The Hard Truths About Building Products in the AI Era: Beyond the Hype

Keith Rabois has been through PayPal, LinkedIn, Square, Khosla Ventures. When he talks about what separates products that work from expensive failures, it’s not theory. It’s pattern recognition from hundreds of investment decisions.

And what he’s saying about AI goes against almost everything the Brazilian market is doing right now.

The problem isn’t the technology. It’s the wrong question.

The question most PMs are asking: “How can I use AI in my product?”

The question they should be asking: “What user problem justifies the complexity and cost of AI?”

Hype creates artificial pressure. Every board wants to know “what’s our AI strategy.” Every pitch needs a slide on AI. This inverts product logic. You start with the solution and then hunt for a problem.

Rabois is direct: most AI use cases he sees don’t survive basic ROI analysis. Not because AI doesn’t work. Because it’s being applied where it shouldn’t be.

The three real categories of application

Let me break down what separates AI that works from AI that’s just cost disguised as innovation.

Category 1: Automation of repetitive tasks with high human cost

Works when: the task is well-defined, repetitive, and currently consumes significant time from qualified people.

Real example: contract review in legal. The task is clear, human cost is high, error is measurable. AI reduces cost and time in a verifiable way.

Category 2: Content generation at scale

Works when: you have genuine demand for content volume that would be impossible to produce manually, and the cost of error is low.

Real example: product descriptions in e-commerce with thousands of SKUs. The volume justifies it, quality is auditable, error isn’t catastrophic.

Category 3: Interface for complex data

Works when: you have valuable data that users can’t access intuitively, and natural language responses solve real friction.

Real example: conversational search across internal knowledge bases. The data exists, it’s scattered, users would spend hours digging.

AI that works

• Solves an existing, measurable problem
• Replaces verifiable human cost
• Error is detectable and correctable
• ROI calculable in under 6 months

AI that's cost in disguise

• Hunts for a problem to fit the solution
• Adds a layer without reducing real cost
• Error is diffuse and hard to measure
• ROI depends on optimistic assumptions

The trap of “copilot for everything”

Rabois makes an observation that should be obvious but isn’t: generic copilots are commodities. You don’t build competitive advantage with something any competitor can implement in weeks.

The wave of “AI assistants” in every type of product is a symptom of lazy thinking. It’s the default answer when you don’t know exactly how AI can solve something specific to your domain.

Real differentiation lives in domain-specific applications that competitors can’t easily copy. That requires deep problem understanding, not technology expertise.

Real cost vs. perceived cost

One of the most ignored truths: AI has ongoing operational cost. It’s not just development.

• Tokens cost money. At scale, serious money.
• Latency affects UX. Slow responses degrade experience.
• Hallucinations require oversight. Someone needs to audit outputs.
• Models change. What works today can break on the next version.

• Have you calculated the cost per use of your AI feature?
• Do you know the scaling limit before it becomes unviable?
• Is there a process for auditing outputs?
• Do you have a contingency plan if the model changes behavior?

Many products launching AI features today haven’t done these calculations. They’re subsidizing the cost with investment or margin from other areas. When they need to scale, they’ll discover the math doesn’t work.

The architectural decision nobody wants to make

AI at the core of your product is different from AI as an auxiliary feature. These are different architectural decisions with different long-term implications.

AI as auxiliary feature: you add a function that uses AI, but the product works without it. Risk is contained. If it fails, you remove it.

AI at the core: the product doesn’t exist without AI. Your value proposition depends on it. If AI fails, the product fails.

Rabois is skeptical about building AI at the core for most companies. Not because AI doesn’t work. Because dependency on external infrastructure (OpenAI, Anthropic, etc.) creates strategic vulnerability.

This doesn’t mean you shouldn’t use AI. It means you need to be honest about where your real advantage lives. If it’s not in the AI itself, don’t treat AI as core.

What separates good decisions from hype decisions

After observing hundreds of products at different stages, the pattern that separates solid decisions from hype-driven decisions is simple:

Solid decisions start with the problem, not the technology. The PM can explain the problem in one sentence, quantify the current cost, and demonstrate why AI is the best solution (not just a possible one).

Solid decisions have success criteria before development starts. What does “working” mean? Which metric needs to move? In how long? If you can’t answer this before you start, you’re experimenting with company money.

Solid decisions account for failure scenarios. What happens if AI doesn’t perform as expected? What’s the backup plan? How much will you have already invested? If the answer is “we didn’t think about it,” you’re gambling, not deciding.

The necessary counterpoint

All that said: AI is transformative. I’m not arguing against AI. I’m arguing against AI without discipline.

The products that will win in the next few years are the ones applying AI surgically, to specific problems, with clear ROI. Not the ones adding AI everywhere hoping something sticks.

Competitive advantage doesn’t come from using AI. It comes from using AI better than your competitors, on the right problem, the right way.

If you don’t yet have clarity on what problem you’re solving, AI will amplify your lack of clarity. If you have clarity, AI can accelerate the solution dramatically.

The question isn’t “should we use AI?” The question is “where specifically does AI solve a real problem in a way that justifies its cost and complexity?”

Most companies will spend the next two years discovering they didn’t have an answer to that question. The ones answering it now will be far ahead when the dust settles.