Beyond the Hype: A Product Builder's Guide to Where AI Is Actually Heading

TL;DR

• AI won't replace most jobs wholesale—it will unbundle tasks within roles, creating new workflows and value chains rather than simple automation. Product builders should design for augmentation, not replacement.
• The real opportunity lies in discovering new S-curves of value creation: AI enables entirely new products and business models that weren't economically viable before, not just cost savings on existing processes.
• We're still in the "what can we build?" phase, not the "what should we build?" phase—experimentation and rapid iteration matter more than perfect strategy right now.
• Distribution and go-to-market will determine winners, not just model quality. The best AI product isn't the one with the best underlying model; it's the one users actually adopt and integrate into their workflows.

---

Every product manager I know is drowning in AI noise right now. Half the takes are breathless hype about AGI arriving next Tuesday. The other half are cynical dismissals that "it's just autocomplete." Neither extreme helps you build better products.

That's why Benedict Evans' conversation with Lenny Rachitsky landed like a glass of cold water. Evans—who spent years as a partner at Andreessen Horowitz and has watched multiple tech waves crest and break—offers something rare: a measured, historically-grounded view of where AI is actually going, stripped of both utopian fantasy and reactionary fear.

For product builders, this matters. Your roadmap decisions today will compound over the next 3-5 years. Getting the framing right—understanding what AI actually changes versus what it doesn't—is the difference between building something that matters and chasing mirages.

Let's break down what Evans gets right, where I think the implications go even further, and what it means for what you're building right now.

The Job Displacement Paradox: Why "AI Will Take All Jobs" Misses the Point

Evans makes a crucial distinction that most AI discourse mangles: AI doesn't replace jobs—it unbundles them into tasks, then automates or augments specific tasks within those jobs.

This isn't semantic hairsplitting. It's the core dynamic that determines how value gets created and captured.

Think about what happened with spreadsheets. VisiCalc and Lotus 1-2-3 didn't eliminate accountants. They eliminated the specific task of manually calculating columns of numbers. But accounting as a profession expanded dramatically because suddenly many more business questions became answerable. The bottleneck shifted.

Evans points to a similar pattern with AI: "The question is not 'will AI take my job?' but 'which parts of my job will AI change, and what new things will I be able to do?'"

This reframing has massive implications for product strategy. If you're building AI products by asking "how do we automate this entire role?", you're likely solving the wrong problem. The better question: "What tasks within this role are bottlenecks? What becomes possible when we remove those bottlenecks?"

My take: The real product opportunity is in workflow reconfiguration

I think Evans is exactly right here, but I'd push it one step further: the biggest product opportunities aren't in automating existing workflows—they're in designing entirely new workflows that were impossible before.

Consider customer support. The obvious AI play is "automate tier-1 support." Fine. But the more interesting play is: what if support agents could handle 10x more complex issues because AI handles all the context retrieval, policy lookup, and draft response generation? What if the support org becomes a product feedback engine because agents now have cognitive bandwidth to actually listen?

That's not automation—it's augmentation that creates new value. And it requires rethinking the entire workflow, not just plugging AI into existing processes.

The product builders who win won't be the ones who automate the most tasks. They'll be the ones who reimagine what's possible when certain constraints disappear.

The S-Curve Problem: Why Cost Savings Aren't the Real Story

One of Evans' sharpest insights is about where value actually comes from in technology transitions. He argues that the transformative impact of AI won't primarily come from making existing things cheaper—it'll come from making entirely new things possible.

He uses the example of machine learning in the 2010s: "The really big opportunities weren't in making existing recommendation engines 10% better. They were in products like Instagram's Explore page or TikTok's For You feed—things that couldn't have existed without ML."

This is the S-curve argument: new technologies don't just slide you up the existing curve of efficiency. They create entirely new curves of value creation.

For AI, this means the most important question isn't "what can we automate?" It's "what becomes economically viable for the first time?"

Where the new S-curves might be

Evans is appropriately cautious about predicting specific new S-curves—they're inherently hard to see in advance. But we can identify characteristics of where to look:

Personalization at scale: Products that require deep customization for each user but were previously too expensive to deliver. Think personalized education, individualized health coaching, or bespoke financial advice.

High-frequency expertise: Scenarios where you need expert-level judgment repeatedly, but the economics never worked. Legal review of every contract clause. Security analysis of every code commit. Editorial judgment on every piece of content.

Coordination across complexity: Problems that require synthesizing information across dozens of sources and stakeholders. Multi-vendor procurement. Clinical trial design. Urban planning.

The pattern: these aren't about doing existing things faster. They're about doing things that were previously impossible at any speed.

The "What" vs. "Should" Inflection Point

Evans makes a subtle but critical point about where we are in AI's maturity curve: we're still firmly in the "what can we build?" phase, not yet in the "what should we build?" phase.

This has profound implications for product strategy.

In the "what can we build?" phase, the primary constraint is imagination and execution speed. The winners are the teams who ship fast, learn from users, and iterate. Strategy is emergent, not prescribed.

In the "what should we build?" phase—which comes later, once the technology stabilizes—the constraint shifts to distribution, brand, and strategic positioning. The winners are the companies with the best go-to-market, the strongest moats, the clearest positioning.

We're not there yet with AI. Which means:

For product builders right now, velocity matters more than perfect strategy. You learn more from shipping a flawed v1 and watching how users actually interact with it than from six months of strategic planning.

The companies winning in 2027 might not be the ones with the best strategy today. They'll be the ones who learned the most between now and then by actually building and shipping.

Your competitive advantage is speed of iteration, not model quality. Evans notes that model quality is rapidly commoditizing. The differentiation comes from product design, user experience, and how well you solve a specific user's specific problem.

Distribution Eats Model Quality for Breakfast

This brings us to what I think is Evans' most important point for product builders: the best AI product isn't the one with the best model—it's the one users actually adopt.

We've seen this movie before. Google wasn't the first search engine. Facebook wasn't the first social network. The iPhone wasn't the first smartphone. In each case, the winner wasn't determined by who had the best technology first—it was determined by who nailed the full product experience and distribution.

For AI products, this means:

Integration beats isolation. A pretty-good AI feature embedded in a tool users already use every day will beat a slightly-better AI feature in a standalone app they have to remember to open.

Workflow fit beats raw capability. An AI assistant that understands your specific workflow and integrates with your specific tools will beat a more powerful general-purpose AI that requires you to change how you work.

Trust and reliability beat bleeding-edge features. Users will choose the AI product that works consistently 95% of the time over the one that's amazing 70% of the time and breaks catastrophically 30% of the time.

My take: The AI product landscape will look like the SaaS landscape

I think we're headed toward a world where AI capabilities become table stakes—every product category will have multiple players with roughly equivalent AI features—and differentiation comes from the same factors that differentiate SaaS products today:

• Domain expertise and workflow understanding
• Integration ecosystem and data connectivity
• User experience and interface design
• Pricing and packaging strategy
• Brand and trust
• Customer success and support

The implication: if you're building an AI product, spend less time obsessing over model fine-tuning and more time obsessing over user onboarding, integration depth, and go-to-market.

What This Means for Your Roadmap

Let's get concrete. If you're a product builder trying to figure out where to place your bets, here's how I'd translate Evans' insights into action:

1. Map the task graph, not the job description

For whatever domain you're in, break down the actual work into discrete tasks. Which tasks are bottlenecks? Which tasks, if removed or accelerated, would unlock new capabilities?

Don't ask "can AI replace this role?" Ask "can AI remove the bottleneck tasks that prevent this role from creating more value?"

2. Look for new S-curves, not efficiency gains

Yes, cost savings matter. But the bigger opportunity is in products that become viable for the first time.

Ask: "What would we build if [specific task] were free and instant?" That question reveals new product categories, not just features.

3. Optimize for learning speed, not perfect strategy

Ship something. Watch what users do. Learn. Iterate. Repeat.

Your strategic clarity will come from empirical learning, not from upfront analysis. The teams that ship and learn fastest will have the best strategies in 12 months.

4. Build for integration and workflow fit

Standalone AI apps face a steep adoption curve. AI features embedded in existing workflows have a much easier path.

If you're building standalone, your distribution strategy needs to be exceptional. If you can integrate into existing tools, that's often the faster path to adoption.

5. Solve for trust and reliability

Users will forgive a lot of missing features. They won't forgive a product that breaks their workflow or produces unreliable output.

Invest in guardrails, error handling, and graceful degradation. Make sure your AI product is reliable enough to integrate into critical workflows.

The Uncomfortable Truth About Timing

Evans doesn't shy away from one of the most uncomfortable truths in tech: timing is everything, and timing is mostly luck.

You can have the right insight, build the right product, and still be too early or too late. The companies that win are often the ones that happened to be in the right place at the right time, not the ones with the best foresight.

This is simultaneously liberating and terrifying for product builders. Liberating because it means you don't need perfect foresight—you just need to be in motion, learning, and positioned to capitalize when the moment arrives. Terrifying because it means even great execution doesn't guarantee success.

The practical implication: build optionality into your strategy. Don't bet everything on a single vision of how AI will evolve. Build products that create value today while positioning you to pivot as the landscape clarifies.

Where I Think Evans Undersells the Disruption

I agree with almost everything in Evans' analysis, but I think he's slightly too sanguine about one thing: the speed at which AI will reshape competitive dynamics within existing categories.

He's right that AI won't eliminate most jobs or industries. But I think it will create brutal winner-take-most dynamics within categories faster than previous technology waves.

Here's why: AI products have unusual scale economies. The more users you have, the more data you collect, the better your product gets, the more users you attract. This flywheel spins faster than traditional SaaS network effects.

Combine that with the fact that switching costs are lower for AI products (users are less locked into specific interfaces when the interface is conversational), and you get a recipe for rapid market consolidation.

The implication for product builders: speed to market matters even more than usual. Being second or third in your category might be fine in traditional SaaS. In AI products, being second might mean you never catch up.

The Real Question: What Are You Building?

Evans' conversation is valuable precisely because it strips away the hype and forces you to think clearly about fundamentals. Not "will AI change everything?" but "what specific changes create what specific opportunities?"

For product builders, the actionable insight is this: the AI revolution won't look like a revolution from the inside. It'll look like a thousand small decisions about which tasks to automate, which workflows to redesign, which new products to experiment with.

Your job isn't to predict the future of AI. Your job is to understand your users deeply enough to see which AI capabilities remove which bottlenecks for them—and then to ship products that capitalize on those insights faster than anyone else.

The teams that do that consistently over the next few years will build the defining products of the AI era. Not because they had perfect foresight, but because they were in motion, learning, and building when it mattered.

So: what are you building? And more importantly, what are you learning from what you're building?

That's the question that matters. Everything else is just commentary.