Dinner at my house is never "on average" ready. The ragù can be four hours in and exactly right, and if the pasta went into the water ninety seconds ago, nobody is eating. Average the two and you get a number telling you the meal is half done. That's true, and it's useless. The thing you actually need to know is which part is behind, because that's the part you go stand over.
Agencies have the same problem with AI, and almost nobody has handed them a way to see it. Most principals I talk to have a rough sense of how far along they are, usually a generous one. What they don't have is the shape of it. They know the score and not the spread, so the next move stays guesswork, and guesswork is expensive in a year when everyone around them is spending.
So here's the map. Five stages an agency moves through, what each one looks like on an ordinary Tuesday, and the single move that gets you off it.
The curve is not a scoreboard
The five stages come out of six things worth measuring in any agency: whether there's a real mandate, how many people actually use the tools, how you introduce AI to a process, what the AI does for you today, how it's governed, and whether you can point at a result.
Average those six and you land on a stage. That average is the least interesting thing the exercise produces.
Two agencies can land on the same stage and be in completely different trouble. One is level across the board, moving slowly and evenly. The other has a strong number in adoption and nothing underneath it, which is a compliance exposure wearing an adoption success story. Identical placement, and the two of them should spend the next quarter doing almost nothing in common. The gap between your strongest area and your weakest one tells you more than the stage name ever will.
Read the shape. The stage is just where the shape happens to sit.
Stage one: Observer
You know AI matters. The agency hasn't decided what to do about it.
On a Tuesday this looks entirely normal. Quotes go out, renewals get worked, and the AMS gets fed by hand the way it always has. Somebody in the office uses ChatGPT for email drafts and hasn't mentioned it. There's no policy, because there's nothing anyone thinks needs a policy yet.
The instinct here is to go buy something. Resist it for about two weeks. The move that actually pays is smaller and duller: pick one week and write down where the hours go. Not a guess, a record. Most principals are genuinely wrong about which task is eating the most time, and the wrong starting point costs you a year. This is Admin Drag before anyone has named it out loud, and you can't redesign work you haven't looked at.
Stage two: Tool Sprawl
You've bought the AI tools. Your people aren't using them in any way that matters.
Tuesday looks like licenses. Copilot came with the Microsoft renewal, somebody expensed ChatGPT Plus, the AMS vendor turned on an AI feature nobody asked about. Three producers are enthusiastic, everyone else opens the tab once a month, and the spend shows up on a line item that nobody can tie to an outcome.
This is the most common place to be, and it's the most frustrating, because from the outside it looks like progress. It isn't. A tool nobody owns just becomes another tab, which is the Toggle Tax with a subscription attached.
The move is adoption, and adoption is not training. Take one task that one team does every week, and redesign that task around the tool in the room with them. One task, one team, one afternoon. What you're buying is a person who can say out loud that it worked.
Stage three: Pilot Purgatory
One thing is working. Nothing is compounding.
The certificate request summarizer is genuinely good. It's been genuinely good for nine months. It hasn't spread to renewals, nobody has built the second one, and the producer who made it has quietly become its support desk.
I've been here myself, on a smaller scale and with less excuse. I spent a couple of hours building what I intended to be an agent in Zapier and ended up with an automation that got more complicated every time I touched it, until I abandoned it. That wasn't a tool problem. I'd skipped the part where you map the work first, so there was nothing to build against, and complexity filled the gap.
The gap at this stage is architecture. You have a working thing and no pattern, so the second one costs as much as the first and the tenth never gets built. The move is to write down how the one that worked actually works, in enough detail that somebody else could build the next one from it. There's more on the shape of this trap in Pilot Purgatory.
Stage four: Workflow Redesign
Architecture has caught up to the technology. The work now is scale, not implementation.
Tuesday looks different here in a way that's hard to miss. Somebody maps the workflow before anyone touches a tool. There are two or three redesigned processes running, each with a named human who answers for what it produces. Governance exists on paper and people have actually read it.
The honest difficulty at this stage is that it stops being a technology project and becomes a management one. You're now redesigning jobs, and some of your best people are attached to the version of their job that you're changing. Keep going, and say plainly what's changing and why, before someone else in the office tells the story for you.
The move is to stop building one at a time and start building from a template, with the people who'll live with the result in the room.
Stage five: Digital Crew
AI is no longer a project. It's how the agency operates.
Named functions run on their own, each one reporting to a person who owns its mistakes. New hires learn the crew during onboarding, the way they learn the AMS. The results sit in your regular operating metrics instead of a deck.
Very few agencies are here, and the ones that are didn't get here by buying better tools than everyone else. They got here by deciding earlier than everyone else who owns what. That's the whole of the Digital Crew argument, and the part people skip.
What the shape tells you that the stage cannot
When your weakest area and your stage's usual weak spot are the same thing, the diagnosis is easy and you already knew it. When they disagree, that disagreement is the most useful sentence anyone can hand you.
An agency at Tool Sprawl is normally told its problem is adoption. Fine. But if that same agency has half its people using AI daily and no written policy anywhere, adoption isn't the problem at all. The problem is that client data is moving through personal accounts right now, in a regulated business, and the adoption number is what's hiding it. Nobody would have caught that from the stage name.
This is also why the stages aren't a ladder you climb by working harder. You move by fixing the specific thing that's behind, which is usually not the thing you've been paying attention to.
Mike Welker, who runs the benefits side at Insurance Center of Utah, is the clearest version of this I've watched. Before a session at the Kellogg Business Summit last May, the thing eating his week was carrier downloads. The benefits side doesn't get the automatic feeds that property and casualty does, so every carrier's production and cancellation reports were collected by hand, analyzed, keyed into Applied Epic, then keyed again into Agency Zoom. He'd been trying to standardize those spreadsheets for a year.
He built his first assistant in that session. Sixty days later, after four or five goes at it working off and on, he had a pipeline that masks the protected health information, collects the carrier reports, standardizes them, restores the data locally, and distills the whole thing into a dashboard his team uses all day. He's since chained it to DocuSign for scope-of-appointment forms.
The detail I'd point at is the four or five attempts, which is the credibility of the whole thing. He didn't find a better tool on the fifth try. He got clearer each time about what the work actually was.
Where this leaves your Monday
If you read the five stages and knew immediately which one you're in, you're probably right about the stage and probably wrong about the reason. That's the normal failure, and it's not a knock on anyone's judgment. You're inside the thing you're trying to see.
The fastest way to get the shape rather than the guess is to answer the six questions honestly and look at the spread. That's what the AI Adoption Reality Check is for. It takes about five minutes, it gives you your placement immediately, and I read every one and write back within two days with the part the automation can't tell you: what's anchoring you, where you're already ahead, and what that gap means in practice.
Bring the honest answers rather than the flattering ones. The flattering ones produce a nicer number and a worse Tuesday.
