By Teresa Torres · producttalk.org · @ttorres on X · LinkedIn
Teresa Torres and Petra Wille challenge the popular claim that AI has made software delivery free, arguing that while building a single feature is far cheaper, producing a real, production-quality product is not. They warn that treating coding agents as free leads to tangled code, bloated and inconsistent data models, and maintenance debt that surfaces weeks later. They distinguish throwaway prototyping for learning from building to earn, and note that AI products are harder to ship than demos suggest because they require error analysis and evals. The core point is that delivery and discovery both keep mattering, and that the last stretch of a product is where most of the real work lives.
01Key takeaways
- Treat cheaper feature generation as a reason to watch total product cost, not as proof that delivery is free.
- Keep skilled engineers observing and steering architecture rather than outsourcing those decisions to a coding agent.
- Use throwaway prototypes for discovery, and commit to actually discarding them rather than shipping them.
- Budget months to years for the last stretch of polish, reliability, and maintainability in any product.
- Plan for error analysis, evals, and iteration when building AI features, since their behavior is non-deterministic.
- Rely on domain experts to review AI-generated outputs, because quality issues are often invisible until a specialist checks them.
02Key sections
- Why delivery isn't actually free
- Cheaper single features don't translate into cheaper products, especially when teams respond to cheap output by building far more features. The hosts argue against the framing that delivery is free and taste or discovery is all that remains.
- The hidden costs of 'free' feature building
- Rapid feature addition produces spaghetti code, duplicated logic, Frankenstein data models, and degraded performance. These problems compound and surface as maintenance burdens a few weeks later.
- Build to learn versus build to earn
- Disposable prototypes are a cheap, useful discovery tool, but only if they are actually thrown away. Production software requires architecture decisions and skilled engineers steering what AI generates.
- AI products are harder than they look
- Deterministic code and AI features differ, and AI features need error analysis, evals, LLM-as-judge methods, and prompt and orchestration iteration. Domain experts are needed to catch quality issues a demo hides.
- The last 30 percent
- Reaching a convincing prototype takes roughly 60 to 70 percent of the effort and happens quickly. Closing the remaining gap to a trustworthy product takes months to years.
03From the post
“Listen to this episode on: Spotify | Apple Podcasts Everyone's saying it: "Now that AI makes delivery free…" But is it? In this episode of All Things Product, Petra Wille and Teresa Torres pull apart one of the most repeated claims in product right now. They”
04Frameworks mentioned
Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.