First-party data is the new ad infrastructure. The cookie was never the real point.

Marketing Cloud is the leading marketing cloud and focuses on sending one-to-one communications with known customers. There has always been an ad side; we had a DMP, Krux, that was acquired by Salesforce, and Datorama for media optimisation. We are a major source of demand for Meta, Google, Amazon, and increasingly Snap and LinkedIn. The model: build a first-party audience in Data Cloud (our CDP), send it over to the walled garden, the walled garden does a match, runs the campaign. That's the post-cookie paradigm for ad execution.

I tell younger people my entire career didn't exist when I was in college. The internet didn't exist; it had just started. What led me into marketing is that it combines psychology with technology and data. There aren't many professions where you combine the left brain and the right brain like that, and I appreciate that. I also like working with creatives; advertising has a lot of creative energy. The data and analytics side has always been attractive to me, and it's grown in importance over time.

I've made the move from analyst to operator. Being an analyst is contemplative and cerebral, putting together reports, the cadence I liked. Becoming an operator is being in meetings, influencing people, getting things done. I can apply the analytical side. I respect operators (and salespeople) more than I used to. The longer I've been at Salesforce the more I realise sales is very difficult and the science of it is real.

Marc Benioff said he'll be part of the last generation of CEOs managing an all-human workforce. The hybrid agent-human future is here already. It's adopted more slowly in marketing and advertising because the work is more complex than the call centre. 70% of call-centre calls are about the same things (returns, status, delivery). In marketing, campaigns have many steps.

We built Agentforce as a platform that lets enterprises assemble trusted agents, swarm them around a problem, test them, and deploy them safely. Put them in front of customers in a way that doesn't ruin the brand. A platform for building trusted agents.

The platform is a year old. Many customers are testing into it. In marketing we have something called campaign creation: an agent that takes a general description of a campaign (your product, your audience), develops a brief, you approve the brief, a different agent pulls a segment or audience from the brief, another agent produces the media plan, another runs optimisation. Efficiency improves through the chain.

There are different types of agents. Employee-facing agents help the workforce but aren't exposed to the customer. Customer-facing agents are higher risk. Then there's agent-to-agent (agent-facing) interaction. Anything in front of the customer is higher risk. A good place to start is employee-facing.

A CMO in wealth advisory I recently spoke to has brokers on the phone with high-net-worth customers about their portfolios. An agent runs on the broker's desktop, listening to the conversation, offering tips on what to say next, surfacing relevant product upsell ideas, pulling everything the firm knows about the customer. The customer doesn't see the agent; the broker does. It enhances the human-to-human interaction.

I was at Gartner covering marketing clouds and the hype-cycle framework describes the trajectory of any new tech category. CDPs went through it around 2018-2019. Tech categories are first overhyped (peak of inflated expectations), then underhyped (trough of disillusionment) because people try it and find it isn't as easy as expected.

Companies are stumbling now because expectations in the short run are too high. It's very easy to build a single agent. It's very difficult to make it work on an ongoing basis in a workflow, to make agents work together, and to prove them. That takes time and patience. Companies become disillusioned too quickly, then overreact and say I don't need agents in my workflow. That's a mistake. The right posture is: expect less in the short run, more in the long run. Twenty years from now the workplace will be very different.

Email and SMS are a large part of our business, much of it non-transactional performance marketing where the messaging is repeated over time. Models trained on the brand's past messaging can generate variants of headlines and body copy. Product descriptions in Commerce Cloud (100,000 SKUs) are well-suited to generative AI; humans don't want to write 100,000 product descriptions; a computer has no objection. Versioning is the helpful application.

The goal is personalisation at scale. Each interaction (website, app, call centre with a human) feels coherent, held together, the way a CDP like Data Cloud would do. The customer doesn't feel they're being targeted or that their data is being weaponised; they feel that everything is right on point.

The strategic emphasis on Agentforce and AI agents is built on Data Cloud. The data foundation comes first, then agents on top.

People forget history, and young people don't know it because they weren't there. Technology disruptions have things in common. When I wrote Agentforce, I had to think about the future of the workforce. I was sceptical: would there be any role for people, what about us hardworking humans?

Looking at history (the internet, the telegraph, the agricultural revolution, the industrial revolution, computers in the 1960s), each was greeted by widespread panic that human workers would be displaced and life as we know it would be over. Overreaction every time. In the long run, more jobs are added. The jobs we have today didn't exist 20 years ago. The jobs the next generation has won't exist now. We don't know what they will be.

In a great many customer organisations, the sluggishness in AI adoption is coming from the top, not from the workforce. Most of us use AI in our daily lives and want to use it at work. People at the top are reluctant for various reasons (legal risk in many cases). CEOs and CMOs tend to be conservative. Being too conservative now is a mistake. Risk can be controlled through testing.

For junior people: get technical. Don't be afraid to take a course in Python, SQL, data operations. You don't have to become a master. Technical literacy helps significantly in the long run.

Marketing is fundamentally human. Marketing is one person presenting a case for a product to another person, communicating human to human. We need to imbue every process and every agent with that human quality. Trust common sense. Common sense is the stuff humans know that AI models don't, and there's a lot.

An AI model is really just somebody who walked into a library and read every book in the library. Every book. But they don't have a life. They don't live in the world. They don't interact with people. They have to fake all of that. Humans have it. Keep the human in. Don't underestimate the power of the human.

Scott Brinker, who covers the MarTech ecosystem, has said there are 4,000 to 5,000 AI start-ups from the past year. They're all pitching the CMO. Nothing wrong with trying a product. The goal is not a single agent that can do a single flashy task or demo well. The goal is to improve the organisation as a whole. That means a repeatable process, a platform, the ability to build multiple agents, a repeatable way to test them, people in the organisation who know what they're doing. More than only a demo.

I rely on human common sense. Psychotherapists say a good cue to how someone is perceived in the world is how you are perceiving them. Different companies have different cultures. Salesforce is hard-driving but more relaxed than (say) Oracle. Microsoft is complex and changing. You have to see the person in the context of the culture.

In the area I work in, technical interest matters. Having at least an interest in the technology, even if you're not a computer-science major, helps.

Hollywood is different people with different skills coming together. They may not know each other, but they all have their own area, and they can work together to make a great movie. Then they dissolve and the team moves on. Software engineers do the same. It's a forming team. Collaboration platforms (Slack, Teams) enable it, when they're genuinely used. Asynchronous communication, plus dialogue on video and not only text.

AI is a tool like anything else. The danger in relying on something like Claude or another sidekick too much is that you'll lose skills over time. If you have it do all your writing, your writing will get worse. That may be okay for people who don't enjoy writing; their emails will sound better than they otherwise would, which is a good thing. The decision is conscious. I'm a writer; I wouldn't have AI write for me. I can do a better job saying what I want to say.

The honest cost of overuse: people can tell. Customers can tell. Coworkers can tell. This guy put no effort into this, put it into a machine, sent it to me. That's almost insulting. The recipient feels they weren't worth putting thought into. Over time the sender erodes their own brand capital.

The flip side: better prompts produce better outputs. Creating a good prompt is art and science. The discipline of writing the prompt may help people become clearer communicators about their value proposition and customer.

Understand at a high level how large language models are built. The LLMs revolutionised AI. GPT-3 was a step change, based on insights from Google and other sources. There's a method to how these models are built. It's complicated; I spent a long time figuring it out. You don't need to be that technical to get the idea of how the models are trained. Understanding the principle demystifies it: it's a process, on-and-off switches at massive scale. Andrej Karpathy has videos on this; they take four or five hours and they're worth the investment.