--- title: "When building gets cheap, deciding gets expensive" description: "What the 2025 data says about AI and product speed, why the gain feels bigger than it is, and where the product role goes once building stops being the limit." url: https://torn.studio/en/insights/when-building-gets-cheap locale: en published: 2026-09-07 --- # When building gets cheap, deciding gets expensive > **Short answer:** AI tooling has made producing software cheap, and the gain is smaller than it feels: METR’s 2025 trial found experienced developers 19% slower with AI while believing they were 20% faster. The constraint moves to deciding what to build, how to know it worked, and whether users can trust it. That is the product role, and it got harder. Every product team has felt it this year: a feature that took a sprint takes an afternoon, and the backlog empties faster than anyone can refill it with things worth building. This article answers what that does to the product role, using the measurements published so far, with each figure’s sample and date attached. ## Is building actually cheaper? Yes, and the honest figure is smaller than the feeling. The 2025 DORA report, published by Google Cloud on 23 September 2025 from nearly 5,000 respondents, finds 90% using AI at work and, for the first time, a positive relationship between AI adoption and software delivery throughput. The same report finds that AI adoption continues to have a negative relationship with delivery stability: more change ships, and more of it breaks. Lenny Rachitsky and Noam Segal’s survey of 1,750 product managers, engineers, designers and founders, published 23 December 2025, found more than half saving at least half a day a week on their most important tasks. The same survey found 92.4% reporting at least one significant downside to the tools. Both numbers are true at once, and a team that quotes only the first is planning on a fiction. ## Why does the gain feel bigger than it is? Because the people experiencing it misjudge it in a consistent direction. METR ran a randomized controlled trial, published 10 July 2025: sixteen experienced open-source developers, 246 real tasks from their own repositories, each task randomly assigned to allow or forbid AI tools. With AI allowed, the developers took 19% longer. Before the study they had forecast a 24% speed-up, and after living through the slowdown they still believed AI had made them 20% faster. Sixteen developers is a small sample and the tools were early-2025 models, which the authors say themselves. What the trial establishes is narrower and more useful than a verdict on AI: perceived speed is an unreliable instrument. A product team that wants to know its real velocity has to measure it — lead time, change failure rate, the DORA four — with the tools on and off. ## Where does the constraint move? To judgment. When production stops being the limit, three decisions set the outcome, and none of them gets faster with a code assistant. - What to build: which problem, for whom, how often it occurs, and what it costs to wait three months. - How you will know it worked: the number that has to move, written down before the build, with the date it will be read. - Whether a user can trust it: for an AI feature, the cases where it may answer, the cases where it must refuse, and how confidence is shown. Marty Cagan wrote on 30 December 2024 that “the PM role becomes more essential but also more difficult with generative AI-powered products, not less”, reserving his enthusiasm for the combination of “someone with very strong judgement” and the tools. Teresa Torres said the same thing from the discovery side in April 2026: “Building is very cheap now. That doesn’t mean we should build every idea we have.” Both are describing a role whose scarce input is judgment about what deserves to exist. ## What does a product team do with this? Three practical moves follow, and each is cheap to start. - Measure velocity with the tools on and off for a month, and read the change failure rate beside the throughput. DORA’s instability finding is the one most teams discover late. - Write the decision before the build: one page with the problem, the alternatives on the table, what was set aside, and what has to be true in two months. - Spend the freed hours on discovery and on evals. An interview synthesized by a person and then by a model, and an AI feature specified as test cases before the first prompt, are where the week now goes. The prioritization method is written up in [deciding what to build next](https://torn.studio/en/insights/deciding-what-to-build-next). For a founder with a runway, the same logic decides what the first build has to prove, and the [founders page](https://torn.studio/en/for/founders) sets out that sequence. ## How the studio works with this Torn Studio’s answer is a delivery model: one person holds the product role and builds with AI tooling, so a decision made on Monday is in code the same week. The cost is stated as plainly as the gain. One person’s week is the capacity, and a team that needs three product roles in parallel is better served by hiring them. Product Management is priced per engagement, with the scope set in advance and the price fixed before work starts. **Read next** - [Keeping quality when AI writes the code](https://torn.studio/en/insights/quality-when-ai-writes-the-code) - [Product Management at Torn Studio](https://torn.studio/en/services/product-management) ## Common questions ### Does AI make our product team faster? Measurably yes on throughput, and by less than it feels. The 2025 DORA report finds a positive relationship between AI adoption and delivery throughput alongside a negative one with stability, and METR’s trial found experienced developers 19% slower while believing they were 20% faster. Measure it with the tools on and off. ### Which numbers should we track when the team adopts AI tooling? The four DORA measures: lead time for changes, deployment frequency, change failure rate and time to restore. Read the change failure rate beside the throughput, because the 2025 report finds AI raises the second and worsens the first. ### Is the METR result a verdict on AI coding tools? No. Sixteen developers, early-2025 models, and tasks in repositories the developers knew deeply. What it shows reliably is that perceived speed is a poor instrument: the same people forecast 24% faster, felt 20% faster and measured 19% slower. ### What should a product manager spend the freed hours on? Discovery and evals. Synthesizing every interview by hand before a model does it, writing the decision page before the build, and specifying each AI feature as test cases with a tolerance. Those are the three places where judgment now sets the outcome. ### Do we still need a product manager if engineers can build anything? Yes, and Cagan’s December 2024 reading is that the role gets more essential and harder. Someone has to decide what deserves to exist, how success will be read, and what a user may trust. Building capacity does none of that. ### How does Torn Studio work when it takes the product role? One person holds the role and builds with AI tooling, priced per engagement with the price fixed before work starts. The gain is a same-week loop from decision to code; the limit is one person’s capacity, which the studio names before an engagement starts. ## Sources - [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) — Backs the 16 developers, 246 tasks, 19% slower with AI, the 24% forecast and the 20% perceived gain. - [Announcing the 2025 DORA Report — Google Cloud](https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report) — Backs the nearly 5,000 respondents, 90% AI use, the throughput relationship and the negative relationship with stability. - [AI tools are overdelivering: results from our large-scale AI productivity survey — Lenny’s Newsletter](https://www.lennysnewsletter.com/p/ai-tools-are-overdelivering-results) — Backs the 1,750 respondents, more than half saving at least half a day a week, and 92.4% reporting at least one significant downside. - [AI Product Management 2 Years In — Silicon Valley Product Group](https://www.svpg.com/ai-product-management-2-years-in/) — Backs Cagan’s 30 December 2024 line that the role gets more essential and harder, and the point about strong judgment combined with the tools. - [“Building is cheap now, but don’t build everything”: Teresa Torres on AI product management — airfocus](https://airfocus.com/blog/teresa-torres-ai-product-management/) — Backs Torres’s April 2026 quote that building is cheap and that this is no reason to build every idea.