Context
Teresa Torres and Petra Wille use this episode to push back on the claim that AI changes everything. Their argument is not that AI is unimportant. It is that many teams are blaming AI for problems that were already there, such as weak testing, poor planning, shallow discovery, and losing focus on the core product. For PMs, that makes this a useful reset: it helps separate genuine product shifts from hype, and it shows where the old craft still matters.
The Big Idea
AI changes the speed and shape of software work, but it does not remove the need for testing, coordination, discovery, or product focus. In many cases, AI is simply exposing which teams already had those fundamentals and which teams never did.
The episode's strongest claim is that the biggest product failures being blamed on AI are often not AI failures at all. They are management, process, and strategy failures that became easier to see once output got faster.
Key Insights
Existing debt gets amplified
AI can make engineering output faster, but it does not fix weak delivery systems.
- Teams in the episode describe code piling up at the end of the workflow because testing and code review became the bottleneck.
- Teresa and Petra's read is that this is usually not a new AI problem. It is a sign that the team never had strong automated testing, test-driven habits, or reliable release discipline.
- That matters to PMs because faster output can create a false sense of progress. If quality checks do not scale with generation speed, the real system constraint simply moves downstream.
- Teresa also points out the irony that agents are often good at writing automated tests. If testing is still manual and handled by QA people feature by feature, AI is revealing how outdated the delivery setup already was.
Rituals survive the Scrum label
The episode rejects the idea that AI makes work rhythms obsolete.
- Petra pushes back on the claim that "Scrum is dead" if what people really mean is "we still plan work in one-week chunks, but renamed the chunk."
- The specific anti-pattern they call out is dropping ceremonies because estimates were wasteful, then pretending that coordination itself is now unnecessary.
- Their point is narrower and more useful: fake precision may be less defensible, but planning, review, demos, retrospectives, and measuring outcomes are still needed.
- Teresa makes the deeper point that multi-agent workflows actually increase the need for coordination. Once multiple agents and humans are contributing work in parallel, someone still needs a shared way to define what is being built, what done means, and who is accountable for what.
Cheap delivery still needs discovery
The episode strongly rejects the idea that discovery is obsolete because building is now cheap.
- Teresa's argument is that shipping without discovery only tells you one thing: whether the shipped idea worked or did not work.
- What it does not tell you is which assumption failed, why users did not respond, or how to iterate intelligently.
- That is why discovery still matters even if prototypes, features, and experiments are faster to create. Discovery is not just about reducing build cost. It is about learning in a way that separates one failure mode from another.
- The same logic applies to customer interviews. Story-based interviewing is not about asking people to imagine a brand-new AI product. It is about learning from the customer's past behavior, pain, and workarounds around the underlying problem.
- Teresa's test is simple: the solution may be new, but the customer is not. People still have lived experience that can reveal unmet needs and hidden opportunities.
Bad strategy kills SaaS
Teresa's most forceful claim is that struggling SaaS companies are often failing because of bad product and capital decisions, not because AI suddenly made software businesses impossible.
- Her Airtable example is concrete. She says views in her own setup now take about 45 seconds to load, even on small tables, which she reads as evidence that the company lost focus on its core product.
- Her explanation is strategic drift: Airtable chased AI product directions while the core product quality deteriorated.
- She also argues that overfunding makes this worse. If a company raises at a valuation that assumes extreme future growth, a solid business can still look like a failure because it cannot grow fast enough to justify the price investors already baked in.
- For PMs, this is a useful distinction. "The market changed" is sometimes true. But it is also a convenient story when the real problem is neglecting the product customers already pay for.
Personalization is the real shift
Where the conversation does see a genuine change is in hyper-personalized software.
- Petra describes using Claude Code to heavily personalize parts of her own business systems, including CRM, invoicing, and estimates.
- Teresa does not think every business owner or consumer will build software for themselves. Her prediction is different: more companies will offer products that are far more configurable for narrower customer needs.
- The example she uses is Notion's building-block approach. A product made of reusable parts can support much more customization than a rigid one-size-fits-all tool.
- For PMs, this means AI may shift product strategy away from giant generic suites and toward smaller, better-targeted products or platforms that let users assemble what fits them.
Mental Models & Frameworks
AI as an amplifier
Use this model when a team says AI "caused" a delivery problem.
- Ask what the system looked like before AI increased output.
- If testing was manual, release confidence was low, or review queues were already slow, AI probably amplified the weakness instead of creating it.
- This is a better diagnostic lens than treating every new bottleneck as proof that AI changed the nature of software work.
Coordination without fake precision
The episode draws a sharp line between useful coordination and useless estimation.
- Drop: story-by-story estimates that pretend the team knows more than it does.
- Keep: explicit time horizons, a shared weekly or short-cycle goal, and a review of whether the team delivered what it intended.
- Use this when a team wants to remove process. The question is not "Do we need rituals?" The question is "Which rituals reduce uncertainty, and which ones create ceremony without learning?"
Assumptions before experiments
Teresa's discovery argument implies a simple framework for faster AI-era shipping.
- List the assumptions: which user problem, behavior, willingness, and context must be true for the idea to work.
- Choose the learning method: interview, prototype, rollout, or usage data.
- Interpret failure correctly: if the launch fails, tie the result back to the assumptions instead of concluding only that the idea was bad.
Use this when delivery is cheap and teams are tempted to skip structured learning.
Building blocks enable personalization
Teresa uses Notion to point at a broader product pattern.
- Start with reusable components instead of fixed workflows.
- Let customers mix those components into workflows that fit their context.
- Use AI to make assembly easier, not to replace product thinking.
This is most useful for PMs building platforms, internal tools, or work software where customer needs vary by role, team, or operating model.
Decision Principles
Principle: Fix bottlenecks first
- When: engineering output increases after adopting AI tools.
- Why: more generated code does not matter if test coverage, review capacity, or release safety stays flat. The system bottleneck simply moves to the next unchecked step.
Principle: Keep cadence, drop estimates
- When: a team wants to remove planning rituals because detailed estimates feel wasteful.
- Why: the episode argues that coordination is still required even if task-level estimates are fiction. Teams still need a short-cycle commitment, a way to review outcomes, and a rhythm for improving how they work.
Principle: Diagnose the real cause
- When: a product, company, or market looks threatened by AI.
- Why: Teresa's SaaS argument is that "AI changed everything" can hide more basic causes, such as core-product neglect, bad capital structure, or a weak strategic focus.
Trade-offs & Nuance
Speed versus review capacity
Faster generation is only an advantage if review and validation keep up.
- More output can mean faster learning when automated tests and code review scale with it.
- More output can also mean more silent risk when manual QA or human review becomes the choke point.
- The trade-off is not speed versus quality. It is speed versus your current ability to safely absorb speed.
Less ceremony, same coordination
The episode is not defending heavyweight process for its own sake.
- Petra agrees that some old rituals, especially detailed estimation, may have always been poor uses of time.
- The nuance is that removing ceremony is not the same as removing coordination.
- A lighter process can be better. A process vacuum is usually worse.
Customization versus broad simplicity
Hyper-personalization creates product opportunity, but it is not free.
- More flexibility can fit niche customer needs better.
- More flexibility can also increase product complexity, support needs, and implementation overhead.
- Teresa's view is not that everyone will build from scratch. It is that products may need to become more modular so they can adapt without collapsing under one-size-fits-all design.
Growth capital versus product focus
The SaaS discussion is really about two pressures that can pull against each other.
- Venture-backed growth can fund ambition and market expansion.
- That same pressure can push teams to chase new bets, justify high valuations, and underinvest in the core product that existing customers already depend on.
- Teresa's Airtable and Miro examples suggest that PMs should ask whether product choices are serving users or serving the growth story.
Common Mistakes
Mistake: Blaming AI for debt
Teams can misread an AI-era bottleneck as a brand-new problem.
- If releases break old functionality because there is no automated testing, that is not a surprising side effect of AI.
- It is a legacy engineering quality problem that AI made harder to ignore.
- The better move is to treat AI adoption as a systems audit: where does faster output hit a weak control point?
Mistake: Renaming process as progress
The episode mocks the pattern of declaring a framework dead while keeping the same behavior under a new name.
- A one-week iteration does not become something radically different because a team stopped calling it an iteration.
- PMs should evaluate whether the team genuinely improved the workflow or just changed the vocabulary.
- Cosmetic anti-process rhetoric can hide the fact that the team still depends on the same basic coordination habits.
Mistake: Skipping discovery after launch
"Delivery is free" can become an excuse for shallow learning.
- Shipping first can be useful when the launch is designed as an experiment.
- Shipping without a clear view of the assumptions behind the idea makes failure hard to interpret.
- The better approach is to use AI to accelerate both delivery and discovery, not to replace one with the other.
Mistake: Neglecting the core product
The Airtable example is a warning for product strategy.
- Teams often get excited by new platform shifts and adjacent bets.
- If that energy comes at the cost of reliability, speed, or usability in the core experience, customers feel the decline quickly.
- PMs should treat core-product health as a strategy choice, not maintenance work that can be postponed forever.
Practical Application
Audit the testing path
Before celebrating faster AI-assisted delivery, map the full path from generated code to safe release.
- List every step: code generation, code review, automated tests, QA, staging, release, and post-release monitoring.
- Mark which steps are mostly human and which steps are automated.
- Look for the new queue. If code review or manual QA is where work now piles up, fix that constraint before pushing for still more output.
Preserve a weekly work rhythm
Keep a short planning and review cadence even if you drop formal Scrum language.
- Set a small time-box, such as one week.
- Agree on the outcome or work chunk for that period.
- End the period by reviewing what shipped, what moved, and what needs to change in the workflow.
This keeps alignment without pretending detailed estimates were ever reliable.
Map assumptions before shipping
For every AI-enabled feature or experiment, write down the assumptions that must be true.
- Which user problem is real?
- Which behavior do you expect to change?
- Which condition would make the idea fail even if the feature works technically?
When the launch underperforms, use that assumption list to diagnose the miss instead of treating the result as a vague failure.
Watch for core-product drift
Create a simple review that protects the product customers already depend on.
- Track core reliability and performance trends over time.
- Pair those signals with a list of major adjacent bets, including AI initiatives.
- Ask in roadmap reviews whether new bets are starving the experience that still earns the revenue.
Design configurable product blocks
If your product serves varied workflows, explore where a building-block model could replace rigid workflows.
- Identify the few reusable components customers combine most often.
- Test whether exposing those components would let a narrower segment tailor the product more effectively.
- Use AI to reduce the setup burden, not to hide unclear product architecture.
Questions to Consider
- Where is AI increasing feature output faster than your team can test, review, or safely release code?
- Which parts of your current planning process are real coordination, and which parts are just detailed estimates dressed up as certainty?
- If a recent AI-enabled launch underperformed, which specific user or business assumptions had you written down before shipping it?
- Is your team using AI excitement to justify adjacent bets while the core product experience is getting slower, less reliable, or harder to use?
- If AI made far more customization feasible in your product, which customer segment would benefit first from configurable building blocks instead of a fixed workflow?
Bottom Line
The episode's reset is simple: AI speeds up output, but it does not erase the need for sound product and engineering habits. The teams that benefit most will be the ones that use AI to strengthen testing, coordination, discovery, and product focus, while also looking for the real new opportunity in more personalized software.
Case Studies Mentioned
Airtable lost core focus
Teresa uses Airtable as the clearest example of a company being hurt by product choices rather than by AI itself.
- She says even small tables now take about 45 seconds to load in her workflow.
- Her interpretation is that Airtable got distracted by new bets, including AI-related directions, while underinvesting in the core product.
- The lesson for PMs is that platform shifts are not an excuse to let the main customer value proposition decay.
Miro hit a valuation trap
Teresa argues that Miro's situation is better explained by capital structure than by AI disruption.
- She cites roughly $600 million in annual revenue and a much higher $17.3 billion valuation.
- Her point is that a healthy business can still be pushed into a bad outcome if investor expectations require unrealistic long-term growth.
- The lesson is that product strategy can be distorted when the company must satisfy the story told at the last funding round.
Notion shows modular design
Notion appears as a positive example of where AI-era products may be heading.
- Teresa describes it as a system of building blocks that customers can mix and match.
- That architecture supports much more customization than a rigid feature set can.
- The lesson for PMs is that modular product design can become more valuable when AI lowers the cost of personalization and setup.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Airtable | Database-style SaaS for workflows and information management | Teresa uses Airtable as an example of a strong product that she believes declined because the company lost focus on core performance and customer value. |
| Miro | Collaborative whiteboarding software | Miro is discussed as an example of a successful business that may still face pressure because of high venture-backed expectations rather than because AI destroyed demand. |
| Notion | Modular workspace software built from reusable blocks | Teresa points to Notion's design as a model for how products can support more customization and hyper-personalization. |
| Claude Code | AI coding tool | Petra mentions using Claude Code to personalize parts of her own business systems, which supports the discussion about customized software. |
| Harvest | Time tracking and invoicing software | Petra briefly mentions replacing Harvest as part of her own move toward a more personalized setup. |
| Coda and Smartsheet | Workflow and database-style SaaS tools | Teresa cites Coda and Smartsheet alongside Airtable to show that markets often support several viable products rather than one winner taking all. |
