An AI Governance Framework That People Actually Follow

Most AI governance frameworks I have seen are compliance documents dressed up as operational frameworks. They describe what should happen in ideal conditions, they get approved by leadership, they get distributed to employees, and then they sit in a shared drive while the actual decisions about AI usage get made informally based on whatever feels reasonable to whoever is making the decision that day.

The reason this happens is not that people do not care about governance. It is that most governance frameworks are designed to be complete rather than to be used. They cover every possible scenario, include exhaustive policy language, and require the reader to hold a mental map of the whole document to understand how it applies to any specific situation. Nobody has time for that when they are deciding whether to paste a client email into an AI tool to get a draft response.

The governance framework that people actually follow looks different. It is designed around the decisions employees actually face, not the scenarios the governance team imagined they would face. It is short enough to remember. And it provides clear enough guidance that employees can apply it without consulting anyone else for common situations. Here is the structure that has worked for me across several deployments.

The three-tier classification

Every piece of information that touches AI needs to be classifiable quickly into one of three tiers. If the classification requires a judgment call every time, it will not be applied consistently.

• Tier 1: Public or general information. This is content that is either publicly available or that the organization has no material interest in keeping confidential. Marketing copy, publicly available competitor information, general industry knowledge. AI tools can process this freely, including external AI tools with standard terms.
• Tier 2: Internal business information. This is information that is not public but is not specifically sensitive: internal process documentation, meeting notes about operational topics, general project updates, internal communications that do not contain client data or commercially sensitive content. AI tools can process this with enterprise agreements in place. External AI tools with standard consumer terms cannot.
• Tier 3: Sensitive or confidential information. This includes: client data, personally identifiable information, financial data, legal strategy, M&A activity, HR records, anything under NDA, and any information that a client has shared in a professional relationship. AI tools can process this only if they run entirely within the organization's own infrastructure. External AI tools, even with enterprise agreements, are not appropriate for Tier 3 content.

The classification should be done by the person initiating the AI interaction. If they are unsure which tier content belongs in, the default is the higher tier.

The five common situations

Rather than trying to cover every scenario, the framework covers the five situations employees encounter most often and tells them specifically what to do.

1. Drafting an email to a client. The email will likely contain client-specific context, making it Tier 3. Do not paste client-identifying information into an external AI tool. Either use an internal AI tool, draft the email without the AI, or use the AI only for general writing assistance on a version with client details removed.
2. Summarizing a meeting. If the meeting involved clients, client names, or confidential project details, it is Tier 3. Record and transcribe internally. If the meeting was internal and covered operational topics only, it is likely Tier 2 and can use AI tools with enterprise agreements.
3. Researching a market or competitor. This is almost always Tier 1 and any AI tool is appropriate.
4. Drafting a proposal. Proposals contain client context and pricing information, making them Tier 3. Use only internal AI tools. Do not use external AI tools even at the draft stage if client-identifying information is included.
5. Analyzing internal data. Depends on the data. Financial projections, HR data, or data that could identify individuals is Tier 3. General operational metrics or anonymized data is Tier 2.

The escalation path

For situations not covered above, the framework provides a single escalation path rather than a decision tree. Employees who are unsure contact the designated AI governance contact for their team. That person has the authority to make a binding classification decision in under 24 hours. If the 24-hour window is too slow for the situation, the default is to treat the content as Tier 3 and restrict to internal tools only.

This matters because governance frameworks that require escalation through multiple levels, or that take a week to produce an answer, get bypassed. The escalation path needs to be fast enough to be useful and authoritative enough to be trusted.

What makes it work

The framework I described above is not comprehensive. It will not cover every situation. There will be edge cases that require judgment calls. A comprehensive framework that covers every case is a framework that is too complex to remember and therefore not used. The operating principle is that a governance framework that is incomplete but followed is more valuable than one that is comprehensive but ignored. The goal is to move decisions from "whatever feels reasonable" to "consistent with a framework that reflects our organizational values and risk tolerance." You do not need to cover every edge case to make that move.

The other thing that makes it work is visible leadership behavior. If the people who created the framework are seen using external AI tools in situations that the framework says require internal tools, the framework means nothing. Governance frameworks require behavioral modeling from the people who champion them, not just policy language in a document. The frameworks I have seen successfully adopted share one property that the unsuccessful ones lack: the people who created them also visibly comply with them. That visible compliance is not a nice-to-have. It is what determines whether the framework is operational or decorative.