Top Product Insights from AI Product Summit

96% of product teams now use AI consistently in their workflows, according to our recent survey of 379 product professionals. Zero, not a single one, reported going without it.

The verdict is clear: we're done imagining AI's potential. For product teams, AI has moved from pilot to production. It's the new baseline, not the exception.

But adoption is just the beginning. The real question isn't whether teams are using AI. It's how they're using it to create lasting value without losing sight of what makes products successful in the first place.

At this year's AI Product Summit, 15+ leaders shared hard-won lessons on navigating a world where pace has accelerated while fundamentals—customer pain, experimentation, trust—still remain the anchors. AI is no longer about if, but how and how fast.

Here's what matters according to them…

(Plus what we announced: Productboard Spark, our specialized product management AI agent).

Lessons for Product Leaders: Turning Hype Into Lasting Value

For product leaders, the challenge isn't just adopting AI—it's doing so without losing sight of what makes products successful in the first place. Speed matters, but only when paired with customer value, disciplined experimentation, and rock-solid governance.

1. Customer Pain + Differentiated Value = Successful AI Product Strategy

Paige Costello from Figma now ships AI features weekly, but speed isn't the only metric. When asked what changed from pre-AI product development, she was clear: "We're building for three years from now today. That is the only way to build."

The formula remains consistent across companies:

• Yashodha Bhavani from Box filters every feature through customer value, security, and confidence before shipping: "We sit for a moment and think: who are you trying to serve? How are you adding value? If the best way to add value is to be part of a link in the chain, then that's the way you should go."
• Emily Silberstein from Instacart prioritizes core user challenges before layering AI on top. She gave a concrete example: "73% of our users have a dietary need. Before AI, that was a huge problem—hard to solve with 17 million items. With AI, we found 1.4 billion data points about dietary needs in our catalog."
• Dharmin Parik from Uber AI echoed this: "Be very clear in terms of what you want to achieve. The key is to focus on outcomes."

The bottom line? Start with the pain point. AI is the "how," not the "why."

2. Experimentation Has Become the New Organizational Norm

Speed without structure = chaos. The question now: can we move fast AND learn fast?

James Evans (Amplitude): "Customers have much greater tolerance for experimentation with AI than pre-AI. They want to lean in with us—partially to learn how to build AI themselves."

Luke Behnke (Grammarly) on why interaction design matters: "Most agentic AI feels like command-line era computing. Grammarly spent 15 years bringing AI to your fingertips without making you prompt. That interaction model is what excites me."

Paige Costello's caveat: Teams must clarify experiment goals or risk wasting energy: "Design a roadmap with small, medium, and large bets. Get into customers' hands weekly and learn at different scales."

3. Trust and Governance Are Non-Negotiable

But experimentation without guardrails creates risk. As teams move faster, the stakes get higher. That's why trust and governance aren't optional add-ons—they're table stakes.

Netta Haiby (Microsoft): "Trustworthy AI means quality, governance, security, safety, privacy, observability, and control. Build trust in from the start."

She warns AI presents a different challenge than previous tech advancements: "It can generate code, take actions. We need to make sure the AI stays within intent—that it's not breaking what we intended."

John Kucera (Salesforce): "Without transparency frameworks, enterprises won't scale multi-agent systems. You need clear exec sponsors, clear KPIs, and humans staying accountable."

Lessons for Product Managers: Building AI Products

Building differentiated AI products requires bold vision, relentless quality focus, and technical depth to guide engineers. Our speakers in the Building AI products session spoke to just that…

4. Set your moonshot product vision in an AGI-world, then work backward

Aashi Jain (Google DeepMind): "Assume Artificial General Intelligence (AGI) is available within the next decade, likely much sooner. Ask yourself: what problems still matter?"

Her framework: Map hypotheses across three key dimensions—technology, users, and business—all built upon a crucial foundation of responsibility and safety.

• Technology: If AGI is possible, what other technical capabilities will / need to exist?
• Users: How will behavior change in the world of AGI?
• Business: How to make your business model viable? Will there be new business paradigms?
• Responsibility & Safety: What safeguards are non-negotiable?

Trace backward from your moonshot vision. “Identifying milestones and leading indicators helps clarify which capabilities are within reach versus far out, and reveals where your biggest risks are.”

On uncertainty: "There's a very good chance many of us will get it wrong, and that's okay. The point is to try our best now, so we've laid a thoughtful foundation for when AGI arrives."

5. Context Engineering Is the New Competitive Edge

Vision without execution = hallucination. Vikash Rungta (Alloi.ai, former Meta) brings it back to execution.

His thesis: Prompt engineering won't differentiate you, managing context will.

Why? LLMs are stateless. "If there's one thing you take away: LLMs don't have memory. Every question requires the entire context."

Your moat = three memory layers:

• Short-term: This session's interactions
• Mid-term: Compressed insights ("booking family trip to Italy, wants Florence, flexible dates")
• Long-term: Deep preferences (vegetarian, budget-conscious personally, prefers United)

Rungta warns against "dump everything": "Just because you have 10 million tokens doesn't mean you should use them. If we can't process tons of files, agents can't either."

The discipline: Compress insights. Isolate context by task. Feed the right information at the right time.

6. Those Who Define the Evals Define What Qualifies as "Good"

Aman Khan from Arize AI warns that when teams optimize for technical metrics like hallucination or retrieval, business metrics get lost.

Khan is direct: "The people that write the evals and think about the metric of what's good are the same people that are defining the quality of the AI product in the first place."

The solution: PMs and subject matter experts—not just engineers—should label data and define evaluation criteria because they understand customer quality.

Three questions Khan says every team should debate:

1. How many human labels do we need to feel confident in our eval system?
2. What happens when the eval is good but the human label disagrees anyway?
3. What happens when the eval is good but the business metric goes down? Who's responsible?

"If evals pass but business metrics fail, you have fundamental misalignment," Khan emphasizes. "You should make sure you're actually measuring the thing that matters to your company."

7. The Feedback Loop Between Production Data, Evals, and Iteration Separates Winners from Losers

But how do you know if your evals actually work? You can't just "vibe check" your way to production. That's where the continuous loop becomes critical.

Ian Cairns from FreePlay outlined the three-stage loop AI products need—and crucially, why traditional software development doesn't apply:

• BUILD (Iterate on prompts quickly): "You probably weren't having to continuously experiment to see if your product worked—you just designed it, built it, and tested it. That doesn't work anymore."
• TEST (Validate with representative datasets): Not just happy paths—abuse cases, edge cases, the weird stuff users actually do.
• Real Example: One S&P 500 company's VP of Engineering told Cairns their highest-leverage ritual is simple: "Every other Friday, we spend two hours locked in a room with domain experts. We just read through logs. That cadence has become key to how we improve quality."
• OBSERVE (Monitor production reality): "Build evals bottom-up from real failure modes, not top-down from theoretical KPIs," Cairns urges.

He emphasizes the continuous nature of this loop: "People who get to production without good evals always say they wish they'd started sooner. This isn't 'design, build, test, ship.' This is a loop you run weekly."

Lessons for Product Ops: Building the Product Operating Model

Product ops teams are uniquely positioned to make AI work at scale. They see across silos, understand workflows, and can orchestrate the systems that free PMs to do their best work. Our speakers in the “The Rise of the 10x PM” sessions shared how they’re building processes and motions that help future-proof their product org and business.

8. Deciding How to Decide Is One of the Most Important Things You Can Do

Chris Butler from GitHub argues that most teams lack deliberate decision-making processes, which kills speed in an uncertain AI era.

He outlined five stages every decision goes through:

1. Identification: When does a decision impact others?
2. Discourse: Generate options and evaluation criteria
3. Decision: One person decides (sometimes a consensus, but preferably not)
4. Communication: Share with stakeholders
5. Learning: Separate process quality from outcome quality

AI can augment each stage—detecting violated assumptions, simulating missing viewpoints, manufacturing dissent—but humans stay accountable.

Butler's key insight: "Consensus-driven cultures loop endlessly between discourse and decision. Being intentional about your method—like veto-based decisions—accelerates everything."

9. Product Ops Must Connect Tooling Across Silos to Unlock PM Productivity

The bottleneck has shifted: It's no longer engineering—it's product managers drowning in tools and context.

Ross Webb from Product Team Success argues that PMs are now the constraint.

His solution? AI agents that synthesize data across your entire stack. 

Webb demonstrated a working system built with N8N and LangChain:

Input sources: Productboard (priorities + roadmap), Linear (dev status), PostHog (user behavior), sentiment analysis tools

Outputs generated automatically:

• Executive summary with 2 critical insights
• Health scores across key metrics
• Prioritized recommendations (red/orange/green)
• Implementation roadmap

Result: "Within 5 minutes you should have insights and know exactly what your next steps are. You don't have to spend hours aggregating data—the agent identified critical funnel drop-off in comment posting and prioritized it immediately."

Webb's message to ops teams: Focus on saving PMs time, build a compelling ROI case for executives, and shift from managing individual tools to orchestrating intelligent systems.

He explains the root cause: "Product managers would rather go to a burnout workshop than actually go to the source of why they're burnt out. The source is busy work. Use AI to automate that so PMs can focus on the strategic work that's actually scary, because that's where they create value."

10. Executive Buy-in and Time to Experiment Unlock AI Success

Matt Johlie from Relativity emphasizes that ops teams must automate routine work to reclaim strategic time—which requires air cover from leadership to experiment.

His approach: "You need to secure air cover from your boss. If you're getting signals from the top that you need to transform how you do AI, you need time to explore tools, understand their strengths, understand their weaknesses, and discover opportunities. You need time to have those aha moments."

His team's focus areas (stack-ranked by impact):

1. Intelligence flows (product feedback → insights → action)
2. Rapid prototyping capabilities
3. Democratized data access
4. Knowledge management
5. Communication transformation
6. Process automation

The guiding principle: AI as co-pilot, not autopilot. "Co-pilot is intentionally distinct from autopilot. For product management, co-pilot thinking is required. Autopilot thinking is bad. Humans must be prepared to defend any AI output."

Product Development Fundamentals Haven't Changed (But the Execution Has)

The verdict? AI is transforming how we build products, but the fundamentals haven't changed. Customer pain, experimentation, trust, and clear decision-making still separate great products from mediocre ones.

What has changed is the speed of execution—and the need for new operating models.

Speed is table stakes. The teams that win will be those who:

• Embrace uncertainty
• Build tight feedback loops 
• Empower their people with intelligent systems 

In an age where AI compounds both speed and risk, the meta-process matters as much as the process itself. 

Want to see these principles in action? Check out Productboard Spark, our AI-powered product management agent that embodies the tight feedback loops, context-aware assistance, and workflow integration our speakers championed.