• ~70% adoption of all orgs with Agentforce activated, since the Spring ’26 rollout.
• >50% weekly active of adopted accounts, a sticky core weekly routine.
• Unlocked exec funding traction that resourced the broader Marketing Goals Agent · patent pending.

Modern marketers are drowning in manual configuration. When teams are stuck fighting legacy tools:

• The customer experience degrades into fragmented, irrelevant noise.
• Teams become burnout-rich and insight-poor.

The mission wasn’t just to build a faster text generator. We needed to simplify the end-to-end campaign lifecycle, empowering marketers to build at the pace of an idea.

To pull this off, we needed an engine built on two non-negotiable design pillars. Every downstream decision was anchored here:

• Robust Grounding: Anchoring AI in brand guidelines, historical data, and audience insights so the agent generates safe, relevant content.
• Human-in-the-Loop (HITL) Control: Giving the marketer a persistent, editable brief. The AI holds the pen, but the human steers the ship.

Building ahead of the curve, we evolved the architecture through four phases, each balancing platform constraints, user needs, and fast-expanding AI capabilities.

The evolution above is the what. This is the how, the hard problems I had to solve to make it real, all rooted in the second pillar: the AI holds the pen, the human steers the ship. That meant giving marketers:

• A surface they can steer, from a cramped 300px panel to a full canvas.
• Output they can trust, provenance, guardrails, and graceful failure.
• Control that doesn’t cost a prompt every time, intent-based direct manipulation.

300px Just 300px wide: barely room to read one message’s copy, let alone see the full set we were drafting or steer what the AI generated.

Phase 2 forced the sharpest decision of the project: marketers were orchestrating campaigns they couldn’t actually see. I scoped two ways out:

I chose the modal even though, at the time, Salesforce modals couldn’t yet host the conversational panel. The clarity win outweighed the constraint, and Phase 4 later closed that gap.

Then I tested it.

Before we wrote production code, I put the wide-canvas modal in marketers’ hands. About 90% said they’d adopt it on paper, but the score wasn’t the real signal; the behavior was. Watching real use surfaced two fixes I made before we shipped.

Before After

1. 1. Prepare the brief

   

2. →

3. 2. Preview & refine

   

4. →

5. 3. Review

   

1. 1. Prepare the brief

   

2. →

3. 2. Preview & refine

   

Watching real use revealed a confirmation step marketers didn’t need, so I collapsed a three-step wizard into a leaner two-step workspace.

Before After

The polished preview looked editable, but marketers couldn’t touch the details they wanted to change. Stripping each message back to raw copy gave them a clear space to edit the words.

Because that same testing surfaced something deeper: two problems no amount of canvas space could fix. Marketers couldn’t always tell which words were theirs and which were the agent’s, and they were worn down re-typing prompts for small tweaks. Those two signals, trust and refinement fatigue, defined the next stretch of work, and ultimately converged in the native Phase 4 agent.

The first Phase 3 signal, “who wrote what?”, was a trust problem. An agent that drafts content is only useful if the marketer trusts it. Three failure modes threatened that trust:

The second signal, refinement fatigue, pushed the design past the chatbox. The agent now bridges conversational speed and visual control: drop a prompt, and it drafts the brief and maps a full multi-channel campaign onto the canvas, with the assistant docked for big-picture orchestration.

But to combat refinement fatigue during micro-edits, we moved beyond the chatbox entirely. We embedded quick actions directly onto the generated campaign nodes as intent-based direct-manipulation controls. Instead of typing prompts for minor tweaks, marketers express intent through contextual actions, right where the content lives:

• Nudge the tone.
• Shorten a subject line.
• Regenerate a single step, one click.

We blended the generative speed of AI with the precision of direct-manipulation UI.

AI doesn’t replace the craft; it clears the busywork around it. The work was less about the model than the scaffolding: an obvious next step at every stage, and trustworthy ways to review, edit, and approve before anything ships.

The center of gravity is shifting from screens to systems. This year I started building with Claude and Cursor, so I hand engineering a working prototype to align on intent, not the production build. That’s where I see AI-native design heading.

As the tech moves, expectations move with it, so I designed the system, not the UI. The patterns that held up I carried into the goals agent as defaults: a clear next step at every stage, review-edit-approve guardrails before shipping, visible provenance, in-place refinement over re-prompting, and one canvas for agent, brief, and preview.

A moving product makes a fixed baseline impossible. Benchmarks kept going stale, so we made benchmarking faster and more frequent, an ongoing rhythm, not a one-time study.