← All topics

AI in Product

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

15 episodes12 timestamped moments15 guests

Every Product Builder episode about AI: how product managers, founders and engineers bring AI into products and daily work, and how the PM role is changing.

From building copilots and agents, through AI-assisted software development and AI hardware, to the mindset it takes not just to survive the shift but to thrive in it.

Start here

Ela Hatzav | Wix | An unconventional career move into AI PM. The product world is changing

Ela Hatzav

מעבר קריירה לא שגרתי משתלם: מפסיכולוגיה ותיירות, דרך עשור ב-non-profit (תגלית), ועד Product Management ב-Wix — הצטיינות בכל תחום (לאו דווקא אקדמי) היא מנבא חזק להצלחה בפרודקט.

Go to the episode →

Learn about AI in Product

AI inside the PM's daily work

AI is no longer just a feature in the product; it's part of the product manager's own workflow, from discovery to specs to running experiments. Ela Hatzav of Wix describes a playbook that's still being written: evaluation, prompting and short cycles are now part of the job, scoping takes up far more of it, and the AI PM works right next to engineering, down to the architecture.

Zeevi Arnovitz of Wix builds a standing copilot for product work on three pillars: general knowledge, personal instructions and per-project context. He treats it like a brilliant intern: without a good brief the output is generic, and without custom instructions that ask for pushback, the model just agrees with every idea.

Shay Shitrit of Augury fed Claude Gong transcripts, Slack threads and CSM summaries and distilled discovery insights in hours instead of weeks. Itai Horn of Resident built a CRO Agent that took concurrent tests from 2–4 to 8 and the success rate from 10–20% to 35%. Shay adds that internal adoption doesn't happen by itself: you need to share success stories and make the tool one link away.

When AI writes the code: understanding, review and ownership

AI writes code fast, but someone still has to understand what was built, and that's the new challenge. Nadav Abrahami, Wix co-founder and founder of Dazl, explains that teams reach 80% quickly and then get stuck on the last 20%, either because they don't understand what was built or because the change simply doesn't fit a prompt. In his view, managing context for the LLM is the most important skill today, and gatekeepers, review and tests matter more, not less.

Dedy Kredo, CPO of Qodo, sees developers becoming orchestrators and reviewers of agents. Vibe coding is great for fast iteration, but in the enterprise it can produce an unmaintainable "Frankenstein", which is why Qodo focuses on code integrity: verifying at scale that model-written code does exactly what the spec says and breaks nothing.

On the organizational side, Ohad Peri, CTO at Bllink, cites up to a 50% improvement in Time to Merge with tools like GitHub Copilot and Cursor. Galit Galperin of Hi Mindset warns about the flip side: unchecked AI output passed along as finished work, AI slop, just creates more work for the team cleaning up after the machine.

Trust, data and reliability in AI products

What slows AI product adoption is usually not model capability but trust and data. Amnon Morag, VP Product at AI21 Labs, says a model that's right 90% of the time is unacceptable when mistakes aren't allowed, so the big value will come from controllable autonomous systems that wrap the LLM and give enterprises guarantees they can rely on.

Rami Segal of Salesforce puts it simply: there's no AI without data, which is what drove Data Cloud. When they launched Agent Force they kept a human in the loop even where the system could run alone, so users would trust the tool and press the button themselves. Ohad Peri of Bllink learned this the hard way: onboarding through an AI chat failed not technically but psychologically, because people handing over money want to know there's a human behind the process.

Tse'ela Mida of Intuitive, the company behind da Vinci, shows the extreme case: in medicine a product has to prove efficacy and safety, and AI regulation demands transparency about how the model was trained and how it's monitored. Medical data is a barrier to entry, so she built a "Data Factory" with a clinical team that labels anatomy and surgical steps.

Guest experts

Frequently asked questions

What changes in product management in the AI era?

According to Ela Hatzav of Wix, the playbook is still being written. Evaluation, prompting and short cycles have joined the role, scoping has become central, and the work sits very close to engineering, including thinking about architecture and bridging non-deterministic input to deterministic output. Go to the episode

How do you build an AI copilot for product work?

Zeevi Arnovitz of Wix builds it on three pillars: general knowledge, like onboarding a new hire; custom instructions that define how it should answer, including pushback; and context specific to each project. In practice it's one project holding the knowledge and instructions, with a separate chat under it for each product initiative. Go to the episode

How can you use AI for product discovery?

Shay Shitrit of Augury gathered Gong call transcripts, Slack context, CSM/RSM summaries and decks, fed it all into Claude thanks to its large context window, and distilled insights in hours instead of weeks. After validating with internal stakeholders, the output went into the backlog and roadmap. Go to the episode

Why is context management such an important skill when working with LLMs?

According to Nadav Abrahami, the hardest problem in working with LLMs is context: how much information to expose, where, and at what cost in tokens and accuracy. Whoever can build and maintain the right pieces of context, for example saving decisions in markdown files and pointing the model to them, gets the most out of vibe coding. Go to the episode

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

Galit Galperin defines AI slop as AI output that wasn't properly checked and gets passed along as if it were finished. She says it costs you twice: someone on the team has to fix both your work and what the machine got wrong. It happens mostly when the machine isn't given enough context. Go to the episode

Why are enterprises hesitant to adopt AI?

Amnon Morag of AI21 Labs points to reliability: a model that's right 90% of the time is unacceptable when mistakes aren't allowed. That's why he expects the value to come from agentic systems that wrap the LLM, use it as one tool among many, and provide controlled autonomy with guarantees an enterprise can trust. Go to the episode

Resources mentioned

All episodes on this topic