Explain a complex AI concept to a non-technical stakeholder

[INTERVIEWER] Explain a complex AI concept to a non-technical stakeholder. This one's different from the other behavioural questions, and people miss that. "Explain a complex AI concept to a non technical stakeholder" is not a story you tell. It's a live test. They're going to watch you teach, right there in the room. Pick a clean analogy, strip the jargon, and check the person actually followed you.

The trap is showing off how much you know. The skill is making the other person understand. They focus on this because half the AI PM job is acting as the translation layer between research and the business. The scientists talk in embeddings and context windows. The VP of Sales needs to make a call by Friday. If you can't stand in the middle and make one side legible to the other, you can't do the job.

In the next few minutes I'll give you the four moves that make an explanation land, then a full worked demo, then the follow up questions. Four moves, and each one prevents a specific mistake. Move one: anchor to what they care about. About thirty seconds. Start from the problem the stakeholder has, not the concept. Not "let me explain retrieval", but "you asked why the assistant sometimes cites the wrong policy, so let me show you how it finds answers." You've earned their attention by starting where their pain is.

Move two: one analogy, chosen well. About ninety seconds. Pick an everyday thing that shares the concept shape. Say the analogy, then map each part back to the real thing, once. And here's the discipline: do not stack three analogies. One that genuinely fits beats three that half fit and leave them more confused than when you started. Move three: strip the jargon.

About ninety seconds. Say the plain English version of each term. If you absolutely must use a word like "embedding", define it in the same breath, in normal words. No acronym gets to sit in the air untranslated. The moment you say "vector similarity" to a sales VP, you've lost them and you won't know it. Move four: check for understanding.

About a minute. Never assume they got it. Ask them to say it back, or ask a question that reveals whether it landed. Something like "so if we add a new policy doc, what has to happen before the assistant can use it?" Their answer tells you whether you taught or just talked. And then close by tying it to the decision they actually have to make.

Teaching with no payoff is just a lecture. Let me show you all four moves in one go, explaining RAG to a sales VP. "You asked why the sales assistant sometimes quotes an outdated discount policy. Here's how it works. Think of the model like a smart new hire. Brilliant writer, but they haven't memorised any of our policies yet.

Retrieval augmented generation, RAG, means we never ask them to answer from memory. It's an open book exam. When a rep asks a question, the system first looks up the most relevant pages from our policy library, hands those pages to the model, and only then asks it to write the answer using what it just read. There are two separate jobs here: the looking up, and the writing.

When it quotes an old discount, that's almost always the looking up part fetching a stale document. It's not the model making things up. The fix is in our library, not the model brain. We keep the docs current and pull the retired ones. Let me just check I explained that well. If Legal updates the refund policy tomorrow, what has to happen before the assistant gives the new answer?" And the VP says: "You'd have to put the new doc in the library and take the old one out."

"Exactly right. Which is exactly why I'm asking for a doc owner on your team, so the open book stays current." Look at what that did. It anchored to his problem. One analogy, the open book exam, mapped cleanly and then dropped. Every term in plain words. A real comprehension check. And it ended on the decision he needed to make, assign a doc owner.

That's the whole technique in ninety seconds. If you want a second concept to keep in your pocket, here's fine tuning versus RAG for the same audience. "RAG is the open book, we hand it the pages. Fine tuning is more like sending the new hire on a training course, so the knowledge is baked into how they think. Open book is faster to update and easy to keep current.

The training course takes more effort and you redo it whenever things change, but the hire gets naturally better at the work style. For our policies, which change monthly, the open book wins." Same shape, same discipline, one analogy, mapped and dropped. Here's what makes them lean in. First, one analogy, mapped cleanly, then dropped, with no jargon left sitting undefined.

That's rare and it reads as genuine command of the material. Second, a real comprehension check, and you used their answer to confirm it landed rather than plough on regardless. And third, you tied the concept to a decision, so the teaching had a point. That's the PM signal: you don't explain for its own sake, you explain to unblock a call.

Now the ways people fail. The first is showing off. Reaching for vector similarity and context window in front of a sales VP, because it feels impressive. It isn't. It's the exact opposite of the skill being tested. The second trap is three competing analogies that muddy each other. You think you're being generous with options, but you've just made it harder.

Pick the one that fits and commit. And the third is no check for understanding at all, so you sail on and never find out you lost them at minute one. Expect the interviewer to play the stakeholder. They might frown and ask if the model is wrong or the document is wrong. Don't retreat into jargon. Stay in the analogy and say the document is outdated, the model just read what we gave it.

If you can hold the plain language under pressure, that's the whole test passed. The whole picture: anchor to their problem. One analogy, mapped and dropped. Every term in plain English. Then make them say it back, and land it on the decision they have to make. Four moves. The last one is the check, which everybody skips but actually proves you taught.

One line to carry in: one analogy, every term in plain words, then make them say it back. If they can explain the next step to you, you taught it. If they can't, you just talked.