Look Outside the Building: Shaun Clowes on Why Most PMs Fall Short and How AI Changes the Job

Shaun Clowes, chief product officer at Confluent and formerly of MuleSoft, Metromile and Atlassian, calls the state of product management a conundrum and "a bit of my life's work." Fifteen to twenty years into the discipline, he says, outcomes, behaviors and individual performance still look random. Something about how the craft is practiced misses the truly value-adding parts: the right way to think about problems and the abstract reasoning they demand.

Clowes draws on his past as an engineer, where people talk about 10x engineers. If those exist, he argues, then 10x product managers must too, and because a PM's job is leverage over other people's work, the return on one could be 100x or more. With outcomes that widely distributed, he finds it surprising that the field has not become far crisper at reliably producing them.

Host Lenny Rachitsky adds that a weak PM experience feeds a meme that PMs are useless, something nobody says about engineers or designers. Clowes agrees the role starts at a disadvantage, because a PM must say no to around 90% of what comes their way and so has to rack up wins quickly with good insight and data. He still thinks it is the most fun job there is, because it is about spotting the unoccupied, valuable part of the chessboard and making decisions under uncertainty.

Clowes's most repeated advice is that time spent looking inside the building does little for you. Echoing Steve Blank, he says PMs should spend about 80% of their time thinking about what is happening outside, yet few do, because they get dragged into internal politics, scrum management and delivery. He says the job is not execution but finding reliable, differentiated value you can uniquely deliver to the market.

He gives two concrete habits. First, write every document from the perspective of the customer, the market and the competitor; PMs who adopt this improve almost immediately. Second, be data informed by backing up statements with anecdotes and bits of data, so colleagues come to believe in the opportunity and start wanting you to win.

When Rachitsky asks how to tell whether you are really doing enough of this, Clowes points to bias. People talk to the same people they always talk to, never seek out the counterfactual, and bring back an "omnibus edition of random stuff I heard on a Tuesday" instead of analysis. Without a structured process, he says, it is just more activity and no outcomes.

Before reaching for tools, Clowes says many teams get the foundations of qualitative research wrong. He cites the Nielsen number, the idea that you stop learning new things after roughly 7 to 14 interviews, and warns against both too little research and too much. Leading questions that coax customers into saying what you want to hear spoil the results before any model touches them.

Once the research is sound, he thinks it is a remarkable time to be a PM. He puts a batch of customer interviews into ChatGPT along with his strategy and asks where the strategy does not fit what customers said. He also pastes in a competitor's public positioning documents and asks whether the customers' words fit that company better, or asks the model to infer the rival's likely strategy.

Clowes says straight-up LLMs are mostly good enough, though Confluent also built internal tooling. It uses LLMs to summarize the many inbound customer requests and cluster semantically similar ones, showing which ideas are growing in popularity across hundreds or thousands of items. He borrows Sachin Rekhi's idea of a "Feedback River": smart PMs surround themselves with interviews, direct feedback, NPS data and competitor information, and the tools help find gaps without everything living in one person's head.

Clowes says models are very smart and also "insanely dumb": they know only what they were trained on or what you hand them in that moment. Information such as customer feedback or competitor moves decays in value very quickly, so an LLM is only as good as the data it is given and how recent that data is. For a PM using LLMs, the job is to gather as much relevant information as possible and bring it together.

The same logic applies to building AI into a product. Models will be somewhat replaceable and prompts matter as an ongoing investment, but the real differentiator is context. Clowes imagines an HR bot at a company like Workday: its smarts depend on employee records, benefits, the legal rules of the country where someone works and company policy, all linked together.

Teams that think they can wire in a model and a small data pipeline, he says, will be "very sad for a very long time." They will constantly struggle to get good, timely, well-structured data to the model, which is where he estimates 90% of the effort goes. Rachitsky links this to a Lenny and Friends Summit talk in which Mike Krieger described Anthropic putting more of its product effort into model and data work than into tweaking UX.

Asked whether AI-made clones will endanger Salesforce or Atlassian, Clowes says people badly underestimate where value is created. Cynically, products like Jira, Workday and Salesforce are forms on databases, so cloning them sounds easy. He says it is unbelievably hard, because the lock-in is not the data model or even the UX but the business rules that customers configure over years until the software becomes their own company's version.

He considers the two usual AI scenarios. If building forms on databases becomes trivially easy, new entrants' value drops toward zero and buyers may as well pick the winner. If agents replace the UI altogether, agents still have to operate within the rules of the system, which define what is allowed, so a headless Salesforce would still depend on those rules. He calls the idea that AI wipes out the differentiation of these products "a crazy fantasy."

His own forecast is that the dominant companies get stronger, or that more platform-like applications emerge covering several business functions' rules, which does not exist today. Rachitsky suggests distribution may matter even more, and Clowes agrees but notes that AI-generated cold email and LinkedIn spam are making breakthrough distribution more crowded and expensive. His answer for challengers is to meet the average bar and be different, for example by embedding data into the workflow, as he says the applicant tracking product Ashby does with hiring outcome data.

Clowes says he once called himself data-driven and has since changed his mind. Data, in his view, is more like a compass than a GPS: if you expect it to give you the answer you will be wrong or slow, and mostly it just tells you whether what you said is ridiculous or whether something might be there. Intuition is itself built from lots of data, so being data-obsessed is easy to overdo, and data should not be used as a weapon to force people your way.

On what to do when data contradicts intuition, he says to first believe your intuition and try to prove yourself right, because the most likely explanation for an insanely unintuitive result is that it is wrong. Check whether the data is representative and free of selection bias. Presenting a flawed analysis with authority costs you credibility, and he says it is better to show no analysis than a dumb one.

His tactical checklist is to look at what came before the result, what came after, and then one level up. If an onboarding intervention that seems to crush retention applies to only 2% of the funnel, it is meaningless. If those users all churn in week three, or their average sale price is lower, it fails the test too, because the final goal is happy customers paying money. Rachitsky shares that Shopify's growth head keeps long-term holdouts and found about 40% of positive experiments turn neutral, while Clowes cautions that a temporary lift is not necessarily a failure.

Clowes built the first B2B growth team at Atlassian around 2012, when growth hacking was a hot term in consumer products and nobody was sure the techniques would work in B2B. At the time, he says, strong distribution could paper over product flaws, and it was unclear whether software that sells itself was worth the investment. That early work, which wasn't yet called PLG, is now widely seen as worth it.

He describes growth teams moving through phases. First is a gold-rush phase of proving that any growth effect exists, where wins are plentiful but come from random searching. Then the team must turn that into a repeatable system, apply a PLG lens from acquisition to activation, retention and expansion, and work out how to fit in among the product, sales and marketing teams whose territory it touches. Many fail because wins look random or no one can explain them.

On the claim that PLG rarely works, Clowes argues that no natural force pushes companies to care about end users early on, because the buyer is the most important person and incentives drive behavior. If no one is measured on end-user success, it becomes a hobby, and that is a competitive opening for rivals. At Confluent the company runs self-serve and enterprise sales together, and he says getting the two motions to feed each other yields a resilient business with both many customers and large revenue.

Asked how he picks jobs, Clowes says that in hindsight his career has been a bingo card: he keeps filling boxes he lacks. He spent time in pre-sales, then at Atlassian learned to sell software at scale with no sales team, at Metromile built a consumer insurance product, and at Salesforce saw what he calls an insane distribution machine. Each stint gave him pattern-matching experience he treats as a superpower.

He thinks people too often stay in one domain or one type of company, when the likelier route to growth is to pick roles adjacent to what you know that still stretch you. A formative moment came at Atlassian when the general counsel, a lifelong lawyer who rarely spoke outside legal matters, recalled the company's earlier product strategy and showed that the current debate contradicted it. Clowes decided he wanted to be that kind of professional who can contribute across the whole company.

Rachitsky agrees that variety early in a career helps, and Clowes says the best people are shaped more like scribbles than Ts, deep in more than one thing. He has rarely regretted going deep on something outside his job, since the worst case is learning something he never uses and the best case is a tool that matters at a key moment. He says the reward for completing the card is getting to spend time on work that is both enjoyable and high leverage.

Clowes recounts a failure at an earlier company: a product to measure and reduce a business's environmental impact, such as by managing power draw. He says it was the wrong company, place and distribution, with no right to win, yet it stayed in the market for two years. The team finally killed it just when a customer wanted to pay millions, because signing would have locked them into it for good.

He feels responsible even though he wasn't directly accountable. After about six months it was obviously a zombie product, and the least he could have done was call it dead then instead of investing another year and a half. Rachitsky relates it to Wiz, where a new PM's questions about what they were actually building prompted a pivot.

For closing advice, Clowes warns that if your calendar rules you, nothing good happens, and nobody will force you to spend your time outside the building. He also cites a Colin Powell idea that deciding with less than about 30% of the data is a mistake, while waiting for around 70% or more means you've waited too long. Product managers are paid to make good decisions, and both too little and too much data can be fatal.