How much can AI actually improve a person's work efficiency?

Recently, a post about Anthropic triggered a lot of shares on social media. The poster, Ole Lehmann, stated that the entire growth marketing team of this company valued at $380 billion has only one person - a non-technical marketer, who single-handedly managed paid search, paid social, app store optimization, email marketing, and SEO for nearly ten months.

Shortly after the post was made, it was questioned in the comments section, but soon, the person involved confirmed it themselves. This growth marketer named Austin Lau replied that when the article was written, he was indeed the only person doing growth marketing, managing it all alone for nearly ten months.

Image丨Related tweet (Source: X)

At the end of January this year, Anthropic released an official case study detailing Austin Lau's way of working. Around the same time, Anthropic also published an internal white paper titled "How the Anthropic Team Uses Claude Code," which covered use cases from ten teams, including data infrastructure and legal departments, with growth marketing being one of them.

The white paper states: The growth marketing team focuses on channels such as paid search, paid social, mobile app stores, email marketing, and SEO, and is a "non-technical one-person team," relying on Claude Code to automate repetitive marketing tasks, building automated workflows that traditionally required a lot of engineering resources.

(Source: Anthropic)

Austin Lau is not an engineer. He mentioned in Anthropic’s official case video that he "has never written a line of code," and when he first encountered Claude Code, he even had to Google "how to open the terminal on Mac." When Claude Code was first released, his initial reaction was "completely unaware of who this product was for;" as a marketer, he found its purpose unclear.

The turning point occurred when a colleague shared a Claude Code installation guide for non-technical employees in the company Slack group. Out of curiosity, Austin installed it, and a week later, he built two automated processes that completely changed the way he worked.

The first process is a Figma plugin. For paid social ads and app store marketing, he needed to handle a large number of visual materials in Figma. The previous process was: when creating multiple copy variants for the same design scheme, he had to manually copy the framework from Figma, constantly switch between Google Docs and Figma, copying and pasting titles one by one. If there were 10 copy variants to fit 5 different aspect ratios, this mechanical labor could easily take half an hour.

Image丨Austin Lau (Source: Anthropic)

He described this pain point in natural language to Claude Code and asked it to help write a Figma plugin. During the process, he had Claude Code refer to the Figma API documentation while developing the prototype. The first version of the generated prototype was not perfect, but it was enough as a starting point. He continuously adjusted it based on that and ultimately created a usable plugin.

(Source: Anthropic)

The plugin works by selecting a static image frame; it automatically identifies components such as titles, call-to-action buttons, and code blocks, and then bulk generates individual Figma frames from a prepared list of copy, with each variant corresponding to a new copy. A single batch generation can create up to 100 ad variants, taking about half a second per batch. The previous 30 minutes of manual operation is now reduced to 30 seconds.

The second process is a copy generation workflow for Google Ads. Google Ads' responsive search ads have strict character limits for titles and descriptions, with a title limit of 30 characters and a description limit of 90 characters. Previously, he had to draft in Google Sheets, manually check the character count, and then paste the content piece by piece into the Google Ads backend.

Austin created a custom slash command "/rsa" in Claude Code, which, when triggered, would prompt for input on campaign data, existing ad copy, and keywords, then cross-reference his preset "Agent Skills," which include Anthropic's brand tone, product accuracy standards, and best practices for Google Ads RSA.

The system uses two well-defined sub-agents, one specialized in writing titles and the other in writing descriptions, each working within their respective character constraints, producing a quality far superior to cramming both tasks into a single prompt.

Ultimately, Claude Code packaged 15 titles and 4 descriptions into a CSV file that can be uploaded directly to Google Ads. Austin emphasized that the generated copy is just a starting point; he evaluates each one: Is the value proposition clear? Is the tone appropriate? Is there a distinction from competitors? But at least the tedious initial draft generation and formatting work has been completely automated.

The efficiency gains from these two workflows are already remarkable, but Austin's system doesn't stop there. He also built a MCP server (Model Context Protocol) connected to the Meta Ads API.

Through this integration, he can directly query ad performance, spending data, and the effectiveness of each ad within the Claude desktop application, without needing to open the Meta Ads dashboard. Questions like "Which ads have the highest conversion rates this week?" and "Where did I waste my budget?" can be asked directly to Claude, receiving real-time data as answers.

More importantly, it closes the loop. Austin has built a memory system that records hypotheses and experimental results from each round of ad iterations. When starting a new round of variant generation, Claude automatically retrieves all previous test data, highlighting which copies performed well and which did not, refining the next round of generation based on historical experiments. This system becomes smarter after each cycle. Systematic experiment tracking across hundreds of ads would typically require a dedicated data analyst in traditional teams.

According to Anthropic's white paper, the results of this working method are: ad copy creation time compressed from 2 hours to 15 minutes, and the creative output increased to ten times the original, with the ad variants he tested covering more channels and quantity than most fully staffed marketing teams.

In that white paper, growth marketing is just one of ten cases. The data infrastructure team uses Claude Code to debug Kubernetes cluster failures, resolving issues that previously required contacting network experts in just a few minutes; members in the inference team without machine learning backgrounds use it to understand model functions and settings, reducing document lookup time from an hour to 10 to 20 minutes; the product design team directly uses Claude Code to modify frontend code, and engineers notice designers making "large state management changes that you wouldn't usually see a designer making"; a member of the legal team created a predictive text assistive application for family members with language barriers in just an hour, despite having no prior programming experience.

The usage methods between technical and non-technical roles may differ, but the conclusion is consistent: Claude Code is blurring the boundary between "can do" and "cannot do," a boundary that was almost entirely determined by technical capabilities in the past.

Austin Lau summarized in the case that, in essence: "The distance between ‘I wish this existed’ and ‘I can create it myself’ is much shorter than most people think."

Of course, it should be noted that growth marketing does not equal the entire go-to-market (GTM) strategy. Anthropic has a complete brand, product marketing, and communications team; Austin Lau is responsible for performance marketing, which includes quantifiable channels such as paid promotion, app store optimization, and SEO.

In February this year, Anthropic ran a television advertisement during the Super Bowl, which is clearly not something one person can manage alone. The copy and brand assets that his workflow depended on were initially produced in collaboration with the product marketing and copywriting teams, with Claude generating variations and scaling tests based on that.

Austin Lau recently added some background on LinkedIn. He pointed out that the widely circulated article describes his experience as the sole growth marketer in the second quarter of 2025, which was nearly 8 months ago. The team did eventually expand, although still on a much smaller scale than people outside may imagine; in his words, "our combat effectiveness far exceeds our numbers."

Even so, the signal is strong enough. A company with a post-investment valuation of $380 billion and an annual revenue of $14 billion, at its fastest growth stage, allowed a marketer who has never written code to manage core growth channels alone for ten months, with positive results. This should already demonstrate that the amplification multiple of AI for knowledge workers may be much larger than what our current organizational structure and hiring inertia assume.

However, it is still unclear how widely this model can be replicated. Growth marketing is highly data-driven, process-oriented, and API-friendly, making it naturally suited for automation. Fields that require more interpersonal judgment or creative intuition may not fare the same.

At the end of the growth marketing chapter, Anthropic's white paper provides three recommendations: look for repetitive workflows with API interfaces to automate; break complex processes into multiple specialized sub-agents rather than trying to tackle everything with a single prompt; and fully think through the overall process design in Claude before writing any code. These three recommendations essentially illustrate that the bottleneck of efficiency often does not lie in technical capabilities, but in whether you are willing to take the time to clarify your workflow and then delegate the parts that can be taken over by machines.