DEV Community

Sumas Keller
Sumas Keller

Posted on

How I Build AI Usage Policies People Actually Follow

Most AI usage policies fail for a simple reason.

They are written by legal teams.

Then handed to employees.

Then forgotten.

A month later, nobody remembers where the document lives.

Three months later, people are doing whatever they want anyway.

I've seen this happen multiple times.

The problem isn't that employees dislike rules.

The problem is that most AI policies are impossible to use in real work.

A policy only works if people can remember it while they're busy.

That's the standard I use.

Not whether the policy sounds impressive.

Whether a normal employee can apply it at 4:30 PM on a Friday.

The Biggest Mistake: Writing For Auditors Instead Of Employees

Many AI policies read like compliance documents.

Pages of definitions.

Pages of legal language.

Pages of edge cases.

Technically correct.

Operationally useless.

Employees don't need a 20-page document when deciding whether they can paste information into an AI tool.

They need a simple answer.

Can I do this?

Yes or no?

The more effort required to find the answer, the less likely people are to follow the policy.

Rule #1: Classify Information Before You Classify Tools

Most companies start with tools.

I start with information.

Because tools change.

Data doesn't.

I typically group information into three buckets:

Public information

Examples:

  • marketing content
  • public product documentation
  • published research
  • public website content

Internal information

Examples:

  • internal processes
  • project plans
  • meeting notes
  • operational documents

Sensitive information

Examples:

  • customer records
  • financial data
  • contracts
  • employee information
  • security documentation

Once employees understand these categories, tool decisions become much easier.

The question becomes:

"What information am I sharing?"

Not:

"Which AI product am I using?"

Rule #2: Make The Safe Path The Easy Path

People naturally choose the fastest option.

Good policies acknowledge this reality.

Bad policies fight it.

If employees must jump through ten steps to use an approved AI solution, many won't.

Instead, they'll find a shortcut.

That's how shadow AI starts.

I've learned that enforcement is much easier when approved tools are more convenient than unapproved tools.

Convenience is a governance tool.

Not just a product feature.

Rule #3: Focus On High-Risk Behaviors

Another mistake I see is trying to regulate everything.

That approach rarely works.

Instead, identify the behaviors that actually create risk.

For most organizations, those include:

  • uploading customer data
  • uploading contracts
  • sharing credentials
  • exposing financial information
  • exposing internal security information

Those deserve attention.

Whether someone uses AI to rewrite a meeting summary usually doesn't.

Good policies prioritize.

Great policies prioritize aggressively.

Rule #4: Give Employees Examples

Examples are more useful than rules.

Compare these two approaches.

Approach A:

"Do not upload confidential information."

Approach B:

"Do not upload customer contracts, financial statements, employee records, or unpublished product roadmaps."

The second version is far easier to follow.

People remember examples.

They rarely remember policy language.

Whenever possible, I replace abstract rules with concrete scenarios.

Rule #5: Assume Policies Will Be Ignored

This sounds pessimistic.

It's actually practical.

Every policy should assume occasional mistakes.

That means organizations need:

  • logging
  • monitoring
  • approval workflows
  • permission controls
  • audit capabilities

Policies reduce risk.

Systems enforce risk.

Both are necessary.

A document alone has never protected a company.

Rule #6: Review Policies More Often Than You Think

AI changes quickly.

What was reasonable six months ago may already be outdated.

I've seen companies create AI policies once and never revisit them.

That's risky.

A simple review cycle works much better.

Quarterly is usually enough.

The goal isn't rewriting the entire policy.

The goal is keeping it aligned with reality.

Because reality changes.

Fast.

What Good AI Governance Actually Looks Like

Good governance doesn't feel restrictive.

It feels clear.

Employees know:

  • what they can do
  • what they cannot do
  • which tools are approved
  • what data is sensitive
  • who to ask when unsure

When those answers are obvious, adoption becomes easier.

And risk becomes lower.

That's the outcome most organizations want.

My Take

If employees constantly ask whether they're allowed to use AI, the policy is probably too complicated.

If nobody reads the policy, it's definitely too complicated.

The best AI usage policies aren't the longest.

They're the easiest to remember.

Because a policy only creates value when people actually follow it.

And people only follow policies they understand.

Top comments (0)