Austin Lau joined Anthropic — the makers of AI model Claude — in 2024 as the company’s first growth marketing hire in the U.S. Since then, he’s built out performance marketing, SEO, lifecycle, attribution and a suite of AI-powered workflows largely on his own, contributing to the company’s meteoric rise from $150 million to $7 billion in revenue in two years.

Remarkably, the technology he markets to the public is rapidly changing the marketing industry itself. According to HubSpot’s 2026 State of Marketing Report, 80% of marketers already use AI for content creation, while 86% of marketing teams use AI in at least some areas of marketing.

So one might expect that Lau’s privileged vantage point would unlock a treasure trove of AI secrets and shortcuts. But during a fireside chat at the AirOps Next Conference in New York City on May 13, he pushed back on that assumption.

“We don’t have a magic playbook to reference,” he explained. “We’re honestly trying to figure out and build these functions and muscles ourselves.” The models are going to get “exponentially better,” he predicted, and he finds it challenging — even internally — to keep up with the new features, model launches and optimal use cases for Claude within the marketing team and other non-technical functions across Anthropic.

Still, Lau’s front row seat to Claude’s evolution as a marketing tool provides enviable perspective. Following are his insights and best practices for marketing teams experimenting with the technology.

Determine what Claude does better and faster than your team — then prioritize the overlap.

These two questions should drive every decision about what processes to automate, according to Lau. “There are basically two dimensions … when I’m thinking about building or spinning up a new agent of some sort. What can Claude do better than me, or what can Claude do faster than me? Ideally we’re working on something [that] crosses both of those off.”

The first step is letting Claude — or your preferred LLM — know what your process looks like today. “You want to be able to encode your entire process end to end,” he said. “Then you can start to understand, okay, what are the areas that we can potentially automate? There’s going to be some areas where … it’s just impossible because this tool doesn’t expose an API so we can’t programmatically pull the data.”

Use your own outputs as a benchmark.

Provide examples of what you consider to be an optimal result and compare Claude’s output against what you’d produce manually. This is a fast, scrappy way to evaluate quality before scaling anything, Lau said, and providing examples of what good looks like allows you to test the process. “Here’s Claude’s output versus my output if I were doing it manually. Does this meet the bar or not?” Lau said.

Don’t try to build everything in one prompt.

A common mistake, Lau said, is dumping a complex brief into a single prompt. Instead, break the command into pieces, remembering to validate the concept first and then build incrementally.

“When I first started using tools like Claude Code, I would literally try to one shot everything,” he admitted, which prevented him from evaluating the LLM’s steps before it performs the work. “Break it down into much more digestible pieces,” Lau said, “one as a proof of concept to see some of the initial bones [and decide if] this actually makes sense, this is actually feasible. And then from there build up in increments.”

Don’t necessarily rebuild what you can buy.

Just because AI can generate code doesn’t mean your team should spend time replicating tools that already exist, Lau said. He recommends evaluating the ROI of building versus buying — and always factoring in the maintenance of that tool.

“Once you build it, you need to think about all the tech debt,” he said. “Somebody has to maintain it. Who’s going to continue building on top of it? Unless you have a ton of dedicated resourcing or a bunch of people on your team that have too much time, which I highly doubt is the case for anybody, then you should really always evaluate what those tradeoffs are.”

Build custom tools only when your use case is truly niche.

When deciding whether to build something in-house or pay for a tool, choose the latter — unless the problem to solve is niche enough and only needed by a handful of people. Anything widely used should be bought, and that philosophy is followed by Anthropic itself. “We’re not going to build our own CRM,” Lau said. “We use Salesforce just like everyone else. We also use Gong just like everyone else.”

The exception for Lau is when the use case is precise, targeted and rare. “There are going to be some subset of tools where it is very bespoke and you’re having a hard time even finding a tool that solves this very specific and niche use case,” he said. “If nobody else has the same problem, nobody’s 1) going to build it and 2) that means that I’m sort of out of luck. So in that case, I’ll build something myself.”

Encode your team’s tribal knowledge into AI skills as soon as possible.

The way teams currently transfer knowledge — training new hires and walking them through processes — needs to become codified, shareable and repeatable, according to Lau.

“I’m not seeing enough non-technical teams actually encode their knowledge,” he said, “so that it shifts from being this sort of tribal knowledge known by few to now being something where … we can actually share this skill so somebody else outside of our team can also access the same previous tribal knowledge in order to help them with their work as well.”

This is particularly important for subject matter experts — skilled in reporting or putting together a campaign brief, for instance — whose knowledge should be passed on more easily to other teams.

Don’t re-prompt from scratch every time.

Without encoded workflows, every new conversation starts at zero. Building reusable skills is what makes AI adoption scalable rather than a series of one-off requests. It’s about “encoding all of this stuff so that you can actually truly leverage AI in a way that’s going to be repeatable and scalable as opposed to trying to re-prompt every single time [you’re] starting a new conversation. That’s a pretty painful process,” Lau said.

Prioritize craft in new hires, then AI curiosity — but require both.

Functional expertise is still a requirement for new marketing hires, Lau said, but candidates should also be actively experimenting with AI, not just using it for basic copy edits or generating punchy headlines.

“We’re not abandoning first principles here. We still need somebody who lives, breathes, eats, sleeps, whatever function we’re hiring for. They have to have experience in that,” Lau said. “But some of the other things that we’re also looking for now are, is this individual AI-pilled or even AI-curious? Especially for non-technical teams where adoption of AI has for the most part generally started and stopped with just chat.”

“That’s where I start to see a lot of the conversations shift,” Lau said. “We’re looking for people who are really good at what they do, but can also understand where AI actually fits into the workflow — whether that’s to supplement areas that maybe they’re not as good in or make areas that they are really solid in excel even further.”