What Not to Paste: Customer Data and the AI Tools Your Team Already Uses

Nobody at your shop made a decision to start feeding customer data into AI tools. It happened the way most things happen: someone pasted a difficult email into a chatbot to get help writing a reply, it worked, and they did it again. Multiply that by a front desk, a dispatcher, and two salespeople, and you have a data practice you never wrote down. This is not an argument for banning the tools — they save real time. It is an argument for deciding, on purpose, what goes in and what stays out.

Assume anything you type leaves the building

The useful mental model is simple: pasting text into an AI tool is closer to emailing it to a vendor than to typing it into a document on your own computer. It goes to a company you have a contract with, or more often no contract at all, because someone signed up with a work email and a credit card. Consumer tiers of most AI products are governed by consumer terms. Business and enterprise tiers usually come with different commitments around retention and training, and some vendors will sign a business associate agreement for regulated data. Those are not the same product, and the difference is invisible from inside the chat window — the interface looks identical whether or not your data is covered.

So the first question is not what should we paste. It is which account are we pasting into. If your team is using free personal accounts for work, the answer to everything downstream is already constrained, and no policy you write will change it. Fixing the account tier is the cheap, one-time move that makes every later rule enforceable.

Three categories that should never be in the prompt

You do not need a twelve-page policy. You need three bright lines that a new hire can remember on their second day.

Identity that lets someone be found. Full name plus home address, plus phone, plus the fact that they are away from the property next Tuesday for the install — that combination is a different object than any one piece of it. Roofing, HVAC, and solar work runs on customer addresses and access windows, so this is the category most likely to slip through by habit.

Health information. For med spas this is the sharp one. Treatment histories, contraindications, medication lists, before-and-after photos, anything a patient told you in confidence about their body: none of it belongs in a general-purpose AI tool unless you have a signed agreement covering it. In the United States this is HIPAA territory, and in Canada it falls under PIPEDA and provincial health privacy law. The practical test is not whether the note looks clinical. It is whether a stranger reading it would learn something about someone's health.

Payment and identifiers. Card numbers, bank details, government ID numbers, financing application data. These have no upside in a prompt. An AI does not need a card number to draft a receipt email; it needs the last four digits at most, and usually not even that.

Everything else — the shape of a complaint, the timeline of a job, a price objection, a tone problem in a review — is generally fine, and that is most of what the tool is actually good for.

Strip it down before you paste it

The reason people paste raw customer data is that stripping it feels like extra work for a worse answer. It usually isn't. The AI is not helping you because it knows the customer's name. It is helping you because it understands the situation.

The habit to teach is substitution, not deletion. Replace the name with "the customer," the address with "the job site," the specific treatment with "the procedure," and keep every detail that actually drives the answer: what went wrong, when, what you already offered, what you are willing to do, what tone you want. A prompt that reads "customer booked a Tuesday install, we missed the window twice, they've asked for a discount, we can offer a partial credit but not a refund — draft a reply that is apologetic without admitting fault" produces a better draft than the same story with a real name attached, because you were forced to state the constraints explicitly.

The one place this breaks down is documents. People paste whole invoices, intake forms, and PDF exports because retyping the relevant part is tedious. That is where most accidental disclosure happens. If your team is regularly feeding whole documents into a tool, that is a signal you need a connected system with proper data handling, not a better paste habit.

Automate the boundary, keep the judgement

What to automate: the account setup, so everyone is on a business tier with retention settings you chose rather than inherited. The templates, so the common tasks — review replies, quote follow-ups, appointment reminders — run through prompts you wrote once with the placeholders already stripped, instead of being improvised from a live customer record every time. The logging, so you can see what your tools were asked and what they answered, which matters far more than most owners expect the first time a customer disputes what they were told.

What stays human: deciding which vendors you trust, and reading the terms of the tier you are actually paying for. Deciding what happens when a customer asks you to delete their data and it may sit in three systems. And every judgement call at the edge, which in practice means the ones involving health, money, or a threat of legal action — the exact three cases where a fast, fluent, confidently wrong draft costs the most.

The businesses that get this right are rarely the ones with the strictest rules. They are the ones where the safe path is also the fast path, so nobody has to choose between doing the job well and doing it correctly.