Governance Before Intelligence:
An AI Rollout Sequence That Sticks

When a company decides to "roll out AI," the first question they usually ask is which tool. It's the wrong question. The right one is which step, and the answer is almost never the one they were planning to take first.

I just finished a full AI rollout inside a mid-size professional services firm — 50–200 people, thirty different tools in flight, individual AI subscriptions paid out of pocket. We didn't buy anything new for the first month. We didn't configure a model. We didn't train a prompt. What we did was put the sequence in the right order.

Step 1: Consolidate the subscriptions

Before anything else, the org has to own its AI. Individual team members paying for AI tools on personal cards is the silent killer of every rollout. There's no usage visibility, no data governance, and no way to build shared context across the company because nothing is shared by design.

Move everything to an org plan. Configure usage limits, set data monitoring policies, turn on admin visibility. Do this before you do anything else. It's not glamorous; it's the foundation every subsequent step depends on.

"AI that isn't governed isn't AI. It's fifty unrelated chat sessions, and most of them will quietly delete themselves."

Step 2: Connect the context

The single biggest quality jump in an AI workflow doesn't come from a better model. It comes from giving the model the right context. On our rollout that meant connecting the AI knowledge base to three systems: Slack, Google Drive, and Calendar. Nothing else. Three.

Suddenly the same tool that was writing generic emails the week before was writing emails that referenced actual project conversations from that channel last Tuesday. The system didn't get smarter. It got informed. That distinction is most of the ground most rollouts never cover.

Step 3: Build the MCP layer

Chat is a demo. The moment you need the AI to do something — write to the CRM, create a record in the PM tool, move something in the recruitment system — you need connectors. We built four. Each one took a few days; each one unlocked workflows we couldn't have done any other way.

This is the unglamorous infrastructure layer. It's also the layer that separates "we use AI" from "AI is how our operations run." Every connector is one more surface where the agent can stop being a suggester and start being an operator.

Step 4: Ship custom skills — and keep shipping

With the foundation in place, the marginal cost of a new AI capability drops dramatically. On this rollout we deployed five custom skills in the first quarter: weekly reporting, vendor review, risk assessment, status reports, process documentation. Each one built once, reused at zero marginal cost thereafter.

This is where the compounding starts. A skill built in week twelve uses the context layer from step two, the connector layer from step three, and the governance from step one. Built in week two without those layers, the same skill would have been a brittle one-off. Built in week twelve, it's reusable infrastructure.

Step 5: Measure adoption, not output

The metric that matters after launch isn't how many tokens the org processed. It's which teams are using which skills, and how often. Low-adoption skills point to friction; high-adoption skills point to where the next investment should go. We track this per-person because that's where support needs show up — and because a small number of power users will tell you more about the roadmap than any survey.

What the shortcut looks like (and why it fails)

The shortcut every org is tempted to take: start at step four. Just ship some custom AI workflows and call it a rollout. This almost always fails, but it fails slowly, so the cause looks like something else — "the model wasn't smart enough," "the team didn't adopt it," "the data was messy." In reality, the earlier steps weren't done, and every problem that surfaced was a consequence.

Governance before intelligence. Context before connectors. Connectors before skills. Skills before measurement. The order is not optional. It's the rollout.

Every AI rollout I've been asked to rescue was one that skipped a step. Every rollout that's still running a year later was one that didn't. The sequence is boring. It's also the entire game.