Using AI Safely for Comms Work
A practical guide to what to share, what to keep out, and which tools to use for which tasks. It sets out the decision-making framework experienced comms professionals use.
Written by Michael MacLennan, AI Communications Leader of the Year (Agency) 2026. Last reviewed 25 August 2026. Reviewed quarterly against current vendor terms.
"If you wouldn't email it to a journalist, don't paste it into a public AI tool."
Public AI tools (including free tiers of ChatGPT, Claude, Gemini, and others) may use your inputs to improve their models. Even where they don't, you've shared data with a third-party service. Treat them accordingly.
What not to put into AI
This list covers the content types that should never go into a public or personal-account AI tool. They go into an approved tool only where that use has been agreed by your communications lead with your data protection lead, legal adviser or, where there is neither, a named trustee or your sector body.
Embargoed announcements
Unannounced products, funding rounds, partnerships, or executive changes before public disclosure.
Personal data
Information about anyone who can be identified – staff, supporters, customers, journalists, service users – including contact details and performance information.
Pre-announcement financial information
Revenue figures, trading updates, or any information that would be price-sensitive if released early.
Legal documents and advice
Draft contracts, settlement terms, litigation strategy, regulatory correspondence, or legal opinions.
Crisis communications content
Holding statements, incident notes, or any communication related to an active crisis – before the situation has been managed and facts confirmed.
NDA-protected information
Client details, partnership terms, or proprietary information covered by a non-disclosure agreement.
Medical or health information
Patient data, employee health records, or any individually identifiable health information.
Beneficiary, service-user and safeguarding content
Case studies, testimonies or images involving beneficiaries, service users, patients or children, and anything touching a safeguarding concern. Permission to publish a story is not permission to put it through an AI tool; that is a separate decision.
Security incident details
Information about an active security breach, vulnerability, or incident response – even after containment.
Which tool for which task
Not all AI use carries the same risk. Use this framework to decide which type of tool is appropriate for each task.
Public information, general research, non-sensitive first drafts
Research on public topics, drafting from publicly available information, editing generic copy, brainstorming messaging frameworks, generating placeholder content.
Tool: Any AI tool (ChatGPT, Claude, Gemini, Copilot, etc.) – free or paid tier.
Internal documents, draft strategies, non-sensitive stakeholder communications
Internal briefings, communications strategies, messaging frameworks containing non-public company direction, stakeholder mapping containing role-level (not personal) information.
Tool: Enterprise version of your preferred AI tool (Claude for Enterprise, ChatGPT Team/Enterprise, Copilot for Microsoft 365). Confirm your organisation has a Data Processing Agreement in place and that inputs are not used for model training.
Crisis communications, M&A, regulatory, board-level, legally sensitive
Security incident response, crisis holding statements, M&A communications, regulatory submissions, board communications, personal data, legal content.
Tool: Enterprise tool with a DPA in place, for a use agreed in advance with your data protection lead or legal adviser. Consider whether AI is appropriate at all – some high-risk content is better drafted without AI assistance and reviewed by a human expert from the start.
Quick triage guide
Match your situation to the recommended approach before reaching for an AI tool.
| Situation | Recommended approach |
|---|---|
| Generic drafts, public information, no sensitive content | Public AI tool acceptable |
| Internal comms, named colleagues, non-public announcements | Enterprise tool with a data processing agreement, approved for this use |
| Crisis content, legal/regulatory matters, financial information | Approved tool, for a use your comms lead has agreed in advance with your data protection lead or legal adviser |
| Embargoed announcements, M&A, security incidents | Only in a tool approved for this use, with sign-off before anything goes in. With no approved tool, draft without AI and review with your legal adviser. |
| Unsure | Treat as the row above – ask your communications lead before anything goes in |
Quick check before you paste
Run through this before putting anything into an AI tool you're not certain is enterprise-secured.
Copy this checklist
[ ] Is any of this not yet public?
[ ] Does it name or identify a person?
[ ] Is it covered by an NDA or a confidentiality obligation?
[ ] Does it relate to an unannounced product, deal or event?
[ ] Would it cause harm if it appeared in a news article tomorrow?
[ ] Is this kind of content outside your organisation's AI use policy, or do you not have one yet?
If you answered yes to any of these, use an approved tool for an approved use, or don't use AI for this task. Unsure means yes.
If you think personal data or confidential material has gone where it should not, stop using that tool for the task and tell your communications lead straight away, the same hour if you can: some incidents have to be reported to the ICO within 72 hours, and the clock starts when your organisation finds out.
What about the tools on this site?
The same standard applies to us. Plan Comms With AI and Message Test send what you type or paste to Claude, the Anthropic model that produces the result. We do not store it, and under Anthropic's API terms it is not used to train models. If you use Plan's voice option, the recording goes to OpenAI for transcription first and is not stored. The one thing we do keep is a Foundations profile, if you save one: it is held in our database in the UK and sent to Anthropic with each Plan or Message Test you apply it to, so write it as you would anything else you put into an AI tool. Saved plans live in your browser only. The detail is in the privacy notice.
Even so, treat Plan and Message Test like any other AI tool: describe your situation in general terms, and leave out named individuals, embargoed detail, and anything covered by the checklist above. A plan built from "we are preparing for a sensitive restructure announcement" is just as useful as one built from the specifics.
If your organisation doesn't have an AI policy yet
Most organisations are behind on this. An AI use policy for communications teams doesn't need to be a 40-page governance document – it needs to answer four questions clearly:
- Which AI tools are approved for use, and which are not?
- What categories of content can and cannot go into AI tools?
- Who owns responsibility for reviewing AI-generated output before it goes live?
- What is the escalation path if someone is unsure?
You do not need to buy anything to answer them. The one-page AI acceptable use note answers all four and is the place to start. When you need the full version, three free templates cover the same ground and are built to work together:
- AI Use Policy (Living) – the policy itself, written as a document you keep current rather than one you file. About 45 minutes.
- AI Tool Register – what is approved, what is on trial, what was declined and why, so the policy never hardcodes a tool list. About 40 minutes.
- AI Use Review Log – the review cadence that stops the other two going stale.
The Manage AI Comms Toolkit sequences all three into one workflow, with the review rhythm and a leadership assurance note. Start there if you would rather follow a path than pick templates.
If you want to talk it through with someone before you write it, a Consult Comms With AI session is 45 minutes plus a written advisory note, £300. If you want the policy, register and tool instructions written for your organisation rather than adapted by you, that is Deploy Comms With AI.
You can also request a template – if there's a specific AI governance tool you're looking for that isn't on the site yet, we'll consider building it.