AI and Client Confidentiality: What Comms Teams Can and Can't Put Into a Chatbot

A communications team handles the most sensitive material a client owns — unannounced news, crisis details, legal exposure, executive movements — and that material now routinely gets pasted into AI tools to speed up a draft. Most of the time nothing goes wrong. The exposure isn't in the average case. It's in the one document that shouldn't have gone in.

Here is the working rule, and the line between what's safe and what isn't.

Quick answer. Treat every consumer AI tool as a public surface until proven otherwise. Don't paste in embargoed news, deal or financial details, unannounced crises, executive personnel matters, contracts, or personal contact data. Public, already-published material is generally fine. What changes the math: turning off model training, and moving to enterprise or team tiers with stronger data terms.

The real risk — in plain terms

Two things can happen to text entered into a consumer AI tool. It can be retained on the provider's systems. And, depending on the plan and its settings, it can be used to train future models. Neither means a leak is likely. Both mean the material has left the team's control and entered a system the team can't audit. For routine work, an acceptable trade. For a client's most sensitive information, it isn't.

What not to put into a consumer AI tool

• Embargoed or unannounced news — a launch, a hire, a deal not yet public.

• Deal and financial material — anything touching a transaction or non-public financials. This can carry legal weight beyond a confidentiality concern.

• Active crisis details — facts about an incident before the team has decided what's public.

• Executive personnel matters — departures, investigations, disputes.

• Contracts and legal documents — client agreements, settlements, anything privileged.

• Personal contact data — reporter and contact lists, private numbers, home addresses.

What is generally fine

• Information already public — published coverage, live site copy, issued releases.

• Drafting and editing with the sensitive specifics removed or genericized.

• Hypotheticals, structure, and format — "outline a crisis statement for a product recall" without the real product or facts.

• Research on public companies, public figures, and published material.

The practical move: genericize before you paste. A model can structure a crisis statement, sharpen a message, or tighten a release just as well with placeholders as with real names. The client's actual facts don't need to be in the prompt for the tool to be useful.

Consumer tier vs. team tier — what actually changes

Data terms and defaults change — confirm current terms with each provider before relying on them. The takeaway holds regardless: for an agency handling client material at scale, a team or enterprise tier is the baseline, not an upgrade. Consumer logins are the wrong surface for client work.

A simple team policy — paste-ready

AI tool use — client material. Before entering client information into any AI tool: (1) If it is not yet public, do not paste it — genericize it or leave it out. (2) Confirm model-training is off on the account. (3) Use the firm's approved team/enterprise accounts, not personal logins. (4) Check the client's own AI policy and follow it where it is stricter than ours. When in doubt, treat the tool as public and leave the sensitive detail out.

That policy costs nothing, takes a paragraph, and closes the most likely way an AI tool becomes a client problem. Build the infrastructure before the crisis — not during it.