What guardrails should a CPG founder set up before using AI with sensitive company information?

Picture a Monday numbers meeting. Your bookkeeper wants to paste last month's bank statement into an AI tool to get the numbers ready, and you cannot say yes or no, because nobody ever told you which company information is safe to share. The answer is to sort your data into four tiers: public, internal, confidential, and never share. Then match each tier to one setup. Public data can go into a free chat. Internal data needs a paid business account with training turned off. Confidential data needs that account plus a read-only link (the AI can look at your files and cannot change them), limited to the folders it needs. Never-share data stays out of AI entirely. That bank statement is confidential: fine on a business account with training off and a read-only link, never in a free chat.

Three terms first. None of this is a job for you personally: your bookkeeper or whoever runs your software does the setup once. A free chat is the personal version of an AI tool, the kind you join with an email address. A business account is the paid team version, where the company's written terms say what you type is not used for training. Training means the company uses your conversations to teach the next version of its AI.

The decision tree

Take one piece of data, for example a supplier price sheet. Start at the top and stop at the first yes. The supplier price sheet lands in step 2: tier three, because a leak would hurt you but would not break the law.

1. Could a leak break a law or a contract, or end a deal? Think formulas, passwords, payment card numbers, and anything under an NDA (a contract promising to keep something secret). Tier four, never share. No AI at all.
2. Would a leak hurt you? Think retailer terms, margins by product, customer lists, bank statements. Tier three, confidential. A business account with training off, plus a read-only link limited to the folders the job needs.
3. Would it be awkward but not harmful if a competitor read it? Think meeting notes, a draft promotion calendar. Tier two, internal. A business account with training off.
4. None of the above? Think website copy, published prices. Tier one, public. Any setup is acceptable, including a free chat.

By hand, you would open every folder, list every file, and sit with your operations lead deciding which tier each one belongs in. That takes a few evenings and ends in a spreadsheet. Or you can give the instructions below to Claude, an AI assistant you type to like a text conversation, and have a first draft in thirty seconds. You paste them exactly as they are. Claude does the sorting and you read the result.

I run a seven-figure consumer packaged goods company. List the 30 kinds of data a company like mine keeps, such as supplier price sheets, retailer terms, customer emails, margins by product, formulas, bank statements and logins. Sort each into one of four tiers. Public: already visible to anyone. Internal: awkward but not harmful if a competitor read it. Confidential: harmful if leaked. Never share: a leak could break a law or a contract or end a deal. Put the result in a table with three columns: the data, the tier, and one sentence on why. Then say which setup is acceptable for each tier: a free chat, a business account with training turned off, a business account plus a read-only link, or no AI at all. Finally, ask me which kinds of data I keep that you missed, and add them to the table.

Which company data should never go into an AI tool?

Anything where a leak breaks a law, breaks a contract, or ends a deal. Formulas, passwords, payment card numbers, and anything a retailer made you sign for sit here. Tier four should be short. If it is long, the tools will sit unused.

What settings turn off training on your data?

Two places. On a business account, the terms usually say your content is not used for training. Read that line in the terms for your plan before you rely on it. On a free chat, there is often a setting you switch off yourself, and its name differs by company and changes over time. Look in the account settings for wording about training or improving the AI. If you cannot find the setting, treat that chat as tier one only.

Should an AI connection be read-only?

For anything in tier three, yes. A connection is a link between the AI and your files or your sales and accounting systems. Read-only means the AI can look and cannot change, delete, or send anything. Whoever sets up the link picks this once, usually as a permission box they tick. It does not stop the AI from seeing what the link can reach, so the person setting it up should point it at the folders the job needs and nothing else. What read-only does prevent is the AI overwriting your numbers or emailing a customer by mistake. Add the right to change things later, for one named job, once you trust the setup.

How do you check what an AI actually touched?

You ask it to list what it used, and you read the list. A session is one conversation with the AI, from the first message to the last. Whoever uses the tool pastes the rules below at the start of each one, or saves them in the tool's instructions box if it has one. The rules work in most tools that let you type instructions.

Follow these rules in every session with me.
1. Before you read or use any file, tell me which tier it falls in: public, internal, confidential, or never share.
2. If something is never share, stop and tell me. Do not read it, summarize it, or repeat it.
3. Only read. Do not change, delete, send, or post anything unless I type the word approved for that one action.
4. At the start of the session, ask me whether this chat is on a business account with training turned off. If I say no, treat everything confidential as off limits.
5. At the end of the session, list every file, spreadsheet, and connection you read, and anything you changed. If you changed nothing, say so.

Rule five gives you a short list to read instead of a system to audit. An AI can get that list wrong, so treat it as a first check. A second AI tool can read the list and tell you whether anything falls outside the tier you approved, as covered in ai-that-audits-the-ai.

Where to connect first

Reporting is a sensible first job, because the Monday numbers are work an AI can take over. Start there, read-only, with the rules above. Related reading: which-number-is-right for when two reports disagree, and ai-starts-with-the-reporting-bottleneck for why reporting comes first. The service page for this work is the Reporting Sprint.