AI-fluent virtual assistants: how to screen for it when hiring

AI-fluent virtual assistants: how to screen for it when hiringAn AI-fluent virtual assistant is not someone who uses ChatGPT. It is someone who can tell which parts of a task are safe to automate, give a model enough context to be useful, verify the output before it reaches you, and document the method so the improvement stays in your business. Screen for it with a short practical exercise containing a deliberate inconsistency, not with a question about which tools they know.Key takeaways

Tool lists on a CV tell you nothing; process reasoning tells you everything

Verification instinct is the single strongest predictor of a good hire

A twenty minute practical exercise beats three interviews

Ask what they would refuse to hand to a model, and why

Every assistant CV now mentions AI. Most of the time it means the candidate has used a chatbot to draft an email, which is a useful habit and not the thing that makes a measurable difference to your week.The difference between using AI and being fluent with it is the same as the difference between using a spreadsheet and being able to model with one. Fluency is a way of approaching work, and it shows up in specific, testable behaviours.The four behaviours that matterNotice that only one of these is about the model. Fluency is mostly process discipline, which is why it correlates far better with strong operational experience than with time spent on AI tooling.

Decomposition: breaking a task into the part that needs judgement and the part that does not

Context: giving a model the source material, the constraints and the intended reader, not a one-line prompt

Verification: checking outputs against a source before anything is sent, by default

Documentation: writing down the working method so it survives them leaving

The twenty minute exerciseGive the candidate a short brief and two source documents that contradict each other in one detail, for example a figure that differs between a spreadsheet and a summary email. Ask for a one page summary and a recommendation.What you are watching for is whether they notice the inconsistency, whether they flag it rather than quietly average it away, and whether they can explain which parts of the task they would automate if it recurred weekly. A candidate who returns a fluent, confident, wrong summary has just told you something important.Questions worth asking, and the answers to listen forAsk what they would never hand to a model in this task. Fluent candidates answer immediately and specifically, usually naming anything client facing, anything with a legal or financial consequence, and anything where the tone matters more than the content.Ask how they check output they cannot verify themselves. The good answer involves going back to a primary source or escalating, not asking the model whether it is sure.Ask what they would document. The best candidates describe a written procedure with the trigger, the steps, the checkpoints and the failure cases, because that is how they already work.Red flags

A long tool list with no example of a process they changed

Talking about prompts rather than about outputs and checks

No instinct to verify, or verification described as an occasional extra step

Unable to name anything they would keep away from a model

No documentation habit, so the improvement leaves when they do

Why this matters more than the toolingTooling turns over roughly every eighteen months. An assistant who has memorised one platform needs retraining each cycle; one who understands process design, verification and documentation absorbs whatever the tooling becomes.This is the whole reason Alvara hires for judgement first and trains on the stack second. Fewer than 1% of applicants are selected, and the exercise above is close to what they sit.A scoring rubric you can useScore the exercise against behaviours, not polish. A candidate who produces a beautiful summary but misses the contradictory figure has demonstrated less fluency than someone who returns a rougher document with the conflict clearly flagged and a sensible next step.Keep the test close to the real roleThe exercise should resemble the work the person will actually own. For a founder's assistant that might be a messy meeting brief and follow-up list. For an investment team it could be a diligence summary. For a back office remit it could be an invoice register with missing fields.Do not ask for confidential data or production access. A redacted, fabricated version of a real process is enough to reveal the habits you need, and gives the candidate a fair test without creating an avoidable data risk.FAQWhat does AI-fluent mean for a virtual assistant? It means they can identify which parts of a task are safe to automate, brief a model properly, verify the output against a source, and document the resulting method so the business keeps it.Can I train an existing assistant to be AI-fluent? Often, yes. Strong operational assistants with good verification habits pick it up quickly. The harder gap to close is process discipline, not tool familiarity.Which AI tools should an assistant know? Broad familiarity with a general assistant such as ChatGPT or Claude, an automation platform such as Zapier, Make or n8n, and whatever CRM and finance tooling you already run. Depth in your stack matters more than breadth across the market.How do I test AI fluency without a technical background? Use the contradictory-documents exercise. You do not need technical knowledge to see whether someone caught the inconsistency and explained their checking process.More insights | Alvara Partners