Content is the frontier layer that makes or breaks a product’s experience. When it’s great, it’s invisible, and when it’s bad, it’s a wall your customers walk into face-first. Every company scaling software hits the same challenges with content: too many surfaces, not enough governance, and quality that degrades as you grow.
About six months ago, our team built an agent called the Content Assistant. Built on Atlassian’s AI platform Rovo, the Content Assistant drafts short-form UX copy, surfaces our content standards, and meets people where they already work — Slack, Confluence, right in the flow. We did this drawing on our own design skills, content domain knowledge and Atlassian AI.
And it worked. Our little agent became one of the most highly adopted at Atlassian. Hundreds of Atlassians now use it every month — not just Designers, but Engineers, PMs, Marketing teams, the list goes on! Teams told us that time on drafting content dropped from a couple of hours to 15 minutes. We saw similar gains when we scaled out Content Assistant to other content types like release notes and support documentation.
But using AI to cut production time and cost is table-stakes. Here’s the thing that surprised us most:
The Content Assistant became a trojan horse for quality and governance.
Every feature team suddenly had access to our standards, our voice, our craft. Not through a style guide locked away in a file they’d never open, but through an agent that was right there when they needed it. Teams that didn’t have a dedicated content designer could now generate standards-compliant copy.
That’s what the Content Assistant gave us: not just speed, but reach. The ability to scale consistent, user-friendly language across surfaces we couldn’t cover without AI. Our agent is bringing one Atlassian voice across every app, and every feature release, without slowing teams down.
Popping the hood
When someone uses the Content Assistant, the agent provides a rationale explaining which sources it referenced and which guidelines it used to generate the draft. We found instructing it to produce a rationale first can improve the quality of its responses.
For example, this prompt asks for a draft notification that lets customers know when their Studio app specification has been drafted:
In response to this, the Content Assistant:
- Triages the intent of the person’s request and determines what standards it needs to satisfy the request. For example, Monica’s request requires structural and writing guidelines for success messages as well as the Rovo glossary.
- Reads the relevant guidelines, applies them to the request, and drafts the requested copy.
- Prepares a response containing the draft and rationale.
- Runs a quality check over the output.
- Shares the response.
From agent-friendly to AI-native
We built the Content Assistant on top of Rovo, powered by built-in integrations with Jira, Confluence, Slack, and every other app in Atlassian’s system of work, so that it felt more like a real teammate, and less like just another chatbot. As we scaled the agent to different workflows and use cases, what we learned pretty quickly is that a great agent is only as good as the system it’s a part of.
If your content standards live in an old document that nobody’s updated since 2023, your teams will spend more time editing than they saved. If the source of truth is buried in a Slack thread, AI will hallucinate. If there’s no feedback loop, quality slowly fades, and you might not even realize it’s happening until the impact inevitably hits your customers.
If we wanted to drive meaningful change and impact, we had to stop thinking about the Content Assistant as a standalone tool and work out how to build something bigger: an AI-native content system.
4 layers any team can build
So what does an AI-native content system look like in practice?
- Context stores. Structure your content standards, guidelines, and product knowledge so that an LLM can read and apply them appropriately. This doesn’t mean creating another style guide PDF. It means developing a well-governed, machine-readable knowledge layer in an AI-native format.
- Agents and workflows. Embed purpose-built agents in the places where people already work. For our team, that meant embedding the Content Assistant in Slack, a Release Notes agent in Jira Service Management, and a Brief Builder that helps PMs request support documentation in a way that the content practically writes itself.
- An operating model with trust tiers. Not everything needs a senior writer. Not every content type should be written with AI. A release note about a minor change and a pricing flow carry fundamentally different risk. We built an operating model to route work into low, medium, and high-touch lanes based on how much human judgment is required.
- Governance and feedback loops. Efficiency and quality metrics are documented and tracked in a dashboard. Automated feedback loops keep content fresh and relevant. Every AI-generated draft that gets edited by a human is a signal. We wire those back into the system so that it gets smarter, not just faster. Every agentic workflow has human governance baked in by design.
Landed impact
The results speak for themselves. Our Content Review Desk, the service layer we’ve wrapped around the system, has processed 1,500+ requests in less than 9 months. Release notes production time has dropped by 88%. We’ve given back 700+ hours of time for our content designers to now spend on strategy and craft instead of drafting multiple rounds of placeholder content in Figma.
And because this is a system, not a tool, the compounding effect is real. Every standard we codify makes the agent smarter. Every workflow we automate frees up a human for more strategic work. Every feedback loop makes the next draft better than the last.
Where do we start?
So, if you’re sitting here thinking, “We need something like this,” here’s how you can get started today:
Start with one repeatable content type. For us, it was UI copy and release notes. Pick the thing that eats up time, follows a repeatable pattern, and has a clear quality bar. Start by proving your model there.
Don’t neglect your context layer. This is one of the key levers of response quality – maybe the most important. This means getting your content standards house in order, and lifting important knowledge out of local files into AI-friendly documents that are structured for LLMs. While this has been a heavy lift, it has helped the Content Assistant become more reliable and trusted by our teams.
Design your operating model around risk and trust, not just automation. Build lanes. Let humans focus where judgment actually matters.
And build the feedback loop from day one. Not as a nice-to-have, but as built-in by default.
The shift to AI-native systems
We are now officially at the beginning of our AI-native content system journey that gives us the opportunity to codify our craft expertise and surface it wherever our customers need it most.
We’re still building, and we’re building openly, across all our teams at Atlassian. As we learn, we’ll keep sharing how we do it — from leveraging Atlassian apps like Jira and Confluence to power the system, to leaning into emergent AI patterns to push the envelope.
While there’s still a ton of unknowns to work through, one thing is for sure: By shifting our thinking towards working in systems, we’ve unlocked our ways of working and raised the quality bar in ways that were unimaginable even a year ago.
