Ask the candidate to share their screen and open something they built with AI that is still running. A builder opens it quickly and can tell you what it costs to run and when it last went wrong. If all you get is a chat history and some prompts, you've learned what you needed.
"Uses AI" covers a huge range of habits. Gallup's workforce survey of 22,368 US workers, reported by the AP in January, found 12% use AI daily in their job and roughly a quarter use it at least a few times a week. Among technology workers it was about 6 in 10 using it frequently and about 3 in 10 daily. And about 6 in 10 of the people using AI at work said they use chatbots or virtual assistants. So daily use is still a minority habit, and chat is the most common way in.
For plenty of jobs a chat window is all anyone needs. When the job is building your pipeline, you're paying for more than that.
The posts ask for building, and mostly for engineers
In the job posts we've collected since 3 May, AI agents or LLMs came up in 26% of the 241 GTM engineer posts and 42% of the 619 AI and automation engineer posts. For the 668 RevOps posts it was 6%.
Those are floors, because a post only counts when the term shows up in its title, summary or stack. Still, the split is the useful part. If you're hiring a GTM or automation engineer, AI building is part of the job description and you should test it hard. If you're hiring a RevOps manager to fix forecasting and the CRM, it's a nice extra, and you shouldn't fail an excellent ops person because they haven't wired an agent into anything yet.
Pick one build and keep asking
In a short piece on screening people who fake AI expertise, Christopher Penn's point is that an AI tool can't answer specific questions about someone's real, non-public work, so you ask for specifics and keep pushing. Vague answers that sound right but have no detail in them are the tell.
So I pick one thing they say they built and stay on it for fifteen minutes. These are the questions I'd copy.
- What triggers it, and what does it run on?
- Roughly what did it cost last month in credits or API calls?
- What's the last thing it got wrong, and how did you find out?
- Which step did you take away from the model because plain logic was cheaper or more reliable?
- Where does the prompt live, and who else has edited it?
- If you left tomorrow, who could fix it?
Good answers tend to be boring. They know which model they picked and roughly why. They know the prompt sits in a Clay column or an n8n node and they've changed it a dozen times. They've caught it classifying a competitor as a hot lead, and they now spot-check a sample every week. My favorite sign is a candidate who moved a step off the LLM and back to a filter or a lookup, because that person has been paying the bill and watching the output.
Weak ones sound like a conference talk, with lots of "agents" and "workflows", every tool on the market named, nothing they can open, and no idea what it costs. A library of saved prompts doesn't count as a build, however long it is.
Run a 30 minute build with AI allowed
Then give them a small live task on a copy of your own data and tell them to use whatever AI tools they normally would. Karat says that banning AI in an interview measures a narrow slice of ability, and that you should judge how candidates test their work and decide when to trust what the model gives them. For a GTM builder, that judgment is most of the job.
A task I like is to hand over 20 real accounts and your ICP notes, and ask them to set up a step that scores each account and writes one line on why. Then ask them to show you three it got wrong.
Watch what they do before they trust it. A builder reads a handful of outputs first, tightens the prompt, asks for structured output so the result can go into a field, and notices the two accounts where the website didn't load. Someone who only uses chat will paste the list into a window, copy the answer back and call it done. You might get a similar-looking sheet from each of them after 30 minutes. My bet is on the first one when your whole database is going through it in month three.
If you'd rather do a take-home, keep it short. Yardstick suggests two to three hours and scoring the live conversation about the work, not the file they send back. I agree with the second half more than the first. The walkthrough is where you learn whether they built it. Our guide to hiring a GTM engineer has a longer paid version of this test.
When a chat window is enough
Not every GTM hire needs to build. A VP of Sales who uses AI every day to prep for calls and pull themes out of lost-deal notes is using it well. Grilling them about API costs tells you nothing useful about whether they can run a sales team. Ask them to show you their prep for a real call instead, and judge whether it saved them time or just added a step.
Be careful about the opposite mistake too. My read is that a very polished, obviously AI-assisted application tells you almost nothing about whether someone builds, so don't reward it or punish it.
If you'd rather we find the GTM engineer, a direct hire is 20% of first-year base salary, paid when they start, with a free replacement if they leave or it isn't working in the first 90 days.
Before the first interview, write down the one build you want this person to own in their first month. Then ask every candidate to open the closest thing they've already made. The ones who can't open anything get one more question, and then you move on.