Yes, by building a Copilot agent that holds the voice and publishing it to everyone in the organisation. A declarative agent in Microsoft 365 Copilot carries its own instructions and knowledge, an owner can share it across the tenant, and an admin can approve each published version. What it can’t do is check the drafts people write with it, so the voice stays consistent only as far as each person checks their own work.
How an agent holds a voice
Microsoft’s Agent Builder, inside Microsoft 365 Copilot, lets anyone with a Copilot licence build a declarative agent without code. The agent has a name, a description and up to 8,000 characters of instructions, which is where the voice description goes: how you open, how you close, the words you use and avoid, how you treat figures.
Its knowledge can be public web pages (four), SharePoint files (up to 100), OneDrive files (up to 50), embedded files (up to 20) and Teams chats and meetings. That’s where the examples of your writing and your approved facts go. The agent respects the permissions on those files, so a person who can’t open a document in SharePoint won’t get its contents through the agent either.
Agent Builder can’t call outside services. For an agent that talks to another system, Microsoft points you to Copilot Studio, which supports connectors and MCP tools and is licensed separately: $200 a month for a pack of 25,000 Copilot Credits, or no extra charge for agent use by people who already hold a Microsoft 365 Copilot licence, within fair-use limits.
Getting it to everyone
A new agent is private to the person who built it. The owner can turn on org-wide sharing, after which anyone in the tenant can find it in the Agent Store, or submit it to the organisation’s catalogue, where an admin reviews it and deploys it to the whole organisation or to named groups.
The catalogue route is the one that gives you control. Microsoft tracks the shared version and the catalogue version separately, and changes to the catalogue version apply only after an admin approves a resubmission, so people who installed it from the store keep seeing the last approved version. Agents can have several owners with equal rights.
Microsoft is also previewing Copilot Tuning, which trains an agent on an organisation’s own documents. Its style-editing template is built for brand tone and needs at least 20 sample files. It’s early access for a limited set of customers, and tuned agents don’t yet take follow-up prompts.
Where it stops working
The agent gives Copilot the voice, and nothing reads what Copilot writes back against it. Microsoft’s pages describe how instructions and knowledge are supplied, and in Copilot Tuning they’re explicit that you’re responsible for verifying the accuracy of any outputs. A colleague who writes a proposal with the agent and never reads it against the brand description has produced a proposal in whatever voice the model felt like that day.
Facts in the agent’s files go stale, and the only way to notice is to read them. And the agent lives in Copilot. The person who drafts in Claude at home or ChatGPT on their phone has none of it.
What people get wrong
They write the instructions as a description of the brand (“warm and expert”) rather than as rules a model can follow and a person can check. They point the agent at the whole SharePoint, which includes every out-of-date deck ever saved. And they share the agent once and treat it as done, with no owner responsible for the next change.
How BeigeCheck does it
BeigeCheck is a Copilot agent your IT team switches on once for everyone, on Business and above. It carries your approved company voice, glossary, house rules and facts, and the same knowledge reaches Claude and ChatGPT, so the voice is one thing in three tools rather than three agents maintained by hand.
Each person also keeps their own voice, built only from writing they’ve done themselves, so sharing a company voice doesn’t make everyone sound identical.
After Copilot writes, the Beige-ometer checks the draft: how authentic it reads, how well it fits the voice, the sentences to change with a fix for each, and every claim against the facts you’ve approved. When people keep making the same edit, BeigeCheck proposes a change to the voice, an approver decides, and every change is a version with a name and a date.