8 minutes
h1
style
article-header-section

Ross Barton, a two-time Adobe Workfront Champion who manages client projects and administers Workfront for a creative and marketing organization, shares how he connected Workfront to a large language model (LLM) through the Workfront MCP server. He covers what he built, what he learned, and where he would start if you want to try it yourself.

How my MCP journey started

As a project manager and Workfront administrator with nearly two decades of experience in creative and marketing operations, I sit on both sides of the platform. I manage client work in Workfront every day, and I also configure, govern, and automate it with Workfront Fusion. That dual view is what drew me to the Workfront MCP server this year, and it has changed how I spend my week.

At the start of this year, I had barely used AI. I began with basic questions, and curiosity turned into experiments quickly.

My first real project was what I called a project management command center, built in Claude Code. I gave it the information I already work with every day and asked it to help me organize and act on it. It produced stakeholder registries, status documentation, playbooks, and a weekly team update showing which deliverables were coming, who owned them, and what to watch.

What made that work was context. An LLM is only as useful as what it can see, so I connected it to the places my work lives: email, calendar, documents, team chat, and time-off data. Workfront was the missing piece. At first I connected it through the Workfront API as a read-only source, which worked, but involved technical setup that made scaling it unrealistic.

The Workfront MCP server changed that. MCP, or Model Context Protocol, is an open standard that lets AI platforms connect to other systems. With Adobe's MCP server, anyone with the right access can connect Workfront to an LLM without configuring an API. That is when I got excited, because this was no longer a one-administrator experiment. An LLM can hold Workfront data alongside everything else and connect dots that no single system sees on its own.

My organization has used Workfront for five years, with Workfront Fusion handling much of our integration and automation. The MCP server is the newest layer on that foundation. Our approach to AI is exploratory: people opt in when they are curious, and some of the most rewarding moments have come from skeptics who try it once and start building tools of their own.

What is a skill?

You will see the word "skill" throughout this article. A skill is a saved set of instructions that tells the LLM how to handle a recurring task: which connections to use, what to look for, and what the finished output should look like. You do not need to write code to make one. I usually build a skill by doing the task once in a conversation, refining the result until I like it, and then asking the LLM to turn what we did into a reusable skill. From there, I can run it on demand or on a schedule. Adobe also publishes ready-made Workfront skills you can install directly (see Resources).

Reading Workfront data

I started with read-only access and a simple question: what could we do if AI could just read project data?

Weekly team updates

Every Friday, a scheduled skill reads my projects in Workfront, including next week's tasks and assignees. It adds context from my calendar and recent project conversations, and it checks that key meetings are on the calendar where the Workfront timeline says they should be. I review and refine the draft, and the update posts to each project team on Monday morning in our team's chat.

Daily digest

A daily briefing covers what is happening today, project burn rates and upcoming deadlines, and who is out. It also flags conflicts I might miss. If someone just requested time off but is expected at a client presentation that week, the digest tells me so I can check whether they are needed.

Status documentation

Status reports have always taken real effort because the information lives in several places. Skills now draft or update them from Workfront plus that surrounding context. The draft is not perfect, but having the context gathered in one place saves real time. It also gets better: when I edit a draft, I build those changes back into the skill, so the next run starts closer to what I need.

Writing data back to Workfront

Once I was comfortable reading data, I started asking what could happen if AI wrote data back to Workfront, with a person still in the loop. I treat this stage as experimental.

Meeting recaps

A skill reads a meeting transcript and drafts a recap in a standard structure. I review it, correct anything it got wrong, and add links before it goes into Workfront. The biggest benefit is attention: I can be present in the meeting instead of typing. Longer term, I see an agent drafting each recap and placing it on the right Workfront task for review.

Timesheet drafts

The idea is to let AI draft a timesheet from the evidence that already exists, like calendar events, messages, and project activity, so people start from a draft instead of from memory. Every hour it enters includes a comment explaining why, which creates an audit trail. It then summarizes how confident it was and what to double-check. It never submits. The person always reviews and submits their own timesheet.

Reporting and insights

Two kinds of requests have changed the most for me. The first is resource management questions, such as how someone's time has been distributed across projects over several months. Instead of building a new report for a one-off question, I describe the request, answer a few clarifying questions, and get the analysis with a summary on top.

The second is questions that span systems. With multiple MCP connections, an LLM can combine Workfront data with other platforms, such as a customer relationship management (CRM) system, to ask whether the effort we invest matches the value of the work. Insights like that used to require an integration project. Now they start with a question.

As an administrator, I also use the LLM to design and optimize Workfront Fusion scenarios based on how Workfront is configured today, and a weekly skill summarizes Adobe's release notes for our admins: suggesting what to test, flagging what is in preview, and noting what does not apply to us.

What changed for my week

The clearest example is the weekly team update. Pulling it together across every active project used to take me about 90 minutes. Now I review and edit a draft in about 10 minutes. Most other gains are harder to measure but just as real: better meeting notes, faster answers to reporting requests, and fewer details slipping through the cracks.

Asking "What am I not thinking about?"

When I think I am done building something, I ask the LLM one more question: "What am I not thinking about?" The answers regularly surface something that seems obvious in hindsight. Not every suggestion fits, and some look good on paper but would never get adopted. That is where human judgment comes in: keep what works, set aside what does not, and keep iterating.

Moving slowly on purpose

We are taking a deliberate, experimental approach. AI capabilities are advancing quickly, and the tools are changing just as fast, including within the Adobe ecosystem. Rather than rolling out a finished system, I run small experiments, share what works with colleagues who want to try it, and learn from their feedback. That pace also leaves room for governance. Starting read-only and keeping a person in the loop for every write has made conversations about security and confidential information much easier.

What I look forward to most is AI inside the tools people already use. Many people who are new to AI freeze when they open a blank chat window. When they can ask a question or invoke an agent inside Workfront itself, adoption will get much easier. Capabilities like CX Enterprise Coworker and AI Collaborators point in that direction.

Where to start

Ask yourself: if I could read information out of Workfront, where would I find value? Think about the work you repeat every week.

Frequently asked questions

Is it the Workfront MCP server or a connector?

Both terms describe the same connection from different sides. Adobe provides the Workfront MCP server. Your AI platform connects to it, and many platforms, including Claude, call that connection a connector.

Do I need the MCP server to connect Workfront to an LLM?

Not necessarily. You can connect through the Workfront API, but the MCP server makes it far easier, especially for people who are not administrators.

Is it safe to let an LLM write to Workfront?

Start read-only and add write access only when you are comfortable. Keep a person in the loop, and align everything with your organization's AI and data policies.

How do I encourage adoption?

A point I heard at Adobe Summit stuck with me: AI adoption is rarely top-down. It spreads peer to peer. Keep it voluntary, share working examples, and learn from each other. A concrete result, like a drafted timesheet, works better than explaining the technology.

Resources

style
article-content-section