How to Build an AI Governance Roadmap in 90 Days

Ninety days is enough to make the AI a business already runs visible and controlled, if an AI governance roadmap splits the quarter into three 30-day blocks. Days 1 to 30 find and rank every model, agent and tool. Days 31 to 60 put controls on the riskiest ones. Days 61 to 90 test those controls and pilot one or two agent workflows under them. You will not solve every future AI risk in a quarter. You need a baseline you can see and a process you can repeat.

Start with what already exists

A 90-day roadmap opens with one question: which AI models, agents, tools, data sources and workflows are already running?

Our example is a made-up home-services company with 150 people. Leadership believes it uses two AI tools. The first week of inventory finds eleven, including an AI writing feature switched on inside the CRM, a quoting assistant a sales rep built on a personal account, and an agent that reorders parts with a company card. The gap between two and eleven is the reason the roadmap starts here.

This roadmap assumes some AI already acts on its own. If your team mainly uses AI tools by hand, start with our 90-day AI governance setup for small business, which covers policy, training and proof. The two calendars can run side by side.

Days 1 to 30: inventory and triage

Create the inventory first. For each system, record the owner, provider, model, purpose, users, data access, tools, outside actions, customers affected and current controls. Our AI governance checklist explains why it comes first.

Then risk-tier each system. Put on top anything touching sensitive data, customer systems, outside communication, money, code repositories or production. Give every system an owner by name, not by team.

Last, pause what is clearly uncontrolled. In the home-services company, the parts-reorder agent can spend on a company card with no cap and no review. It stops until it has both, which can take a single conversation.

The tiers become your risk register: each risk, what you decided to do about it, who owns it, and the residual risk left after your controls.

Days 31 to 60: put controls on the top tier

Define a lightweight approval process for new AI systems, so the next quoting assistant gets reviewed before it goes live. Then give each high-priority workflow a unique agent identity (never a shared login), permissions scoped to the task, logging of what the agent saw and did, a tool allowlist (the short list of tools it may call) and an escalation path to a named person.

For the reorder agent, that could mean its own supplier login, a weekly cap of $2,000, and any order over $500 waiting for the purchasing lead. Those figures are illustrative; set yours from your real order history.

Build a small evaluation set for each critical use case: twenty or thirty real past tasks with the right answers written down. Define what triggers a review, a rollback or a shutdown. Adopt one shared incident process, so nobody invents a response under pressure.

Days 61 to 90: test, pilot and measure

Run regression tests, re-running the evaluation set to confirm nothing got worse. Then run a kill-switch exercise: stop an agent mid-task, and time how long it takes and who had to be involved. Review access and evidence trails. Then pilot one or two high-value agent workflows under the new controls.

Measure the pilot on task success, cost per accepted outcome, latency, human-review time, policy violations, tool errors and user feedback. Cost per accepted outcome puts the model bill and review time in one number.

Review time is more than half that bill. Our guide to AI cost per task walks through the full calculation.

What you hold on day 90

That is a minimum operating system, not a finished program. Our post on the AI governance operating system covers what keeps it running after day 90.

Where 90 days runs short

The kill-switch exercise is the step most likely to break the calendar. If stopping the reorder agent takes 40 minutes because only one engineer knows how, fixing that can eat the pilot's time. Let it. A pilot under a stop control nobody can use is autonomy without a brake.

The pilot has its own limit: a month at low volume may not give enough runs to trust a success rate. Twenty runs with two failures is a 90% success rate. Three failures in the same twenty runs is 85%. Treat month-three numbers as a first reading, and say so when you report them. If the inventory turns up thirty systems instead of eleven, the calendar stretches. Decide on day 30, once you have the count, whether to plan a second quarter.

Pairs well with

The Stop-Authority Assay ($109) grades what each stop control has been shown to do rather than what it claims, returning STOPS, SLOWS or NAMED ONLY. The Prompt Regression Lab ($89) snapshots a baseline, diffs every prompt change and fails your build pipeline on any regression: ship, hold or regressed. The AI Agent Quiet-Failure & Drift Monitor Kit ($49) computes the expected end-to-end success rate of multi-step agents: an 8-step agent at 85% per step succeeds all the way through 27.2% of the time (0.85 to the eighth power, treating the steps as independent).