For years, Parry Malm sent millions of emails. He’d write subject lines, test them, try to spot patterns in what worked. Some performed well. Others flopped. He attempted to build a simple heuristic system. Some internal logic to understand why certain language worked and other language didn’t.

There was no way to do it.

Then he joined Adestra, where he worked with thousands of marketers worldwide. And every single one of them asked the same question: “What should I put in my subject line?”

Every single time, his honest answer was: “I don’t really know.”

Phrasee was founded on a simple realisation: the challenge of identifying high-performing subject lines wasn’t unique. Every marketer struggled with it, yet no systematic answer existed. This was 2015, before the advent of generative AI and the rise of ‘AI-powered’ as a marketing buzzword.

If you have ever received emails from Sainsbury’s, Tesco, or Currys, you have likely seen Parry’s work. He built Phrasee, grew the company to about £10-11 million in revenue with 100 employees, and sold it in 2022 for roughly $100 million. After that, he retired, ate kimchi, and went on walks.

In September of this year, he launched Drumbeat:an AI system designed to handle the mundane, yet effective B2B marketing tactics that actually work, but nobody wants to do themselves.

During our recent Browse Basket Buy conversation, Parry shared how he solved these problems, explained why language is a data issue, and exposed key mistakes companies make with AI.

Language is data (we just don’t think of it that way)

People assume language isn’t a database problem. It is. It’s just that there are enormous amounts of data involved, and because it’s human-created, the nuance is significant.

Parry understood this: certain words prompt specific responses in our brains. With enough examples, you can model and predict how humans will react to language at scale. Enabling you to forecast responses before sending any message.

This insight wasn’t obviously correct in 2015. When Phrasee launched, the objections were predictable: “There’s not a business in that. Nobody cares about subject lines.” And then: “It’s never going to work. Machines can’t write better than humans.”

Parry thought they were wrong. He was right.

The approach was methodical: build generative language models on a brand-by-brand basis, output human-sounding subject lines, run extensive tests, and use that test data to create optimised models that could predict language performance before deployment.

Marketers lacked a systematic, reliable method for predicting which language would perform. Phrasee worked because it provided those systematic answers, eliminating the need for expensive trial and error.

The Einstein principle: understand the problem deeply

Parry often paraphrases Einstein: if you truly understand a problem, implementing a solution is just a technical process.

This philosophy explains why Phrasee succeeded where others failed. The company didn’t start with “let’s use AI.” It began with a clear problem: there were no effective heuristics or mathematical models for determining which subject lines would be successful, despite the universal belief that subject lines were critically important.

Understanding this problem deeply meant recognising several truths:

Subject lines represented the highest-leverage decision in email marketing. Everything else – the design, the offer, the copy – becomes worthless if people don’t open the email.

No systematic solution existed. There were best practices, certainly. But these were generalisations, not brand-specific guidance. What worked for one company might fail for another.

Language patterns could be modelled. This was the crucial insight. Given sufficient data on how specific audiences respond to specific language choices, you could build predictive models that are effective.

Start with the problem, not the solution

The pattern repeats constantly in AI implementation: companies decide they need to “use AI” and then search for applications. They start with the technology and retrofit problems to justify its use.

Parry’s observation is direct: “Don’t worry about AI. Worry about the problem you’re trying to solve.”

This inversion matters more than it appears. When you start with AI as the solution, you’re asking: “Where can we apply this technology?” When you start with problems, you’re asking: “What’s the best way to solve this?”

The technology should always serve the problem. Not the other way around. Sometimes AI offers a better solution; often it doesn’t.

This discipline explains why Phrasee succeeded: rather than beginning with “How can we use machine learning in marketing?”, it started with the problem—”Marketers can’t systematically optimise subject lines; how do we fix that?” Machine learning ultimately proved to be the most effective solution to this challenge.

Understanding stochastic parrots

Today, when most mention AI, they’re referring to large language models (LLMs), though many other forms of AI also exist.

Parry’s explanation of how LLMs work is worth understanding precisely: companies scraped the entire internet and built models that predict what word is most likely to follow another word. They’re sometimes called “stochastic parrots.”

The term captures two characteristics. “Parrot” because they speak back to you in something that sounds human, even though it’s not human. “Stochastic” because the output is random within certain boundaries.

Take the example of flipping a coin. The outcome is randomly either heads or tails, but always within those possibilities. This is a random process with clear limits.

LLMs work similarly. Ask one to complete “my cat’s breath smells like…” and it might say cat food, canaries, or numerous other options. But the answer will be within a bound of reasonableness most of the time.

Most of the time, the stochastic system follows a normal curve. Usually, answers fall within expected ranges, but occasionally, the results are entirely unpredictable.

This is why these systems usually provide correct or acceptable responses—but sometimes, unpredictably, they do not.

Parry’s comparison is pointed: these systems resemble those of McKinsey consultants. Always confident. Sometimes right. A dangerous combination.

From Phrasee to Drumbeat: solving different problems

After selling Phrasee in 2022, Parry retired for two years. Parry spent time walking and eating kimchi before launching Drumbeat.

The pattern holds: both companies started with specific problems, not technological capabilities.

Phrasee addressed a clear gap: marketers couldn’t systematically optimise email subject lines. To solve this, generative language models were built for each brand, tested, and optimised to predict performance before deployment.

Drumbeat tackles a different problem: the boring, unsexy B2B marketing tactics that actually work but nobody wants to do. The tedious tasks that marketers spend time wishing about, complaining about, and dreaming about. Instead of actually executing.

Historically, businesses faced three options: do the work themselves and stay up until 2 a.m. on Sunday, only to see their efforts fizzle when something else demands attention. Hire someone who takes sick days, enjoys the creative tasks, but consistently resists sending newsletters. Or engage an agency, where you’re pitched by the founder but the work is done by a new graduate, resulting in mediocre output for a high price.

The fourth option, enabled by AI advancements: an end-to-end system that plans, creates, and orchestrates marketing where it matters.

Both Phrasee and Drumbeat began by identifying a genuine business problem, rather than focusing on the technology. Technology only entered the picture when it offered a better solution than existing alternatives.

What to do right now

Parry’s practical advice is deceptively simple:

When solving a problem, focus on understanding the problem itself. Gather all relevant information and examine it carefully. Once you’ve done this, take a walk in the forest. Often, clarity and the answer will emerge during reflection.

This isn’t mystical. It’s practical neuroscience. Human brains process information differently when we’re not forced to provide immediate answers. The walk provides space for pattern recognition to operate without conscious pressure.

Whether you’re considering AI implementation, refining marketing strategy, or tackling complex business challenges, use a consistent method: first, thoroughly understand the problem. Next, allow yourself time to reflect, rather than pushing for immediate answers. Once you truly grasp the issue, consider possible solutions.

Starting with a specific, measurable, and clearly understood problem occasionally leads to valuable solutions. By contrast, starting with a solution results in expensive, misaligned experiments.

The gap between these approaches determines whether AI serves as marketing theatre or as a tool that works.

This article was written with the assistance of AI.