Introducing AI without breaking trust
The complexity doesn't disappear. It just moves somewhere the user never has to see it.
How we brought AI into an enterprise authoring tool without breaking the workflows, or the trust, marketers already relied on.
Context
Message Designer is Oracle Fusion Marketing's email authoring tool, the system marketers already trusted to get every campaign out the door. It is a powerful editor: custom components, multiple layouts, personalization, revision history, variant panels, and rule builders. People depended on it daily.
Then the mandate arrived: add AI. This was roughly two years ago, before "generative" and "agentic" were everyday words, and a moment when almost no one trusted AI. Dropping it into a tool people already relied on was a trust problem before it was ever a design problem. Get it wrong and you don't just ship a weak feature. You make people distrust a tool they were using fine yesterday.
The Tension
Generate the entire message.
The one button leadership wanted. It demos beautifully.
So the question was never "can we add AI." It was how to add it without breaking trust, workflow, or brand control, in an environment where a free-range chatbot is a non-starter.
Earn AI one surface at a time
Rather than chase the one-button dream, we proved the integration on the smallest, most demanding surface first. Tiny surface, high stakes: the subject line is the make-or-break on whether an email gets opened at all. Small enough to build safely, important enough to prove the AI was worth trusting. It let us get a real LLM into the product, generate genuine variations, and learn how the interaction should feel before betting the whole message on it.
Slower, and far less impressive in a demo than "generate everything." What we got for that patience was something real and controllable, instead of magic we couldn't govern.
The carousel
This is where the design decision, the business constraint, and the human constraint all collided, and where one move resolved all three. Redwood said one output: trust Oracle's data, give the single correct answer. Reality disagreed.
Give the user the single correct answer. We were the first team to put AI in the system, so there was only that conviction to work against.
Clients were billed in geographic pools. Users regenerate ten or more times hunting for the right line, so runaway regeneration was a budget problem.
"What was the one three generations ago?" Each click piled on cognitive load, with the best option already scrolled out of memory.
Genuine variety to react to, without breaking Redwood's bar for trustworthy output.
Cut the paid-generation churn. Five held options instead of endless regenerates.
Nothing scrolls out of memory. Step back to the option you liked.
We didn't fight the one-output rule. We gave the Redwood team the scenarios and context they hadn't had in front of them, and seeing the real use case, they opened up to it. The first team to bring AI into the system ended up informing how they approached it going forward.
Governance
The last decision was about control: no free-range chatbot in an enterprise environment. The AI had to be governed by each enterprise's own brand standards, not a generic default, so the marketer never had to hold the guardrails in their own head.
AI generated text, subject lines and text content, and was structurally kept away from everything it couldn't safely author: layouts, other components, variant panels, custom HTML blocks. It worked where it could, and couldn't reach what it shouldn't.
The template creator sets intent and constraints when the template is built; the template user works within them. Two different personas, two different levels of control.
One team locked headlines down with character limits and capped section counts; another gave the template user full control to edit anything. We built for both ends of that spectrum, with each client's brand standards enforced.
More scaffolding up front and less apparent "magic." In return, the output stays on-brand and on-strategy, which, in an enterprise tool, is the whole game.
The Honest Outcome
Message Designer's AI reached client organizations in production and was in active use, gathering live feedback to iterate. The cross-agent direction, connecting authoring to campaigns, was later announced publicly at Oracle AI World 2025.
The subject-line integration was designed with the text component already in mind, reviewed with engineering along the way, so nothing went in that couldn't also work in a text component. The trigger stayed consistent, the same AI button and dropdown, while the options behind it fit each component's job.
The Landing Page designer and Forms designer were both based on Message Designer. Other teams reused my patterns and components while I oversaw their work and guided the extras their applications needed. Because that foundation was consistent, the AI features carried across all three with little friction.
The complexity of safe, on-brand AI didn't disappear. It moved into the scope limits, the layered instructions, and the governance. The marketer just gets a good first draft.