
The full programme — 30 days of applied AI workflows for your career field — launches soon, R499 once-off. Tell us your target field and we’ll let you know when it’s ready.
No charge until the Sprint launches. We’ll only email you when it’s ready.
If someone you know has an interview coming up, this brief is more useful in their hands than in your tabs.
If you’ve been wondering whether you’re ready for the AI question, you’re not behind — you’re early. Most candidates aren’t preparing for this yet. The craft is in the answer, not in the enthusiasm. Use the prompts, build the muscle, and walk into the next panel having already thought it through.
A division of De Vlamingh & Associates Consulting CC
PositionMeAI develops capability — it does not promise outcomes.