Meet teams where they work
What building an agency's AI guidance taught me about access, ownership and an experimental mindset.
I’d spent months writing down how our agency works with AI: the standards, the procedures, the templates. It was careful work, and I kept it in a code repository that most of the team couldn’t open and the rest couldn’t comfortably read. I’d written guidance for people and put it where few of them could use it.
I lead UX and AI operations at a small, boutique agency, and spotting usability problems is part of my job. It still took the team to show me this one. By this fall everyone had the same AI tools, and when we asked where their work still got stuck, the top answer wasn’t the tools. It was gathering information that lived in several places. My guidance was one more of those places.
Even good guidance gets in the way if people have to stop what they’re doing to go find it. Ideally, people shouldn’t have to think about the rules at all, and when they do have a question, the answer should be one question away. So now I ask whether the guidance shows up where the work happens, or whether someone has to go looking for it. In other words, meet teams where they work.
Put the guidance where the work happens
When people have to go find a rule, that’s an extra step, and on a deadline it’s the first thing to go. So we built the standards into the recurring work itself. Briefs, decks and audits each got a skill: a set of instructions the AI loads for that kind of job, with our standards built in, so a first draft starts from how we work instead of a blank prompt. It’s the same thing I’d do for a new teammate: show them a good example, give them the context and set the standard.
Everything else went into one place, built straight from the source documents and organized by the questions people bring to it: where do I start, where are we headed, how do I do the work, what should it look like, and how do I get set up? When a document changes at the source, the change shows up there too, so nobody has to keep separate copies in sync. That place opens to the whole team this month.
Figuring out where the guidance lives also meant figuring out who it was for. Some of what I’d written was for leaders: costs, platform decisions, where the program goes next. It was useful, but it got in the way of people looking for how to do their work, so we moved it out of the team’s guidance and into a separate place for leadership.
The downside is upkeep. The instructions need maintaining as the work changes, and I’d still take that over the alternative. I think of an AI policy a lot like a brand guide: it’s a living document. It can keep evolving, as long as people know what’s settled and what’s still being tested, so every piece is labeled as active, in pilot or still a draft.
Keep a person at the points that matter
Building the guidance into the work doesn’t mean taking people out of it. Our written procedures have validation gates: required checkpoints where, before work moves forward, a reviewer checks its sources, its fit with the brief and whether it meets our standards. The gates explain themselves in plain language. Every agent also has a named person who reviews what it produces. That person is the one setting the standard, so they need to know what good looks like.
That friction is on purpose. The difference is where it sits: the checkpoint is part of the procedure people are already following, so it’s not one more rule to remember.
Let the people doing the work shape it
The team told us more than where they get stuck. We asked what they want to keep doing themselves, whatever AI can do, and they named strategic thinking, creative concepting, the final creative calls, writing in a client’s voice and the conversations with clients. Nobody was asking AI to take over their judgment. They wanted help with the gathering, the reformatting and the status updates.
We also asked what would get them to try something new. The most common answer was a demo on one of their own tasks, and a lot of people asked for someone to sit with them for an hour on their real work. So that’s how the rollout works: department by department, starting from a task its people already do every week. Leadership has set time aside for it, and department pilots are under way. The plan also has to fit the time people have: I lead AI operations alongside my UX director role, and everyone else fits it in around client work.
Not everything is up to each team. What can go into an AI tool for client work is a standard we set for everyone. How a workflow runs belongs to the people doing it. They decide which recurring tasks are worth automating and help define how: show the AI how it’s done, sort out which steps keep a person, and coach it on real jobs until the results hold up. The people doing the work should own and define those workflows, since they’re the ones who have to live within them.
What I’m after across all of it is an experimental mindset, the kind a beginner brings. The goal isn’t one perfect workflow per department. It’s giving people room to try an approach on real work, see where it falls short, and help decide what becomes standard. That includes permission to drop an approach that isn’t helping. The ones that hold up are worth making standard, because the people doing the work have already proved them.
What we still need to learn
So far, all this shows is that the plan responds to what the team told us. It doesn’t show yet that it helps people do their work, and that’s the part I care about.
Between the two surveys, people reported using AI more often and feeling clearer on the rules. That gives us a baseline, but it doesn’t tell us whether the new guidance helps.
The open question is whether people can find and apply the guidance with less help. In January 2027, a short check-in asks whether people’s weeks feel any different, and I want it to ask where they still lose time finding context or checking their work. That should tell us where people are still getting stuck.
Why it matters to me
In infantry training I learned that slow is smooth, and smooth is fast. In this work, that means giving people time to learn, with a beginner’s mindset, before deciding what becomes standard, and then writing down what good looks like and putting it where people will use it.
As AI takes on more production work, roles are moving toward strategy and harder problems. I want the people on my team to move with that shift, not get caught behind it. Building the rules into the work frees their time and attention for learning the new parts of the job. That’s what I’d bring to any team: a way to adopt AI that grows the people using it.
