Five decisions to make before you write an AI policy

At one manufacturer I work with, someone used a chat tool to turn company material into an instruction manual, drawings included. The manual was good. Nobody could say which tool made it, whether it ran on a paid account, or where the source material ended up once it was uploaded.

The quality is the reason to pay attention. Useful work continues and spreads, and this manual will not be the last one. In my experience that is how AI arrives at most companies this size – through individual initiative, long before anyone writes a rule about it. The people doing it are trying to work faster. None of them are trying to create risk.

So when an owner asks me for “some parameters around AI,” what they usually want is enough structure that the company is protected, without making it so rigid that people cannot do their jobs. That is the right goal. Five decisions get you there, and the policy is how you write them down.

What is at stake for a manufacturer

Product drawings and designs. Customer information, and anything held under a customer’s confidentiality terms. Pricing and cost data. Supplier terms. Employee records.

Company material put into the wrong tool, on someone’s personal account, is outside your control from that point forward.

Start with what is already happening

You are not starting from a blank page. The first job is an honest picture of current use – which tools, whose accounts, what information is going in, and what is coming back out. Every decision below is easier once that picture exists, and none of them can be made well without it.

The five decisions

These belong to leadership rather than IT or an outside advisor, because each one is a judgment about how much freedom to trade for how much protection.

Each of these is a setting, and the right setting can differ by situation. You can be permissive about using AI to draft an internal email and strict about anything involving a customer’s drawings. That balance is the difference between rules people follow and rules people work around.

Decision one – what information can go in

Three lists, short enough that people remember them.

Open. Information that would not hurt you if it were public. Marketing copy, general industry questions, help writing a routine email.

Restricted. Company information that is fine in an approved tool on a company account, and nowhere else. Product descriptions, internal procedures, quotes, supplier correspondence.

Never. Information that does not go into any AI tool without a specific decision. Customer drawings and designs held under confidentiality terms, cost structures, employee records, and anything a customer contract restricts.

The hard line is customer-supplied material. If a customer sends you a drawing under an NDA, putting it into an AI tool may breach that agreement however secure the tool is. That is a contract question first and a technology question second.

Two more things to settle here. Whether the tool stores what people put in, and whether the vendor trains on it, since business accounts generally handle this differently from free personal ones. And what someone should do when they are unsure. Every workable policy has an easy answer to that, usually one named person to ask.

Decision two – which tools, on whose account

One owner asked me whether AI should be part of the company the way a laptop is. That is the right instinct, and it turns out to be mostly a licensing question.

On a personal account, you have no visibility into what went in, no way to remove access, and no claim on the output. On a company account, all three change.

A workable setup is one general assistant plus whatever AI is already built into systems you own. That second category is growing on its own, because ERP and other business platforms keep adding AI features that arrive without anyone requesting them. A short approved list is easier to support than an open field. It also needs a request path, because if there is no way to ask for a new tool, people will buy it themselves, which is how this started everywhere it exists.

In my experience, consolidating scattered personal subscriptions onto company licenses comes out close to cost-neutral, and it buys control.

Decision three – drafting versus acting

AI that helps a person produce something is a productivity question. AI that does something on its own inside your systems is a control question.

Most use today is drafting. Someone asks a tool for help and decides what to do with the answer, so a person is always in the loop, which limits how wrong things can go. Acting is different. An automation that reads an email and updates a record, or a system that answers a customer before anyone sees the reply, carries the company’s authority. That can be entirely appropriate. It is also where mistakes stop being private.

The workable rule is that AI drafts by default, with a short list of cases where it may act. Repetitive, high-volume, low-judgment work is the usual candidate. Each of those cases gets its limits written before it goes live – what it can touch, whether it can change data or only read it, where a person approves in the middle of the sequence (anything touching money, customers, or production), and how you would know afterward what it did.

Decision four – what gets checked before it leaves

AI output is confident whether or not it is correct. It will produce a specification, a dimension, or a policy statement that reads perfectly and is wrong. For a manufacturer, that matters most where the output becomes something physical or contractual.

Go back to that instruction manual. The drawings were good. If they had been subtly wrong and reached a customer, who would have caught it, and at what stage?

The categories that need a second set of eyes are usually easy to name – customer-facing documents, specifications and drawings, quotes and pricing, and anything that becomes a work instruction on the floor. The harder part is making sure the reviewer has the expertise to catch a plausible-sounding error, and deciding whether AI-assisted work should be marked internally so reviewers know to look harder. Some customer contracts already require you to disclose when AI contributed to what they receive.

Internal drafts, first passes, and brainstorming do not need a review gate. Putting one there is what makes people bypass the whole policy.

Decision five – the off switch

When someone leaves, their laptop comes back and their accounts are disabled. AI tools should work the same way, and on personal accounts they cannot.

An off switch that works needs four things to be true. The account belongs to the company, someone knows it exists, that person has administrative control, and access can be removed without the user’s cooperation. From there it is ordinary discipline. AI tools go on the onboarding and offboarding checklists, each tool has a named owner and a backup, and someone has actually tested how fast access can be shut off.

If a customer asked you in writing who has access to the material they supplied, could you answer it today?

Data quality sits underneath all five

AI output is only as good as the information it works from. A model asked about your products gives better answers when product data is clean, complete, and consistent, and no amount of governance changes that.

Many manufacturers are already doing this work inside an ERP rollout, bringing naming, descriptions, and structure into order. That effort decides whether AI is useful in your business at all, which makes the ERP data work and the AI program the same investment.

The open question is everything outside the ERP. Quotes, drawings, procedures, customer correspondence, and supplier terms mostly live in files and inboxes. Whoever owns data quality in the ERP is the natural starting point for the same question everywhere else.

What the program looks like

Once the five decisions are made, the program itself is small. For a mid-market manufacturer it is six pieces.

Build the picture first, because decisions grounded in what is actually happening are better decisions. Make the five decisions in a single working session with leadership. Then write the policy and set up the accounts and controls, which is mostly administrative work. The whole thing takes weeks. The slow part is scheduling.

The picture needs more than leadership in the room. Spend time with the people who actually use these tools. In my experience the usage picture is wider than leadership expects, and that is no criticism of anyone.

How you know it worked

A policy written without that picture, or written to cover every theoretical risk, gets ignored quickly.

If you want a place to start, answer one question with your leadership team this week. If the choice were between someone being slowed down and something leaving the building, which error is worse?