How Do I Check AI Skills If I'm Not an AI Expert?

For an ordinary office role, start with work you understand: a customer reply, a schedule, a summary or a checked calculation. Ask the applicant to inspect a flawed AI-style draft against supplied facts, correct it and explain what still needs approval.

You do not need to judge impressive-sounding prompts. You need to judge whether the person can deliver an accurate, appropriate result and recognise when not to trust a tool.

This approach is for using AI in a job, not building AI systems. If you are hiring an AI engineer, evaluating a high-stakes automation or assessing work you cannot judge yourself, involve a qualified subject expert. A short office exercise is not a substitute.

Decide whether AI experience is needed on day one

“Must know AI” is too vague to assess. Name the actual task and what the person must already be able to do. If you intend to teach the tool, do not turn unfamiliarity with one interface into a hidden rejection rule.

OPM's work-sample guidance recommends tasks resembling the job and cautions that work samples may be inappropriate for skills you plan to teach after selection. The exercise below is an editorial example, not an OPM-endorsed or validated hiring test.

A ready-to-adapt customer-email exercise

Everything in this example is fictional. The flawed draft is author-written to illustrate common checking problems; it is not an observed output from an AI model.

Use this only if checking and writing customer updates is genuinely part of the job. Give every candidate the same packet and tell them what you are evaluating. Pilot it with someone who understands the work so you can set a reasonable time allowance and clear instructions.

Source facts supplied to the candidate

• A customer ordered six replacement filters at $18 each. Delivery costs $12 for the order. Ignore tax for this exercise.
• Four filters are in stock. Two are expected next Tuesday, but the supplier has not confirmed that date.
• You can offer to send the four available filters first. The customer must agree before the order is split.
• Discounts and refunds require the owner's approval.
• There is no confirmed delivery date for the full order.

Draft to check

Your six filters are ready to ship today. The total is $108 including delivery. The remaining items will definitely reach you next Tuesday. We have applied a 10% discount for the delay and will send the first four now.

Candidate instructions

Check this draft against the source facts. Write a corrected customer message, list the changes you made and identify any action or approval still needed. Do not send a message or invent missing information. You may refer to the packet throughout.

You can run this checking version without any AI account. It assesses verification and communication, not the candidate's ability to operate an AI tool. Be explicit about that limitation rather than calling it a complete AI proficiency test.

Your review key

The order total is $120: six times $18, plus $12 delivery. Only four items are available now. Tuesday is an unconfirmed supplier expectation, not a promised customer delivery. The discount has not been authorised, and a split shipment needs the customer's agreement.

One acceptable correction could be:

Four of your six filters are currently available. The order total is $120 including delivery. We expect the remaining two next Tuesday, but are still waiting for supplier confirmation and cannot yet confirm delivery of the complete order. Would you like us to send the four available filters first? We will wait for your confirmation before splitting the shipment.

That is an example, not the only acceptable wording. The candidate could also flag the discount request for the owner and identify the supplier confirmation as a follow-up. Do not penalise a different writing style when the facts, boundaries and customer choice remain clear.

Use a simple evidence record instead of an unexplained score:

Record “shown,” “partly shown” or “not shown,” with a short example from the work. This rubric is a discussion aid, not a validated cutoff or a reason for automatic rejection. Agree in advance which mistakes require clarification and use the same follow-up approach for comparable gaps.

If the job really requires operating the tool, add a separate step

Provide the same approved tool and access conditions, explain permitted use, and use fictional inputs again. Do not require applicants to buy a subscription or open personal accounts merely to participate.

Ask them to draft a response, check it against the packet and hand over the final version with any unresolved questions. You are now observing both tool use and checking. Model outputs can differ, so judge the person's decisions and final work against the source facts—not whether their first draft matches someone else's.

Tell candidates how to request an accessible format or accommodation. Keep the exercise short and job-related; do not use it to obtain unpaid productive work for your business. The appropriate assessment process depends on the role and your applicable requirements.

Ask one question about information handling

Use a hypothetical situation rather than asking for a former employer's confidential prompt history:

You have a real customer complaint containing personal details. Before putting it into an AI tool, what would you check?

A useful answer recognises that the employer's approved tools and data-handling rules matter. The candidate should not assume that removing a name makes every document safe to upload. Ask how they would proceed if permission is unclear; pausing and escalating can be the right action.

Do not ask candidates to demonstrate with real customer, employee or applicant records. Agree any live-tool assessment setup with whoever owns privacy and security in your organisation.

What the result does not prove

One corrected message does not establish broad AI expertise, reliable performance across a job or the ability to design safe automation. It gives you a concrete sample to discuss alongside other job-related evidence.

On a resume, “used AI to improve productivity” is still a claim needing context. Ask what the person produced, what they checked, what went wrong and which decisions stayed human. Resume Ranker can help locate stated experience in PDFs; it cannot verify AI proficiency from those words.

Fit the exercise into a short, purposeful hiring process, not another round added simply because AI is fashionable. The practical question is whether the person can use the available support and still take responsibility for the work.