← All topics

AI in Product

How AI is changing the way products, teams and careers are built

Eyal David · Updated:

AI on Product Builder covers how AI changes the way products get built: both the work itself and the products that come out of it. In these conversations, founders, product leaders and engineers explain why a product manager's value is moving from execution to judgment, why context and planning matter more than any single prompt, how to move fast without giving up ownership of quality, what makes customers trust and pay for an AI product, and where a real moat comes from when the model belongs to someone else.

What founders said

Key takeaways

From execution to judgment

  • A product manager's value has moved from execution to sense: AI gives you five good solutions in a second, and the PM's job is to know which four to kill. AI won't do that for you. — Shira Moyal
  • Decision-making is the new core skill: you used to choose between A and B, today between A and H, fast and at scale. Running fast is only worth something if you know where you're going. — Shira Moyal
  • When AI raises engineering velocity, the bottleneck moves to product managers. That's why Daniel plans to hire fewer engineers and more product managers and designers in 2026, changing the 1-to-10 ratio he used to run with. — Daniel Marcous
  • The boundary between Product and R&D is blurring: a PM who understands the technical complexity (where data is stored, how the system is built) gets a 'superpower'. — Nadav Abrahami
  • The future is a conversational interface everywhere, and product people no longer write requirements with their legs crossed — they're half salespeople who get on calls, do discovery, and accompany deployment. — Rami Segal
  • Code has become cheap, so a good developer is one who can identify customers' problems and solve them, not one who just writes code. — Ilan Peleg

Context and planning before the prompt

  • A product copilot is like a genius intern: without context it returns generic outputs. Build it as a project with knowledge (what you'd give a new hire in onboarding), custom instructions that ask for pushback rather than people-pleasing, and a separate chat for each product project. — Zeevi Arnovitz
  • Setting custom instructions solves the people-pleaser problem: you ask the model for pushback instead of flowing along with every idea. — Zeevi Arnovitz
  • Managing context for the LLM is the new, most important skill — building, saving and pointing the model at the right pieces of information (e.g. markdown files). — Nadav Abrahami
  • When building with AI, planning beats intuition: one big prompt plus spot fixes leads to spaghetti. Write a PRD, convert it into a set of prompts in a logical order, and define for each prompt what you expect to get and its acceptance criteria. — Rani Bar Hama
  • To prevent regressions, ask the AI for the list of tests that prove each feature works alongside the feature itself, and freeze parts that already work (for example in Base44) before moving on. — Rani Bar Hama
  • When working with agents, the most important thing to read is the reasoning, not the code. Almost every failure comes from missing context, and every wrong decision by the agent shows what needs to be added to the knowledge base. — Daniel Marcous

Speed without giving up ownership

  • Fast is slow, and slow is fast: AI speeds up the first step, but if you don't measure the whole cycle (quality, QA bugs, reviewer load) you pay for the shortcut for years. — Amir Frankel
  • AI is not a substitute for ownership: Zero still requires two reviewers and expects every developer to review their own code hard before pushing the pain onto others. — Amir Frankel
  • Vibe coding is magical for fast iteration but can produce an unmaintainable 'Frankenstein'; in the enterprise you need processes, best practices per team/organization, and ownership of AI-written code. — Dedy Kredo
  • There's a big gap between "using AI" and working with it well; unmonitored use creates AI slop — more work for the team and cleanup after the machine. — Galit Galperin
  • Speed isn't the goal: move fast in ideation, not with eight features into production; gatekeepers, review and tests become more important, not less. — Nadav Abrahami
  • You own your outputs: "the AI put that in" is no excuse, and every weak output also erodes the trust of people who aren't believers yet. — Zeevi Arnovitz

AI products customers trust and pay for

  • People won't pay for a mental-health chatbot, but organizations pay for a hybrid service: Kai cracked its business model only when it added a human clinical team for intake and plan tailoring, with the AI accompanying users day to day. — Alex Frenkel
  • With AI products, show customers the output, not the UI: Gong sends customers real summaries and even spreadsheets very early, and runs 300-400 betas in parallel behind feature flags. — Eilon Reshef
  • Measure GenAI with an Elo rating: when two answers are both correct, people pick the better one, and each model gains or loses points as in chess. That yields a number that quantifies improvement between versions. — Eilon Reshef
  • The main barrier to enterprise adoption is reliability: a model that's right 90% of the time is unacceptable when mistakes aren't allowed. — Amnon Morag
  • Not every automation should reach 100%. Honeycomb samples part of every model's decisions and routes them to human underwriters, to measure accuracy and correct the model over time. — Roy Feig
  • To build trust when launching agents, you keep a human in the loop even when the system can run on its own — the user needs to believe in the tool and press the button themselves. — Rami Segal

Moats and big bets

  • A moat in the AI era doesn't come from the model: even if 99% of the value is the model, your 1% is only worth something if it rests on industry-specific knowledge or data that is very hard to obtain. — Yaniv Golan
  • In an AI-Embedded company it's sometimes better to wait for a capability in the model than to fight for it for months: more than once a new model version ships it built-in two days after the team finished building. — Yaniv Golan
  • There's no AI without data: Salesforce's strategic move (Data Cloud, bringing Datorama and Tableau inside) was meant to bring the customer's data back into the platform, so it wouldn't become merely a back-end System of Records. — Rami Segal
  • When anyone can build anything, working closely with the world's biggest customers is the real moat, not the technology. — Ilan Peleg
  • In a GenAI startup the game is win big or fail big: you need a strong point of view, an insight people aren't paying attention to today, and to go with it through a big bet. — Dedy Kredo
  • The sweet spot is usually a few months ahead: don't sink the bulk of your effort into something that will be obsolete in two months, but don't build too far out either. — Dedy Kredo

Frequently asked questions

Where does a product manager's value lie in the AI era?

According to Shira Moyal, a PM's value lies in sense: the ability to ask the right questions, understand the business, customers and strategy, and make decisions quickly. Execution, like tracking timelines and writing tickets, keeps shrinking. Go to the episode

Why is managing context for the LLM such an important skill?

According to Nadav, the hardest problem when working with LLMs is context — how much information to expose, where, and at what cost (tokens and accuracy). Whoever knows how to build and preserve the right pieces of context (e.g. saving decisions in markdown files and referencing them) gets the strongest power in vibe coding. Go to the episode

What is AI slop and why is it dangerous for teams?

AI slop is an insufficiently-vetted AI output passed along as if it's finished. Galit explains it "costs you twice": someone else on the team has to fix both your work and what the machine did poorly, and it also signals you're still at a basic usage level. It happens mainly when the machine isn't given enough context. Go to the episode

Why agentic systems and not just a better language model?

A language model predicts the next word token by token, so it struggles to plan, execute, correct itself along the way, and know how deep to search. An agentic system wraps the LLM, uses it as one tool out of a broad set, and enables controlled autonomy that breaks down a complex, multi-step problem in a way that provides guarantees an enterprise can rely on. Go to the episode

What is Vertical AI and why does it have a moat?

Vertical AI companies have teams with deep knowledge of a specific industry, and usually unique data that is very hard to obtain. Even when they use off-the-shelf models, that knowledge and data create value that is hard to copy. Examples from the episode: Aidoc in medicine and Eleos in psychological therapy. Go to the episode

All episodes on this topic