Plan requirement
| Subscription | Suite Professional or higher, Support Professional or higher |
| Also required | Copilot |
| Access | Admin |
Telling the AI how your company writes. Two things shape the work: guidelines apply per brand, and they govern how something is said while procedures govern what to do. Keeping those apart is the whole trick.
Set them
- In Admin Center, click AI in the sidebar, then select Agent copilot and Writing.
- Select Communication guidelines.
- Click Manage guidelines and write your rules.
- Optionally select one or more brands where they apply. If a brand is already connected to other guidelines, a message tells you.
- Look at a few suggestions to hear the effect, then adjust after the team has used it for a week.
They change the rewrite in your tone tool and auto assist, and when enabled they take priority over the style the system would otherwise infer from the conversation.
Try them in a sandbox first
Since 14 August 2026 communication guidelines are part of version management, along with the rest of Agent Copilot: auto assist settings, procedures, suggestions and the writing tool. So you can write them in a sandbox, hear what they do, and deploy them once you are happy, rather than editing the wording your agents are working with while they are working with it.
A disclaimer is a guideline too
If you want customers told that an answer came from AI, that is a rule like any other: instruct the agent to end its message with the line you want. You can scope it, which matters here, because a rule aimed at generative replies leaves your scripted dialogue replies alone.
Worth deciding deliberately rather than by default. A disclaimer on every reply becomes wallpaper; on the generated ones it still means something.
Where they work
On any channel within Agent Workspace, and on the languages OpenAI supports. So a team writing to customers in Dutch and English is covered by one set of guidelines, and the channel a reply leaves by does not change whether the rewrite happens.
Where the guidelines stop
On an email AI agent you can set your own opening and closing lines. Those are sent exactly as written: not generated, not translated, and not touched by your guidelines, your tone of voice or your formality setting. So a formal register everywhere else and a breezy opening line is a combination you can produce without noticing.
They are per language too, and they are not translated for you, so each language needs its own. Leave them unset and the agent writes the openings and closings itself, in the tone you configured.
Guidelines are not procedures
Zendesk's own advice is to keep them separate: a guideline says how to say something, a procedure says what to do. A rule like "always check the order status first" belongs in a procedure, and putting it here dilutes both.
The test is simple. If following the rule changes the outcome of the case, it is a procedure. If it only changes the wording, it is a guideline.
Be concrete
"Friendly and professional" produces nothing, because it describes every support team ever. "Short sentences. Address the customer informally. Never apologise twice. Always end with what happens next" produces a recognisable style.
Write each rule in simple, direct words, and add a short example or explanation so the system understands what you meant. Vague terms and jargon are read as loosely as you wrote them.
Take the rules from your best replies
Read ten replies from the agent everybody agrees writes well, and write down what they do. That is your guideline, and it is far more accurate than anything composed in the abstract.
Say what to avoid
Jargon, internal terms, hedging, over-apologising. Prohibitions are easier for a generative system to follow than aspirations, and they are what stops the corporate register creeping back in.
Cover formality explicitly
Guidelines refine tone, length, formality, structure, formatting and word choice, so in languages that distinguish formal and informal address, say which you use. Getting this wrong is immediately noticeable to a customer and invisible to somebody reviewing the settings in another language.
Per brand means per voice
Because a set of guidelines can be attached to specific brands, two brands with genuinely different voices can each have their own. That is worth using where it is real, and worth resisting where the difference is only a logo, since two sets are two things to keep in step.
Revisit after the first week
Agents will report suggestions that sound wrong. That feedback is exactly what the guidelines are for, and it is worth one revision after real use.
Comments
0 comments
Article is closed for comments.