Most companies approach AI with one of two settings: lock it down, or let it run wild. Neither holds up. Lock it down and you lose the productivity, plus you create the shadow-AI problem where people just do it quietly anyway. Let it run wild and eventually something sensitive ends up somewhere it shouldn't.
There's a third option, and it's a lot simpler than the policy templates make it sound. Tier your rollout, and match the level of oversight to the level of risk. A tool that reformats emails does not need the same review as a workflow that touches your general ledger.
The two bad defaults
Prohibition feels safe and isn't. It assumes you can stop a behavior that's already happening on everyone's phone. The wild west feels modern and isn't safe either — it treats a marketing intern's chatbot experiment and a finance automation as if they carry the same weight. The fix isn't to pick a side. It's to stop treating all AI use as one category.
The three tiers
Think of it as three lanes, each with a clear entry requirement. Most of what your team does will live in the first lane, which is exactly the point.
Personal experimentation
Anyone can use AI for their own work, guided by a short set of standards. Drafting, summarizing, brainstorming, cleaning up a spreadsheet on their own machine. No approval needed. The goal is to remove friction from learning.
Business-unit deployment
The moment a tool is shared across a team or touches real systems, it gets connected to version control and reviewed before it goes live. Someone other than the builder looks at it. This is where most of the value — and most of the risk — actually lives.
Company-wide systems
Anything that touches the whole organization, customer data at scale, or core financials runs through your full enterprise process: security review, documentation, ownership, the works. Few things reach this tier, and that's fine.
The magic is in the sorting. When people know which lane they're in, they stop either over-engineering small experiments or under-protecting big ones.
Guardrails that travel across all three
A few rules apply no matter the tier, and they're worth writing on a single page so people actually retain them:
- No company data goes to external hosting or consumer tools that train on your inputs.
- Access to live systems is gated — experimentation happens against copies, not production.
- Training comes before access to anything beyond Tier 1, not after.
- Every shared tool has a named owner and lives in version control.
If you want a credible backbone for this, NIST's AI Risk Management Framework organizes the whole thing around four functions — Govern, Map, Measure, Manage. You don't need to adopt it wholesale, but it's a useful spine when someone asks where your model came from.
Who owns it
Each tier needs an owner, but the one that matters most is leadership ownership of the whole thing. This is the part companies skip, and it's the part that determines whether any of it sticks.
3x
more likely to have a mature program when the C-suite owns AI governance directly
IAPP, via Responsible AI Labs, 2025
70%
of enterprises say they lack optimized AI governance today
Acuvity, 2025
Governance has a reputation as the thing that slows AI down. Done like this, it's the opposite. Clear tiers are what let you say yes quickly to the small stuff and yes confidently to the big stuff.
Where to start this week
You can have a working draft by Friday. Write down your three tiers in plain language. Then take an inventory of the AI tools already in use across the company and sort each one into a lane. The exercise alone will tell you more about your real exposure than any policy document, and you'll end the week with something your leadership team can actually act on.
Key Takeaways
- Don't pick between banning AI and letting it run wild — tier it instead.
- Tier 1 is local experimentation, Tier 2 is anything shared or connected to real systems, Tier 3 is enterprise-wide.
- A handful of guardrails travel across every tier: no external hosting of company data, gated access, training first, named owners.
- Leadership ownership is the multiplier — it correlates with a 3x more mature program.
- Start by writing your three tiers and sorting your current tools into them.