How to Tell If a Job Candidate Is Truly AI-Native

Almost every candidate can now say, “I use AI every day.” That tells an employer very little. Access to ChatGPT, Claude, Codex, or another assistant is becoming ordinary; the important question is whether AI has changed how the person learns, works, and solves problems.

A good AI-era interview therefore tests behaviour, not vocabulary. It should reveal how a candidate handles unfamiliar work, decides what to delegate, notices a plausible mistake, protects sensitive information, and produces evidence that an outcome works.

Microsoft’s 2025 Work Trend Index used an “agent boss” frame for people who build, delegate to, and manage agents. The report drew on a survey of 31,000 knowledge workers across 31 countries plus Microsoft 365 and LinkedIn signals. It is a vendor-authored forecast, not a universal job description, but it captures the capability employers increasingly need to examine: directing AI while retaining human judgment.^[1]

1. Ask how they completed something they had never done before

A conventional interview rewards what a person already knows. An AI-native interview also examines how the person behaves when they do not know.

Tell me about an outcome you had never produced before but completed with AI. Where did you get stuck, what did you learn, and how did you know the result worked?

A strong answer names a concrete deliverable and reconstructs the path from uncertainty to evidence. The candidate can explain the missing knowledge, the sources or experiments they used, the corrections they made, and what they can now do without retracing every step.

A weak answer stops at “AI helped me.” It describes generated output but not the candidate’s decisions, tests, understanding, or responsibility.

2. Test whether they understand AI’s boundaries

Being AI-native does not mean trusting AI more. It means knowing that a fluent answer can still misunderstand the requirement, omit business context, introduce a subtle defect, expose private data, or optimise the wrong objective.

AI gives you a solution that looks 80% correct. How do you find what is wrong with the remaining 20%?

The NIST AI Risk Management Framework makes this operational: organisations should define human oversight, test systems in conditions similar to deployment, validate outputs in context, document limitations, and keep monitoring after release. Candidates do not need to recite the framework, but their method should show the same habits—context, verification, escalation, and accountability.^[2]

Listen for risk-based verification. A candidate might reproduce the result independently, compare it with an authoritative source, add tests, inspect edge cases, ask a domain expert, or require approval before a consequential action. “I ask the model if it is sure” is not independent verification.

3. Ask what the human should do and what AI should do

Give the candidate a complete but realistic scenario: launch a small feature in two weeks, including customer research, a specification, a prototype, testing, documentation, and a launch plan.

How would you divide this work between yourself, other people, and AI—and where would you keep a human approval gate?

AI may be useful for gathering an initial market map, clustering feedback, proposing approaches, drafting a specification, generating a prototype, producing test cases, and preparing documentation.

The human should still decide which customer problem matters, conduct or interpret important conversations, identify missing context, make trade-offs, judge the prototype, protect sensitive data, and approve decisions with meaningful consequences.

The best answer is not the one that delegates the most. It is the one that assigns work deliberately according to context, risk, reversibility, and the evidence required.

4. Look for outcomes, not AI vocabulary

Agents, retrieval-augmented generation (RAG), Model Context Protocol (MCP), embeddings, context windows, and reasoning models can all be relevant. None of those terms proves that a candidate can create value.

What useful result can you now produce that would have been difficult for you before AI?

A useful answer might involve deploying an unfamiliar application, analysing a large body of customer feedback, automating a recurring process, creating and testing a campaign, or launching a small business experiment. The scale matters less than the completeness of the loop: a real need, a delivered result, and evidence from use.

For a broader capability model, see What Is AI-Native Talent—and How Do You Become AI-Native?

5. Find out how they recover when AI is wrong

Tell me about a time AI produced a bad result. What caused the failure, and what did you change?

A strong answer diagnoses the failure. Perhaps the goal was vague, context was missing, the task was too broad, the source was weak, the tool lacked permission, the test covered only the happy path, or the model was unsuitable for the work.

Then the candidate improves the system: clarify the specification, add examples, reduce the scope, retrieve authoritative data, introduce deterministic code, add automated tests, change the model, or create a human approval step.

Changing prompts at random until one output looks acceptable is experimentation, but it is not yet reliable supervision.

6. Test whether they can redesign an old workflow

What is something you used to do manually that you no longer believe deserves the same amount of human time?

The answer is less important than the reasoning. Strong candidates continually update their view of what people and machines should each do. They preserve the parts that require judgment, trust, human contact, or professional responsibility and redesign the rest around current capability.

If you designed this workflow from scratch today, what would stay human, what would become automated, and what new failure mode would you watch?

7. Give them a small unfamiliar task

A short work sample can reveal more than a long quiz about memorised tools. Give the candidate a problem they have not seen before, allow reasonable AI use, and make the evaluation criteria explicit.

You have not used this API before. You have 30 minutes. Show how far you can get, what you verified, and what you would do next.

Observe whether the candidate frames the goal, asks focused questions, finds authoritative documentation, tests assumptions, notices when the AI drifts, recovers from errors, and communicates uncertainty. Do not score only the final output; the process is the evidence.

Keep the task relevant to the role, accessible, time-bounded, and consistent across candidates. Do not use AI fluency as a proxy for confidence, accent, personality, age, disability, or familiarity with one branded product.

8. If the job uses AI, let the interview reveal AI use

A company cannot learn how someone collaborates with AI by banning AI from every practical exercise. Foundational knowledge can still be tested separately, especially where the role carries safety, security, legal, or technical duties.

For the applied task, ask candidates to disclose what they delegated, preserve relevant prompts or artefacts where appropriate, explain the checks they performed, and identify what they would not trust without further review.

Provide a clean interview environment and clear data rules. Candidates should not be encouraged to paste confidential company information, personal data, or another employer’s proprietary material into a public model.

A practical AI-native candidate scorecard

Score the evidence under six headings rather than asking whether the candidate “seems good at AI”.

• Outcome: Did they define and complete something useful, or only generate fragments?
• Context: Did they give AI the goal, constraints, audience, resources, and definition of done?
• Judgment: Could they choose between options and explain the trade-offs?
• Verification: Did they test important claims, behaviours, edge cases, and sources?
• Learning: Could they explain what they learned and transfer it to a new situation?
• Ownership: Did they protect boundaries, escalate appropriately, and accept responsibility for the result?

Weight the headings by role. A security engineer needs deeper independent verification; a marketer may need stronger audience judgment and experiment design; a customer-support operator needs careful policy boundaries and escalation. One generic AI test should not replace role expertise.

Four hiring mistakes to avoid

1. Rewarding the longest tool list instead of evidence of a complete outcome.
2. Rewarding speed without checking correctness, privacy, maintainability, or customer value.
3. Banning AI in every exercise while claiming the role is AI-native.
4. Treating “AI-native” as a personality type or permanent label rather than an observable working practice.

The new hiring question

Employers still need to ask what a person knows. AI adds a second question that deserves equal attention.

What can this person accomplish when they do not already know how to do it—and can they produce that result without surrendering judgment?

The strongest candidate understands the goal, collaborates with AI, tests the output, learns what is missing, communicates limits, and keeps moving until the result is useful. That is more durable than expertise in any one interface.