The shift I've been building toward
For most of the last two decades, "automation" meant a human painstakingly encoding every branch, every rule, every exception — and then maintaining that brittle logic forever. Agentic AI breaks that contract. An agent is automation that can reason about an unfamiliar situation and act, which means the human is no longer required to build the automation. The job changes from authoring the workflow to directing and supervising a system that figures out the workflow itself.
That's the lens I brought to Salesforce's first generative AI experience for Marketing Cloud, and to the multi-agent interoperability strategy I bootstrapped after it. It reframes what a product even is: not a fixed set of features, but a capable collaborator the customer learns to delegate to.
The pattern moves faster than the roadmap
The hardest truth about agentic products is that user behavior is changing month to month. The way someone prompted, trusted, and chained agents a quarter ago is not how they do it today. A traditional roadmap — scoped, committed, shipped two releases later — is structurally too slow for a pattern that mutates this fast.
So my product strategy optimizes for one thing above almost all else: removing friction so adoption can outrun the change. Two concrete bets make that real:
A visual experience inside the conversation
Marketers think visually — a text-only chat is a friction wall. I brought a rich visual experience into otherwise purely conversational agents for building campaigns, so the agent meets people in the medium they actually work in (the thinking behind aligning to A2UI declarative rendering).
Activation on Day 1, not month six
I made activating the agent a core part of the Day-1 experience in Marketing Cloud — not a feature buried three menus deep that a customer stumbles onto months later. If adoption has to outrun the change, the agent has to be the first thing you meet, not the last.
The same instinct is why I pushed Model Context Protocol so early — authoring the vision for both an MCP client (Agentforce consuming external tools and context) and Salesforce-hosted MCP servers (exposing Marketing Cloud to external agents like Claude and ChatGPT). Interoperability is friction removal: you meet agents — and the developers building on them — where they already are, instead of forcing them onto your island.
Agents are not the end-all
I'm a pioneer of this technology, and I'll still say it plainly: agents are not the answer to everything. Reasoning is not omniscience. There is a whole class of problems — finding subtle patterns across enormous volumes of data, where the signal is statistical rather than linguistic — that both humans and LLMs are genuinely bad at.
That is exactly where classic machine learning, vectorization, and embeddings continue to earn their place. The sophisticated systems aren't "LLM or nothing" — they're an orchestra: ML surfaces the patterns at scale, vector search grounds the model in what's true, and the agent reasons and acts on top. Knowing which layer should own which job is the difference between a demo and a platform.
Reasoning still needs a keeper
Delegating the building of automation to AI does not mean delegating accountability. Agentic systems demand oversight, maintenance, and governance — they drift, they encounter inputs no one anticipated, and they will confidently do the wrong thing if nothing is watching. Pioneering responsibly means designing the guardrails and the human-in-the-loop checkpoints into the product from day one, not bolting them on after the first incident. That's a whole discipline of its own →