Before you can build the roadmap, you have to know where you're starting from. Most enterprises don't. They think they do. But "we have an AI strategy" is not the same sentence as "we have governed AI in production".
Here's a five-rung maturity ladder. Be honest about where you land.
Rung 1 — Shadow AI# #
Agents exist in production. Nobody knows where, who deployed them, or what they touch. This is the default state for most enterprises right now, whether or not the security team knows it. If you can't answer "what did agents touch last quarter" in one query, you're here.
Rung 2 — Tracked AI# #
An inventory exists. Policy exists too, on a wiki nobody reads. You know agents are running. You don't know what they're doing in real time, and you can't enforce your own policy at the point of action. You have visibility without enforcement.
Rung 3 — Governed AI# #
Single intake. Defaults beat docs. The audit trail captures every tool call. Policy enforces at request time, not retroactively. This is where most organizations say they are. In my experience, far fewer actually are.
Rung 4 — Federated Reuse# #
Teams publish agents, workflows, MCP tools, and policy templates and consume each other's work; reuse is measured and rewarded. Existing people enable; they don't gate. The economics start to compound. This is rare. When you find it, it's usually in a business unit, not enterprise-wide.
Rung 5 — Compounding# #
The platform is the policy. Self-service is the security boundary. The audit log is the product. Every additional agent makes the next one easier to build, govern, and trust. Marginal cost of governance trends toward zero per agent. This is the destination. Almost nobody is here yet.
The diagnostic questions# #
Where you actually are isn't a feelings question. It's an evidence question. Four questions that tell you the truth:
How many agents are running in production right now that you know about? Now estimate the ones you don't. If those numbers are close, you're rung 2 or better. If the second number is a guess, you're rung 1.
Can you answer "what did agents do last quarter" in one query? One audit log, one query, under 60 seconds. If the answer is no, or if it requires pulling logs from five different systems, you're not rung 3, regardless of what the policy deck says.
When something goes sideways, do you have attribution back to the human originator? Not "service account 47." The actual person who originated the chain. If the answer is no, the audit log isn't working for you yet.
Is the safe path faster than the unsafe path? This is the one most enterprises fail without realizing it. If a developer can get a credential faster by down a JSON file to their laptop than by going through the governed path, they'll download the file. Every time. Not because they don't care about security. Because they have a job to do.
Why the honest assessment matters# #
The rung you're actually on determines which moves to make first. Organizations that misread rung 1 as rung 3 skip the foundational work: the intake funnel, the audit log, the identity graph, and go straight to building a marketplace with nothing underneath it. That's how you get a catalog full of assets nobody uses and governance that looks good on a QBR slide and breaks at the first production incident.
The honest meeting (CIO chairs, CISO and CTO in the room, two or three business unit leaders) is the most valuable hour you'll spend this quarter. Not to perform an assessment. To name what's actually true, in writing, and commit to not repeating the past failures that got you here.
Pick the rung you want to be on in 18 months. Build backward from there. That's the executive question this methodology is designed to answer.
Part three of a series based on the Agentic Adaptation Playbook. Previously: seven ways enterprise AI programs die. Next: the ADAPT phase, the one honest meeting and written charter that starts the whole thing.