Guides
AI-assisted business writing

Writing an AI policy for a communications team

A communications team’s AI policy works best when it’s short and settles only what matters. We look at what it should require and what it can leave to the team’s discretion.

Many communications teams already use models, whether or not anyone has given them permission to. The question a policy needs to address is what the organization requires of the documents that result. A good AI policy for a writing team resolves that clearly and leaves the details to the people actually doing the work.

Length matters more than you might expect. A policy that runs to a dozen pages tends to be skimmed once, filed, and forgotten. One that fits on a page can be remembered and used as a checklist. For a team that writes reports, proposals, press releases, and web copy, a single page is usually sufficient.

Decide what the policy is really for

Name the risks the policy guards against at the top, and check that each rule traces back to one of these priorities. For a communications team, there are usually four. Accuracy comes first: nothing should be released without being verified. Confidentiality keeps client and internal material from leaving the organization through a prompt. A consistent voice makes every document recognizably the organization’s. Accountability means that a named person stands behind every document and can explain how it was made. If a proposed rule doesn’t protect accuracy, confidentiality, voice, or accountability, you probably don’t need it.

What to require

Start by requiring a named author for every document: the person whose name goes on it, who has read every line and can answer for its contents. This is the most important rule in the policy. A model can produce a draft, but it can’t be held accountable for one, so the author becomes the person responsible for the document’s accuracy and voice.

Next, require every factual claim to be verified before release. Figures, dates, names, quotations, and references should all be checked against a source that a person has read. The policy should name the person responsible for the checking and say when in the process it happens. A rule that says “everyone checks” leaves the job to no one in particular. If your team produces documents with references, make checking that each reference exists a named step with an owner.

The policy also needs a clear rule on confidentiality that specifies which material may go into which tools. One workable version permits public and general internal material in approved tools. It excludes client material, personal data, anything under an NDA, and anything you wouldn’t email to a stranger, unless the tool is covered by a contract that protects it. Name the approved tools as well as the person who can add new tools to the list. That way, the policy can accommodate a better tool when one appears.

Require a record of how each document was made. This can be a habit more than a form: the drafting history stays in the file, and if a model drafted a substantial part of the text, a line in the project notes says so. If a client or a regulator asks, the answer is then readily available. Where the model you use marks its output with a watermark, this record also shows something the watermark can’t, which is how much of the thinking was done by the people on your team. Researchers studying detection tools have argued along similar lines for judging work by a record of how it was produced instead of by detector scores (Kim, 2025).

Finally, write the house voice down. Where the organization already has a style guide, the policy can refer to it and state that model output is edited to match. If you don’t have a style guide yet, your AI policy is a good reason to create one. Our guide to keeping your house voice when a model writes the drafts explains what the description should cover. A team that uses models without a written voice tends to drift toward the model’s way of writing and can end up sounding like other organizations doing the same thing.

What to leave open

A list of the tasks a model may be used for becomes outdated quickly. It can also encourage people to conceal their use of a tool. It’s more durable to specify what has to be true of the output, however it was produced, and to trust the team to judge where the tools are helpful and where they’re best avoided. If you’d like to offer guidance on that, our guide to where AI helps in a business writing workflow is a reasonable place to start.

A limit on how much of a document a model may write is just as difficult to enforce. It also measures the wrong thing: even a document that a model drafted almost entirely can be perfectly acceptable if it has been verified, rewritten where it matters, and adopted by a named author. On the other hand, a single model-drafted paragraph can do real harm if it contains an unverified figure or an invented reference. A percentage limit wouldn’t flag that paragraph, whereas the verification step and the named author should catch it.

Leave detection tools out as well. A policy that requires documents to pass a detector risks penalizing staff for their own work. Detector results vary substantially from one tool and setting to another (Maghami et al., 2026), and the tools are least consistent with mixed and edited text, which is what AI-assisted documents usually are (Van Vlasselaer et al., 2026). Our guide to what AI detection scores can and can’t tell you explains why. Rely on the author and the verification step instead.

Disclosure

Whether the organization tells clients and readers that models were used is a separate decision. Our guide to whether your firm should disclose AI use to clients examines it in detail. The policy should state your position on disclosure and say who can make an exception for a particular client. It should also note whether any client contract or regulation you work under requires disclosure.

Keeping it up to date

Give the policy a date and an owner, and review it on a fixed schedule instead of waiting for something to go wrong. The tools will always change faster than the document can, so describe the outcomes you need without naming specific products. When it’s time to review the policy, ask your team members what they’re actually doing and update it to match.

If your team works to a written standard and produces documents throughout the year, a group account can help. The same small team of editors applies your house style to everything you send, however the first draft was produced. Don’t have a house style yet? We can help with that too.

References

  • Kim, H. N. (2025). Detecting the undetectable: The need for a new paradigm for academic writing evaluation in the AI era – Addressing inconsistencies in AI and plagiarism detection tools. In C. Stephanidis, M. Antona, S. Ntoa, & G. Salvendy (Eds.), HCI International 2025 posters (Communications in Computer and Information Science, Vol. 2523, pp. 229–239). Springer Nature. https://doi.org/10.1007/978-3-031-94153-5_22
  • Maghami, M. R., Ho, C. K., & Camilleri, P. (2026). AI content detection in student text assignments, false positive rates, and academic integrity. Discover Education, 5, Article 1192. https://doi.org/10.1007/s44217-026-02307-x
  • Van Vlasselaer, M., Van Droogenbroeck, F., & Spruyt, B. (2026). Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity, 22, Article 16. https://doi.org/10.1007/s40979-026-00226-w

Working to a policy?

Tell us the rules your team writes under, and we’ll agree the scope with you before quoting.

Get a quote