Three Moves Into AI Work, Three Different Interviews

Read the fine print on Google's own AI Essentials page: 67% of workers who used gen AI estimated using AI saves them 2+ hours per week. Estimated. That single word is the whole interview. Google is selling a course and still won't call it a measurement — and yet "saved me a couple of hours a week" is the exact sentence sitting on thousands of résumés right now, including the four other people in your loop. It is the market average, dressed up as a personal result.

The second mistake is treating "the AI interview" as one thing. There are three moves people make into AI work, and each gets screened for something different. Prep the wrong one and you give a good answer to a question nobody asked.

The three moves, and what each panel is really testing

• Move 1 — same role, plus AI. You stay a dev, or stay an EM, and AI is now in your workflow. They already assume you can do the craft. What they are testing is judgment: did AI make you faster, or faster at shipping things you didn't check?
• Move 2 — pivot to an AI-adjacent role. AI product, solutions/forward-deployed, enablement, internal tooling. Nobody expects you to train a model. They are testing translation: can you turn a fuzzy business complaint into a scoped capability, with a cost attached?
• Move 3 — go deep as a specialist. Applied AI engineer, agent platform, evals. Here nobody cares that you use AI daily. They are testing tradeoffs and verification: why this design, and how you know it works.

Three moves, three interviews. Which also means the evidence that helps in one actively hurts in another.

Bring the evidence your move asks for

Move 1 — a measured before/after inside your own workflow. Not the certificate. Google AI Essentials is 5 modules and, by Google's own description, under 5 hours to complete. It is genuinely useful for vocabulary, and it is a bad thing to volunteer as proof — you have just told a hiring manager you spent one afternoon. Lead with a change you made to your own work; mention the course only if someone asks what you have studied.

Move 2 — a credential whose last domain is the job. The Google Cloud Generative AI Leader exam is 90 minutes, 50–60 multiple-choice questions, $99, no coding, and explicitly "for anyone in any job role, with or without hands-on technical experience". Its four domains end with business strategies for a successful gen AI solution — which is the move-2 interview, almost word for word. Prep from the exam guide's domain list and you are rehearsing the actual conversation. One warning for Brazilians: that exam is offered in Portuguese. The interview will not be.

Move 3 — the thing a certificate cannot be. AWS Certified AI Practitioner (AIF-C01) is 65 questions in 90 minutes for $100 USD, and AWS describes its audience as people "familiar with, but do not necessarily build, solutions using AI/ML technologies on AWS". Read that again: the vendor is telling you it does not certify building. Offer it as depth in a specialist loop and you have argued the other side's case for them. For move 3 the evidence is a system you shipped, plus the eval that told you it was actually working.

The five-line block that survives a follow-up

Every loop eventually lands on one number. Fill this in for one real thing you did, before anyone asks:

CLAIM:       <task> went from <before range> to <after>.
BASELINE:    how I know the "before" — I timed it / calendar blocks / file timestamps
AFTER:       median of <N> timed runs, INCLUDING the minutes I spend checking output
SAMPLE:      <N> runs across <M> weeks
NOT COUNTED: build time, the weekly review I still do, the runs I threw away

Four things make it hold up. Quote the "before" as a window, never a point — "somewhere between 40 and 55 minutes" reads like someone who was paying attention, while a clean round figure reads like someone who rounded. Put verification inside the after-number: if the run takes three minutes and you then spend twelve reading behind it, the honest figure is fifteen, and a good interviewer asks about review time anyway. State the sample size before they ask for it — five runs, admitted as five runs, beats a confident claim with nothing underneath. And name what you left out, out loud. Almost nobody does this, and it is the line that moves you from "candidate with a number" to "person who does the arithmetic properly".

Then rehearse the push. When they say the number sounds high, don't defend it and don't retreat from it — shrink it on purpose: "my 'before' is reconstructed from calendar blocks, not a stopwatch, so hold me to 40 minutes rather than 55. Even at that floor it's a few hours back every month, and I'd rather you judge me on the floor."

The practice-interview prompt

Paste this into Claude or ChatGPT. Its job is to be a worse audience than the real one:

You are a hiring manager at a US company, interviewing me for <role>. I would
work remotely from Brazil, UTC-3.

I am making move <1|2|3>: <one sentence>.
My evidence: <paste your CLAIM block, plus a link to the artifact>.

Ask 6 questions, one at a time, in English, and wait for each answer.
After every answer:
1. probe the weakest claim in it — especially where any number came from;
2. ask one follow-up I could only answer if I had really done the work;
3. score it 1-5 on specificity, provenance of the numbers, and whether I
   named a limitation myself.

Reject "I review it carefully", "it feels faster", and any recital of course
modules. Push once, then move on.
End with the three answers I should rewrite, and why.

Two rules for using it. Answer out loud, not in the chat box — typing hides the thing that actually breaks in a real screen, which is holding a technical argument in English while being interrupted. And run it before your first real interview. Burning the first version of your story on a company you want is expensive; burning it on a model costs nothing.

One question per move you should answer cold

• Move 1: "Tell me about a task you deliberately kept away from these tools." The boundary itself is the answer. A candidate who doesn't have one has just told me they have no review standard, only enthusiasm.
• Move 2: "This team wants a chatbot. Talk me out of it." Anthropic's engineering guidance is to "find the simplest solution possible, and only increasing complexity when needed" — adding that "this might mean not building agentic systems at all", because agentic systems "trade latency and cost for better task performance". The candidate who reaches for the smaller solution reads as senior. The one who reaches for an agent reads as someone who has never paid a latency bill.
• Move 3: "When is a workflow better than an agent?" Anthropic's split is clean: workflows orchestrate models and tools "through predefined code paths", while agents "dynamically direct their own processes and tool usage". Add complexity "only when it demonstrably improves outcomes" — and the obvious follow-up is how would you demonstrate it. That answer is your eval, and it is exactly where most move-3 candidates go quiet.

Do this today

Write your move in one sentence. Then fill in the five-line CLAIM block for exactly one thing you have already done — including the NOT COUNTED line — and say it aloud in English until it lands under 30 seconds without wobbling. That block, spoken cleanly, is worth more than the next certificate you were about to start.

Sources

• Google AI Essentials
• Google Cloud — Generative AI Leader certification
• AWS Certified AI Practitioner (AIF-C01)
• Anthropic Engineering — Building effective agents