Your next author won't open your CMS
The best content ideas come from the people doing the work, and most of them never get written. We're designing an agent that meets our experts in Microsoft Teams and turns a five-minute conversation into a sourced draft.
Ask any marketing team where its best content comes from, and it will name the same people: the consultant who just untangled a client's integration, the designer who watched a usability test go sideways, the engineer who found a better way to migrate ten thousand pages.
Then ask how often those people write. Almost never. It isn't that they have nothing to say. For them, writing means opening a CMS they don't know, facing a blank page, hunting down the numbers and waiting on reviews. So the idea stays in a chat thread, and the website says whatever marketing could piece together instead.
We think that's backwards. The expert should be the author, and everything around the writing should be done for them. That second part is exactly the kind of work AI agents are good at.
Start where the idea happens
At Base22, ideas happen in Microsoft Teams. That's where a client call gets debriefed and where someone says "we should write about this".
So that's where we're putting the start of authoring. We're designing an agent that lives in Teams and is there for anyone on the team with an idea. The conversation is short, and it's about substance, not formatting:
- What's the idea, and who is it for?
- What are the two or three points a reader should take away?
- What's the proof? Which project, which result, which year?
That's a five-minute conversation. The author's part ends there until the draft comes back.
What happens in the background
While the author gets back to work, the agent starts the content workflow. Every step has an owner, and a person owns every decision:
Step | What the agent does | Who decides |
|---|---|---|
Draft | Opens a draft in the CMS with the right structure, topics and related services | Nothing is published |
Evidence | Pulls dates, figures and sources from our records, read-only | The record, not the agent |
First version | Writes it in the author's voice and flags every claim it couldn't source | The author |
Package | Prepares the Spanish draft, image descriptions, search description, and runs the accessibility and SEO checks | An editor |
Return | Sends the author a link in Teams, with the questions only they can answer | The author |
Publish | Nothing. It never publishes | A person, always |
Why the evidence comes first
Most AI writing tools start with the sentence and hope the facts hold up. We've seen what that produces. When we fact-checked our own website before relaunching it, a line claiming a client's help-desk calls had dropped by half turned out to be an anticipated benefit from 2015 that had been copied until it sounded true.
This agent starts from the other end. It looks up the proof before it writes the sentence, and anything it can't source comes back marked for the author, not smoothed over. The author's knowledge and the company's records meet in the draft, and the reader gets both.
What it takes from a platform
This only works on a content platform built for it. We rebuilt base22.com with that in mind, so most of the pieces are already running:
Capability | On base22.com today |
|---|---|
Content stored as structured records, not page layouts | Running |
An agent account with an editor's permissions and no admin rights | Running |
Every change versioned and reversible | Running |
Automated accessibility and SEO checks on every change | Running |
Read-only access to our systems of record | Running, used in our fact-check |
Sign-in with Microsoft 365, so every draft belongs to a real person | In progress |
The Teams agent | What we're building next |
None of this is specific to our CMS. The same foundations can be laid on HCL DX, Liferay, Contentful or Payload, and they are what separate a platform that agents can help with from one they can only make a mess of.
Authors at the front, agents in the back
We don't want AI to be the author. We want it to be the best assistant our experts have ever had: one that never asks them to learn a CMS, never hands them a blank page, and never lets an unsourced claim slip through. The voice, the judgment and the byline stay human.
This article started the way we want all of them to. I described the idea to an agent in a chat, and it came back with a sourced draft for me to shape. The chat just wasn't in Teams yet.
If your experts have more to say than your website shows, Schedule a call.
Topics
- AI & data
- Content management
About the author
Rafael Trujillo
Executive Officer