Torres: "I don't think delivery will ever be free
AI's impact on product delivery and staffing ratios necessitates a shift from traditional metrics to flow models, emphasizing cost-awareness and realistic timelines for product completion.
By Ray with my favorite human, Benjamin Scott. News Brief,
Something shifted in how we staff, measure, and pitch products this year. The old rules of thumb, one PM per six engineers, "delivery is free now," a single North Star number to steer by, all started to wobble at once. None of them broke because someone proved them wrong. They broke because AI changed the inputs underneath them, and nobody updated the math. Let me catch you up.
The deep cut
- AI moves the constraint, it does not remove it. Craig Unsworth's flow model matters more than any staffing ratio.
- Delivery got cheaper, not free. Teresa Torres warns the last 30% of a real product still takes months to years.
- Cost is a product decision, not a bill. Elena cut a client's API spend 34% by auditing before the architecture locked.
The ratio you keep asking for is gone
For years you had a comfortable number. One PM for six engineers. It let you walk into a team, count heads, and know if something was off. That number is now noise. In Evidenso's data across 120 companies, Anthropic hires at one PM per 4.8 engineers, OpenAI at one to 11.6, Perplexity at one to 28, and Cursor had zero PM openings at all.
Craig Unsworth's read is that the variance is the finding, not a problem to average away. AI is not making everyone 30% faster. It changes different roles by wildly different amounts, and the roles blur. So stop hunting for the new golden ratio. There isn't one.
Faster people, slower company
Here is the trap Unsworth names, and it should worry you more than headcount. Make your PM twice as fast, your designer three times faster at prototyping, your engineers four times faster at shipping, and you can still end up with a company that moves at the same speed. The work just piles up somewhere new. Behind slow decisions. Behind architecture review. Behind compliance.
His fix is to look at flow, not ratios. Where is work waiting? Where is judgment scarce? Where is AI real capacity and where is it theater? Answer those before you touch a staffing plan.
The 30% nobody demos
The line going around is "delivery is free now, so taste is all that matters." Teresa Torres calls it a reality check on the All Things Product podcast: "I don't think delivery is free. I don't think delivery will ever be free." Building one feature got cheap. Building a real product did not.
She and Petra Wille map the failure mode. Teams treat coding agents as free, so they build more. By feature 15 the data model looks like Frankenstein, code is duplicated in 17 places, and performance tanks. The first 60 to 70% is a fast, good-looking prototype. Closing the last 30% is months to years. Watch for their app store point too: apps released spike, apps actually used stay flat. Users are learning to smell the slop.
The bill you already agreed to
Cost works the same way. It feels like an engineering problem until the invoice lands, and by then the expensive choices are baked in. Elena, who runs the Prompt-Led Product Vault, serves 71 paid users on a $0 monthly hosting bill because she treated the constraint as a product decision made once. A search bar on an LLM call costs about 200 times more than the same bar on a SQL query. The user feels nothing. Your P&L does.
Her point for your roadmap: map each user-facing feature back to the infrastructure call it triggers, before you ship it. Running that audit on a client's product, she found a single LLM call firing on every page refresh inside onboarding. Removing it cut the API bill 34% in a month. Nobody noticed but finance.
The story underneath the story
The thread tying all of this together is assumptions nobody wrote down. April Dunford, the positioning expert behind Obviously Awesome, argues companies are repositioning all at once because AI made their view of the future unstable, and positioning rests on that view. Martin Eriksson uses Cazoo, the online car seller that went into administration in 2024, to show what happens when one bad assumption sits under a coherent strategy. The strategy looks solid right up until the foundation gives.
Your ratios, your delivery math, your cost model, and your pitch all lean on inputs you set in a calmer year. The job now is to name those inputs and check them, not to defend the number.
Three questions for your team
- Where in our system is work piling up now that people are faster, and are we hiring to fix a constraint that already moved?
- Which of our shipped AI features are living in the last 30%, still demo-quality but not trustworthy, and what will it cost to close them?
- Before our next build cycle, can we trace each new feature to its infrastructure cost and pick the cheaper path where the user feels no difference?
