The Conversations You're Not Having
Every agency I work with has three or four people who are excellent with AI. The names come up quickly when you ask. "Oh, she's building things nobody else understands." "He's got Claude integrated into our whole briefing process."
Those people are your real AI capability. The tools are replaceable. They're not.
And most of them are actively considering leaving.
What I Hear From Them
When I run sessions inside agencies, I always spend time with the power users separately. Not to promote them. To understand why they feel the way they do.
The pattern is consistent. They're tired. Not of the work — of the lack of support for the work.
They're pushing AI forward on their own. They're building custom workflows on evenings and weekends because there's no time in the week to do it. They're using personal accounts because nobody's approved the enterprise tier. They're documenting nothing because nobody asked them to, and because writing up their process for leaders who don't understand it feels like a tax on their own productivity.
They feel invisible. They're producing five to ten times the output of their peers and nobody in leadership can describe what they actually do.
The Spark Data Calls This Out
Spark AI's 2026 report names this exactly. The people driving AI adoption inside agencies feel legally and ethically exposed, because they're pushing into territory where no rules exist. They want guardrails. They want visible backing. They want their work to show up in performance conversations as real contribution, not "a side project."
The report's language is direct: "The enthusiasts burn out or leave. The sceptics dig in. The middle majority reads the room and decides it's not worth the effort."
That's three groups of people all reaching the same conclusion. And it's being driven by leadership silence.
The Quiet Disengagement You'll Miss Until It's Too Late
The thing about losing enthusiasts is it doesn't look like a crisis when it's happening.
They don't complain in town halls. They don't write Slack posts about their frustration. They get on with the work, because that's the kind of person they are. And then six months from now, one of them announces they've got a new role at a competitor. Or an in-house team. Or they're going freelance.
The workflows leave with them. The prompts. The custom GPTs. The hour-saving automations built on top of the client's brand guidelines. All of it.
You're left with a team that produces at a slower pace and doesn't know why. You're also left with a hole in your capability that will take a year to refill — assuming you can hire someone who can do the same thing, which you probably can't.
What They Actually Need
This isn't complicated. The people pushing AI forward inside your team need four things, and none of them cost much.
Protected build time. Not evenings. Not weekends. A named block in the working week — a two-hour Friday morning, a monthly half-day — where experimenting and building is the job. If you don't ringfence it, they don't get it.
Approved tools, paid for by the agency. The enterprise tier of the model they actually use. Custom GPT or project licences. Proper API access if they need it. The whole thing costs less per month than one billable hour. Stop making them pay for their own productivity.
Visibility to leadership. A monthly show-and-tell where they demo what they built. Not for validation — for context. Your leaders should be able to say out loud, in their own words, what their best AI people are actually doing.
Promotion and compensation that reflects the work. If someone automates a workflow that saves the team twenty hours a week, that's leadership. Treat it like leadership. Promote accordingly. If the only person who benefits from their work is the agency, and nothing about their role or pay changes, they will eventually take the capability somewhere it does get rewarded.
The Math Is Brutal
Losing one AI power user costs you more than losing three average performers. That sounds harsh. It's not my opinion — it's what the operational data consistently shows.
Their workflows are worth weeks of productivity per month across the team. Replacing the person is one problem. Rebuilding what they built is a bigger one.
The fix is cheap. The loss is expensive. The agencies pulling ahead aren't the ones with the most AI tools. They're the ones that noticed which three or four people were actually moving the work forward — and made sure those people had every reason to stay.
Find yours. Today. Have the conversation before they have it somewhere else.