What Your Team May and May Not Put Into AI

Two bad answers

Most organisations land on one of two rules, and both fail.

The first is "do not put anything confidential into AI". It sounds responsible and means nothing, because almost every useful piece of work at your company is confidential to somebody. Applied literally it bans the work. Applied loosely, everyone draws their own line and you have no rule at all.

The second is "use your judgment". That is not a policy, it is a shrug in a font.

What people actually need is a short red list they can remember, plus permission to do the thing that makes most of the red list irrelevant.

The red list

Five categories. Keep it to five, because a list of fifteen is a list nobody recalls at the moment it matters.

Client or customer identifiable data. Names, contact details, account numbers, case details specific enough to identify someone.

Staff personal data. Performance notes, salary, health or wellbeing information, home addresses, anything from an HR file.

Credentials and access. Passwords, API keys, tokens, internal system links, anything that opens a door.

Unreleased financials and commercially sensitive plans. Results before publication, pricing strategy, deal terms, acquisition or restructuring plans.

Anything covered by a confidentiality agreement or a client contract. If a client made you sign something about their information, that information is theirs to govern, not yours.

The middle path most policies miss

Here is the part almost every policy leaves out, and it is the part that decides whether your rules get followed.

Most work that touches the red list does not need the identifying details at all.

You want help drafting a difficult reply to an unhappy customer. The AI does not need her name, her account number or her address. It needs to know that a customer has complained about a delayed delivery for the second time and is asking for a refund you cannot fully give. Call her Customer A. The draft comes back just as good, and you paste the name in yourself afterwards.

You want a summary of a messy internal dispute between two colleagues. Person A and Person B. You want a pricing structure sense checked. Product X at price point Y. You want a contract clause explained in plain English. Strip the party names first.

The rule is short enough to say in a corridor: if it does not need the name, do not paste the name.

Below is some text I need help with, but it contains details I should not share.
Rewrite it as a version I can safely work from:

Replace every person's name with Person A, Person B and so on
Replace every company or client name with Client A, Client B and so on
Replace account numbers, emails, addresses and reference codes with [REDACTED]
Keep amounts only if they are essential to the question, otherwise use [AMOUNT]
Keep the meaning and the tone intact
Then list what you replaced, so I can put the real values back afterwards.
Text:
[paste it]

Consumer and business plans are not the same thing

This is where teams get caught out. The same brand name can behave very differently depending on which plan you are on and how the account is configured. Consumer tiers and business or enterprise tiers commonly differ on how long inputs are kept, whether inputs may be used to improve models, whether an administrator can set policies centrally, and what contractual commitments the provider makes to your employer.

We are not going to state any vendor's terms here, because they change and because yours may have been negotiated. What matters is the habit: your team must check the plan they are actually using and the rules their employer has actually set, rather than repeating something a colleague heard about a different product on a different plan.

I am writing a short AI usage guide for my team of [number] people in [function].
Draft a list of no more than ten questions I should ask our IT, legal or
compliance contact before I publish it. Focus on:

Which tools and which specific plans are approved for us
What our agreements say about retention and use of what we type
Whether client information may ever be entered, and under what conditions
Who to contact when someone is unsure or has made a mistake
Phrase them as plain questions. Do not answer them yourself and do not
guess what any vendor's terms say.

Make the safe path the easy path

A rule people cannot follow is a rule that gets routed around. If your policy says client information may never be entered but gives no anonymising habit, your team will either stop doing useful work or quietly do it anyway on a personal account. Give them the placeholder habit and the safe path becomes the fast path too.

Here is the list of things my team must never paste into an AI tool:
[your list]
And here are the five tasks they most want to use AI for:
[task 1]
[task 2]
[task 3]
[task 4]
[task 5]
For each task, tell me:

Whether it can be done at all under my list
Whether replacing names with placeholders would make it workable
Where my list is so broad that people will simply ignore it
Be blunt about the last one.