TL;DR. Selecting AI talent means finding one champion per function with the disposition to delegate and the design instinct to build workflows the team can trust. The right champion lifts a team and makes it visible, within sixty days, who will engage. They need room, budget, and visible standing to keep doing the work.

You are about to make a hiring decision, or a promotion decision, or a quiet decision about who to staff on the work that matters next year. AI is in the conversation. Either the role explicitly mentions it, or your boss asked, or a recruiter slid you a candidate with “GenAI” in their title.

This hiring, promotion, or staffing decision shapes the function. One kind of person lets it compound. The wrong one leaves you running training programs that produce no observable change while the people who could have rebuilt the function leave for a competitor that gave them air cover.

You can get this decision right. Most of the work is knowing what you’re actually selecting for.

What the resume can’t show

Open ten current job descriptions for senior roles in your industry. Count how many require “experience with GenAI tools,” “familiarity with prompt engineering,” or “Copilot certification.” Most of them. Now look at how the requirement is phrased. It’s in the same paragraph as “proficient in Microsoft Office.” That’s the level of seriousness most orgs are bringing to this.

The corresponding interview process matches. A candidate gets asked if they have used ChatGPT. They say yes. They’re asked to describe a time they used it. They describe summarizing a document. The interviewer nods. The box is checked.

This produces hires who pass the AI screen and behave, on the job, exactly like every other hire. They open the chat tool occasionally. They don’t change how the function operates. Six months later the team’s usage stats look fine, the output looks the same, and leadership can’t point to a single thing that’s now possible that wasn’t possible before.

Meanwhile, the one person on your existing team who would have transformed the function is sitting in a mid-level seat. They’ve already automated three pieces of their own job. They built a script their team relies on. They’re not on the slate for the next promotion because they don’t have the title that matches the JD. You didn’t hire them, because they weren’t on the market. You won’t promote them, because their wins don’t show up in the standard review template.

That’s the actual problem. AI hiring as currently practiced is selecting against the trait that matters.

Disposition matters more than tool fluency

The reason is the same one named in Recognizing Leverage. The consumption gap between AI power users and everyone else, better than tenfold by volume, comes from disposition and increasingly from design, more than knowledge. The Q3 2026 briefing puts this among the three highest-leverage moves leaders can make this year: find the people who design workflows, not just the people who work fast. The trait that produces leverage is a temperament: willingness to delegate work you used to own, willingness to treat the model as a collaborator instead of a search engine, willingness to read a wrong answer as a prompt to refine instead of evidence the tool is broken. Comfort being in the loop without being the bottleneck.

Tool fluency is a week of practice. The disposition is something closer to a personality trait. You can reliably select for it. You cannot reliably teach it.

There is a second axis the disposition frame alone misses, and it is the one that matters most this year. The champion who automates a piece of their own job is a power user. They make themselves faster. The champion who spots a recurring, checkable task and sets it to run on a schedule for a whole team is a workflow designer. They make a function faster without anyone else changing a habit. The first is leverage for one desk. The second is leverage for the org. Your filter now has to detect both: the disposition to delegate and the instinct to design workflows that run without you, not sessions you drive.

This means your hiring filter has to do two things your current process does not. Surface the disposition, which a resume can’t show and a behavioral interview rarely catches. And ignore the credentials that pretend to prove it. “Prompt engineer” was a burst of search interest, not a durable hiring category. Indeed searches rose from two per million to 144 per million in April 2023, then settled at 20 to 30 per million.¹ There is no published first-party time series that proves prompt-engineer postings fell. The defensible finding is title displacement: prompting survived as a skill inside broader roles, where it belongs.

The req has to name the job you actually need. Use AI engineer when the person will build, operate, and improve AI-enabled systems inside your company. LinkedIn put the title first in its 2026 Jobs on the Rise, with postings up 143% year over year. Use forward deployed engineer when success means embedding with a business unit or customer and turning its work into deployed systems. That title’s postings grew 729% year over year in an April 2026 Indeed snapshot, and Indeed’s current salary data puts its US average at about $172,000.² Robert Half’s 2026 guide also names agentic AI engineer/developer, AI strategy consultant, AIOps engineer, and LLM engineer as emerging title families.

AI agent architect, orchestration engineer, AI workflow architect, and AI workflow automation specialist are showing up in individual postings. The category is young. Use those labels as search terms, then hire against a clear charter. AgentOps describes a real operating problem, but not yet a labor-market category you can price confidently. The same is true of context engineer. Indeed counted 822 distinct US AI-touched job titles in the first quarter of 2026, 63 percent of them outside tech.² Title sprawl is why scope and evidence matter more than the label.

The same trap applies inside your organization. Whoever volunteered for the AI working group is not necessarily your champion. Champions are often quiet. They are usually shipping. They have rarely been given air cover, because the kind of work they are doing is not yet on the dashboard.

Champions leave visible evidence in the work

Concrete signals. These hold whether the person is a candidate, a current employee, or a referral.

A champion has already built something nobody asked for. A script that automates a piece of their own job. A prompt template their team uses. An internal tool, a dashboard, a checklist, a workflow document, a spreadsheet that does something the team used to ask BI for. They built it before there was a budget for it, before there was a working group, before anyone signed off. They built it because the friction was bothering them.

A champion keeps the chat open while they work. They do not open it like a search engine when they get stuck. It is one of the surfaces they think on. They will tell you what they have it set up for, what custom instructions they wrote, which model they switched to last month and why. The answer is specific and slightly opinionated.

A champion has a story about a wrong answer. They do not tell you the model is amazing. They tell you about the time it confidently invented a citation, and what they did next. That second half is the signal. They corrected, they refined, they kept going. They did not write the tool off. They calibrated their trust.

A champion is doing the work of someone who is no longer in the org chart. The marketer drafting in their own voice instead of briefing an agency. The analyst running the ad-hoc that used to take BI two weeks. The engineer closing tickets that belonged to three of them. They are not asking for headcount in a place they used to ask for it.

A champion has set up a scheduled task or recurring agent job that fires on a cadence and lands in other people’s inboxes whether or not the champion is at their desk. The Monday report that now writes itself. The reconciliation that runs overnight. Ask who benefits, then ask how the output is checked. A report that requires a careful senior read to tell whether it is right has moved a session onto the calendar. A workflow designer can name the reconciliation, threshold, or exception path that lets it run unattended. They are the highest-leverage person in this guide. They think in workflows, not sessions.

A champion is not necessarily senior. They are often not the loudest person on the team. They are rarely the one with “AI” in their job title. They are sometimes the person you almost passed over because their resume looked unremarkable for the level. They are almost always the person whose teammates, when asked, say “oh, you should talk to them about this.”

A consultant can bring the vocabulary. A champion has artifacts from their own work.

Four questions reveal the work a candidate has done

The standard AI-fluency interview misses the signal. “Have you used Copilot” is a yes-or-no question that screens for nothing. “Walk me through how you would use AI for this role” is a hypothetical that rewards vocabulary over experience.

Four questions reveal the difference.

One. “Walk me through the last thing you built or changed in your own workflow because of an AI tool. What does it do for you now. What was the version before it.” You’re listening for specificity. The answer should include a tool, a workflow detail, a before-and-after, and ideally a slight irritation in their voice about the part that still doesn’t work. If the answer is generic (“I use it to summarize emails”) they haven’t done the thing. If the answer is precise and slightly nerdy, you have a candidate.

Two. “Tell me about a time the model gave you a wrong or bad answer. What did you do.” You’re listening for the second half. The candidate who says “I stopped using it for that” is telling you they hit a single wall and bounced. The candidate who tells you how they reframed the prompt, or fed in an example, or switched models, or just kept iterating until it worked is telling you they have a working relationship with the tool. That is the trait.

Three. “If I gave you budget for tools, time, or hires in this role, what would you spend it on first. What would you stop doing because of it.” You’re listening for an opinion. A real champion has one. They’ve already thought about which seat is the wrong tier, which workflow is the next one to automate, which task they would gladly delegate. The candidate who turns this into “it depends on the strategy” is not in the cohort.

Four. “Tell me about a piece of recurring work you got a tool to do on its own. Who benefits from it now, just you or others.” Then ask: “How did you know it was right before people relied on it? What can two people check and agree on? What happens to an exception.” You’re separating the personal power user from the leverage-multiplier, then separating a real workflow designer from someone who has scheduled unverifiable work. The candidate who automated their own inbox is good. The candidate who set up a job that produces a report the whole team relies on, on a schedule, with them no longer in the loop, and can name its reconciliation or release check, is the workflow designer. If the answer is that a careful senior reader has to decide whether each output is good, it is assistance, not unattended work. Verification Is the Bottleneck explains why that distinction decides what can run.

Pair the four questions with one structured exercise. Give the candidate a real, non-confidential piece of work the role would actually involve. A messy data set, a draft document that needs analysis, a strategy memo that needs a counterargument, a pile of customer feedback that needs synthesis. Tell them they may use any AI tool they want. Give them ninety minutes. Watch what they do. The strongest signal is whether they choose a tool with intent, iterate, course-correct when an early pass is thin, and produce something better than a person without tools could in the same time.

Call the ninety-minute portion what it is: a test of the power user in a driven session. It cannot, by itself, tell you whether the person designs work that outlasts the session. At minute seventy-five, give them a second prompt. Ask them to name one part of the task they would put on a cadence, its trigger and audience, the check two people could mechanically agree is right, and the exception owner. Five lines is enough. “I would review it” is a session. The workflow designer names a test and where a failed test goes. Score this separately from the output. It is the part of the exercise that tests design.

Credentials hide more than they reveal. The AI bootcamps and short-form credentials that proliferated over the last two years are, with rare exceptions, a tax on the credentialed and a signal of nothing. “Prompt engineer” as a title, unless it names a broader engineering role with clear ownership. Portfolio pieces produced by an agency or a course. The work was done by someone else; the candidate was a client. Vendor logos. “Used Copilot at a Fortune 500” tells you they had a license. It does not tell you they opened it.

This filter will eliminate candidates who look strong on paper. A strong AI hire often looks like a generalist with a slightly weird side project. That is the candidate to hire.

The current team answers the same questions

The same filter applies internally, with one inversion. You are not selecting from a slate of candidates. You are looking for the one or two people per function who already meet the bar, and giving them explicit air cover to do more of what they are already doing.

One champion per function, surfaced and supported, will produce more downstream change than a quarterly training program for the whole team. They will rebuild a workflow other people then copy. They will show what is possible. They will quietly raise the standard of what gets shipped without you having to mandate it.

They will also make assessment of everyone else observable within about sixty days. Watch the diffusion. Once a champion is producing visible artifacts (a script, a prompt library, a workflow that closes a recurring task in a fraction of the previous time), the rest of the team sorts itself into three groups. The first group picks up the artifacts and adapts them. They are your second wave, and the next group worth funding. The second group asks the champion how to do the thing, takes the answer, and tries it once. They are your trainable middle. Most adoption work belongs here. The third group quietly avoids the champion. They do not adopt the artifacts. They do not ask the questions. When the work that used to be theirs becomes someone else’s faster output, they reframe it as a quality concern.

That third group is the assessment problem. A sixty-day record of the diffusion is more informative than a skills test.

Access and training will not make everyone AI-fluent

A meaningful fraction of your team is not going to develop AI fluency, regardless of training, time, or tools. Naming that out loud is the leadership job most leaders avoid, because it sounds like the wrong thing to say in a townhall.

This is a staffing decision. The AI-leveraged future of the function belongs with people who, by demonstrated behavior over a reasonable window, will produce it. Others can work where their experience compounds without that disposition, and advance on the work they do well.

The signals that someone is not in the cohort are quieter than the champion signals, and worth being honest about. They have been licensed for six months and still talk about AI in the future tense. They escalate small wrong answers as evidence of a categorical problem with the tool. They have not opened the chat in the last week, and when reminded, they explain why their work is the kind that does not benefit. They route around colleagues who are using it visibly. They reach for the previous version of the workflow when given the choice. None of these are character flaws. They are dispositional facts. Treat them as such.

The trap to avoid: confusing seniority with disposition in either direction. Some of your most senior people will turn out to be your strongest champions. Some will turn out to be the ones routing around the chat tool. Same for your most junior. The trait does not correlate with tenure in either direction, and assuming it does is one of the more expensive mistakes a leader can make right now.

Work patterns reveal champions more reliably than surveys

Champion signals. A person is a champion if you can observe at least three of these.

They have built something nobody asked for, in their own workflow, because of an AI tool.
They keep the chat open while they work, and can describe specifically how they have it configured.
They have a story about a wrong answer, and the story ends with them iterating, not bouncing.
They are doing work that used to belong to someone else, in volume.
Their teammates name them, unprompted, as the person to ask.
They have an opinion about the next tool, plan, or workflow change, and the opinion is specific.
They have set up something that runs without them: a scheduled task or recurring agent job that produces value for others on a cadence, with a mechanical check and an owner for exceptions. They think in workflows, not sessions.

Non-champion signals. A person is not in the cohort if more than one of these is consistently true after six months of access.

They talk about AI in the future tense.
They cite a single bad answer as a reason the tool does not work for their role.
They have not opened the tool in the last week, in a role where the tool plainly applies.
They route around the colleagues who are using it visibly.

These are observation questions. A sixty-day record of the work is more useful than a form.

A one-page talent map makes the next decision visible

Name a champion candidate in each function. Some functions will not have one yet, which is useful information. Where you can name them, get them on your calendar within the week. Give them three things in this order: explicit air cover to keep doing the work, a small budget to expand it (a better seat tier, an hour of their week officially carved out, a tool they’ve been wanting), and a public expectation that the rest of the team will adopt their artifacts.

This is retention work too. The champion you finally see is often the person the outside market sees next. Your Best AI User Is the Most Likely to Leave explains why compensation is the weakest lever here. The harder problem, more room, and visible standing to redesign the function are what make staying rational. Finding the champion and keeping them are the same problem.

Funding a workflow designer creates standing automation the organization now depends on. That is more consequential now that scheduled tasks are available across paid plans and can run remotely with the laptop closed.³ Inventory what they have built when you fund it, not after they leave. An unseen scheduled job can go dark when the designer departs or remain as an unowned compliance gap. It needs ownership, access, logs, checks, and shutdown. Managing Risk covers the controls; the Q3 2026 briefing covers why ungoverned scheduled automation became this quarter’s defining failure mode.

Where you can’t name a champion candidate in a function, the hiring brief is someone who has built things nobody asked for. The four questions and design prompt above belong in the next loop for that function, whether or not AI is in the job description.

The functions where you can name a champion will compound first. The ones without a candidate need a hire. That distinction, drawn across your org chart on a single page, is the AI talent strategy.

Once named, a champion needs a workflow, a budget, and a thirty-day expectation. Driving Adoption explains how to set those terms. The Q3 2026 briefing identifies the sequence, finding workflow designers and funding the work they design, as one of this year’s highest-return leadership moves.

Hannah Calhoon, Indeed’s VP of AI, on Indeed’s prompt-engineer search data. Last verified August 11, 2026. The figures describe searches, not a posting time series. ↩
LinkedIn, Jobs on the Rise 2026 (US); Refolk’s April 2026 Indeed snapshot for forward deployed engineers; Indeed’s forward deployed engineer salary data; Indeed Hiring Lab’s AI job-title taxonomy; and Robert Half’s 2026 technology salary guide. Last verified August 11, 2026. ↩ ↩²
Anthropic, “Schedule recurring tasks in Claude Cowork”. Last verified August 11, 2026. Local-file tasks are an exception: they require the connected computer to be available. ↩

Selecting Talent