How to Create Company-Wide AI Agreements

One of the things I’ve learned about bringing AI into a company is that the early experimentation is the easy (and fun) part. Give your people room to learn, encourage them to find ways to work better, and it won’t be long before you see AI helping them do better work.

The harder part comes next. Your Customer Success team figures out which details in an AI-generated account summary have to be checked before a customer call. Your Finance team learns that an AI-drafted forecast needs a named reviewer before it goes to the board. Someone in Operations discovers that an agent handles one step of a process well and the next step badly.

Each of those is a lesson. And a lesson that stays inside one team is a lesson the rest of the company has to pay for again.

Throughout my AI Leadership series, we've talked about leading AI adoption across your company, deciding where agents belong, and making sure every agent has one accountable human. This article is about the next leadership responsibility: turning what your teams are learning into agreements your whole company can follow.

Learn Before You Write the Rules

There’s only so much you can decide about AI before your people have spent real time with it. Write a complete AI policy on day one and you’ll get a long document built on guesses. Much of it will be wrong, and the parts that are right will be buried.

That doesn’t mean starting with no boundaries. A few belong in place from the beginning: which AI systems are approved to work with your company’s information, which general-purpose tools are for everything else, and who can see what. You don’t need experience to know those. You already know them.

For everything else, let experience do the writing.

Take a Customer Success team that starts using AI to prepare account summaries before calls. After a few weeks, it’s clear the summaries save significant preparation time. It’s also clear that certain details, like renewal dates and open support issues, have to be checked before anyone repeats them to a customer. Now the team knows something it didn’t know when it started: AI can do the preparation, and there’s a specific point where a person has to step in.

That’s an agreement waiting to be written. And because it came from experience, the people who follow it will understand why it exists.

Decide Which Lessons Are Bigger Than the Team

Not every lesson should become a company-wide agreement. Your Finance team is going to work with AI differently than your Marketing team, and forcing both into one rulebook helps neither of them.

Here’s the test I’d use: Does the lesson protect something the whole company shares?

Customer trust is shared. So are employee privacy, the commitments you’ve made, and your accountability for outcomes. If one team learns that an AI tool will happily show a new hire information their Seat has no reason to see, that lesson protects something every team touches. It belongs company-wide.

If another team finds a faster way to prepare for meetings or organize research, that’s a better way of working. Share it. Celebrate it. But it doesn’t need to become an agreement. Turning every good idea into a rule is how companies end up with a policy manual nobody reads.

Most of the company-wide agreements you’ll need will answer questions like these:

• Which AI systems are approved to work with company, employee, and customer information, and which are for general use only?

• Who can see what, and does your AI respect the same permissions your people do?

• At what point does AI-generated work get reviewed by a person?

• What does someone need to do before trying a new AI tool for company work, and who approves it?

And underneath all of them, every agent still has one accountable human.

Your department leaders will bring you the lessons. You and your leadership team decide which ones are bigger than the team that found them.

Agreements About Information Will Change the Most

Of all your AI agreements, the ones about information are going to change the most.

Early on, the instinct is to keep sensitive information away from AI entirely. That made sense when the only AI available was a general-purpose tool outside your walls. But the most valuable AI agent in your company will be the one that knows your company best: your goals, your numbers, your people, your customers, and years of what you tried and what happened. An AI that can’t see any of that can only give you advice that’s true of companies in general.

So the question is no longer about whether your information goes into AI. Now it's about where it lives, who can see it, and who owns it when the model changes. As I wrote in the last article, models will come and go. Your context, your history, and your permissions should stay with you.

That’s how we’re building Maz inside Ninety. Where it makes sense, Maz runs on frontier models. The choice depends on a host of factors, including cost, speed, and how complex the work is, and the models stay swappable on purpose. What Maz knows about your company, and what it’s allowed to show each person, stays home.

Your agreements should head the same direction. Bright lines on the general-purpose tools that sit outside your system. Clear permissions on the ones inside it.

Write Agreements People Can Follow

“Be careful with AI” isn’t an agreement. Neither is “make sure someone reviews important outputs.” Ten people will read those ten different ways.

A good agreement answers four questions:

• What did we learn?

• What do we do differently because of it?

• When does it apply?

• Who is accountable for it?

Here’s the difference in practice. “Review AI output before it goes to customers” leaves people guessing about what counts as a review and who does it. “Any AI-generated content that goes to a customer is reviewed by the person sending it, and that person owns its accuracy” tells them.

Most agreements should be about that short. If an agreement takes a page to explain, it’s probably two agreements, or it’s trying to do a job that belongs to judgment.

Know Where Agreements End and Guiding Principles Begin

That last point is important because agreements have an edge. An agreement covers the situations you foresaw. It can’t cover the ones you didn’t, and with AI, those show up every week. A new tool nobody has evaluated yet, two agreements that pull people in opposite directions, or a decision someone has to make today with no one around to ask.

You can’t write your way out of that with more agreements. Try, and they’ll multiply faster than anyone can read them. What people need at the edge is a short set of Guiding Principles: rules for how to decide, not rules for what to do.

I'll talk more about Guiding Principles in the next article of this series. For now, the point is the order. Build your agreements first. A principle is a way to decide at the edge of an agreement, and there has to be an agreement for there to be an edge.

Make Every Agreement Easy to Find

An agreement that comes up once in an All-Hands meeting, or sits at the bottom of a long email, doesn’t exist for most of your company. Everyone needs to know the agreement exists, understand why it was created, and be able to find it the moment they need it.

Use the systems you already have. At Ninety, the Knowledge Portal gives every one of our employees and all of our clients one place to keep agreements accessible to the people who need them. It’s where your AI agreements should live and be updated as they change.

Then bring new and updated agreements into the conversations your teams are already having. Add them as Headlines in your weekly meetings with enough context that people understand why the agreement exists. People who understand the why can apply an agreement with good judgment when a similar situation comes up.

Revisit Your Agreements Every Quarter

The agreements you write this quarter won’t be the ones you keep. AI is moving too fast, and your team is going to keep learning. Some agreements will need tighter boundaries because the consequences turned out to be bigger than you expected. Others will loosen as you learn where AI can take on more.

Build the review into your leadership cadence. Make AI a long-term issue you revisit during Quarterly Planning so your leadership team has a regular place to step back and ask three questions: 

1. What did we learn?

2. Which agreements no longer fit?

3. What’s missing?

Turn What You’re Learning Into How You Work

None of us know exactly how our people will use AI over the next several years. But we don’t have to. We can build companies that learn well by giving people room to experiment inside clear boundaries, turning the most important lessons into agreements everyone can follow, and continuing to revisit those agreements as we learn more.

In the next article in this series, I’ll look at what happens at the edge of those agreements, and why AI makes Guiding Principles more important than they’ve ever been.

AI will keep changing, and our companies will keep learning right along with it. We won’t get every decision right the first time, and we don’t need to. If we build companies that learn from experience and turn what they learn into how they work, we’ll be ready for whatever comes next.