Most People are Building AI Products Wrong - Here's How to do it Right
We're all tasked with building AI products these days as PMs and product builders. But most people are getting it wrong. Today, we're going to break down how to build AI products the right way.
We're living in the era of compound AI.
Huh?
It’s the era where any product can be enhanced with AI.
But here's the problem: most products are bolting AI onto existing features. They're not reimagining their core experiences from the ground up.
I’m sure you can relate to this "AI theater" — impressive demos that generate buzz but deliver minimal value to users.
AI Assistants are Failing
Consider the wave of "AI assistants" flooding the market.
Many sit awkwardly in sidebars, waiting to be summoned for tasks they perform with varying degrees of competence.
They feel disconnected from the core workflow — more like party tricks than essential components of the product experience.
Here’s the dirty truth about them (that product leaders will only tell you behind closed doors): users try them once or twice before returning to their established habits.
Their Mistake is Your Opportunity
The bolted-on approach that the vast majority of your competition is doing represents a massive missed opportunity for you.
You see:
The most powerful AI experiences don't announce themselves with flashing "AI" badges.
They seamlessly integrate into the product, solving problems naturally.
They don't require users to learn new interaction patterns or develop new habits.
They simply make existing workflows faster, easier, and more powerful.
That’s what we’ll help you build today.
Why Study Attio's Approach to AI
I covered Attio last year. They’re taking market share from giants Salesforce, Hubspot, and Microsoft.
And the metrics speak for themselves:
200M+ customer records
2,000+ paying customers
$64M series A
What makes Attio particularly remarkable is that it's happening in one of the most competitive software categories in existence. CRM is as “red ocean” as it gets.
How has Attio done it?
For my money, one of the most important ways Attio has is by fundamentally rethinking what a CRM should be in the age of AI.
While the big, established CRMs are simply adding AI features to their existing products, Attio is building with AI at its core.
So I wanted to study this for you.
Today’s Post
Two weeks ago, we’ve covered how to build AI product strategy. Now, it’s time to see how that works in action with a case study. I’ve met with 4 of Attio’s team and 3 customers to bring you:
The power of invisible intelligence
3-part framework to build AI-native products
Beyond feature factories: building for the AI era
1. The Power of Invisible Intelligence
What makes an AI feature truly magical isn't raw technical capability — it's invisibility.
The most powerful AI implementations are the ones users barely notice.
They're so naturally integrated into the workflow that they feel like extensions of the user's own capabilities rather than separate tools.
Example 1
Consider this example from Attio (look in the top left)
When a user creates a new workflow automation, the system automatically generates a descriptive name and summary based on the workflow's function.
It’s a small thing in the top left. There's no flashing "AI" badge. No special interface to learn.
Just a small moment of delight when users realize they can skip a tedious step in the process.
Example 2
Another example: when someone responds to a meeting request with "I'm busy all week," Attio's system automatically adjusts the scheduling sequence without fanfare.
The technology disappears into the background, solving a problem so seamlessly that users might not even register it as an "AI feature."
This approach represents a fundamental shift in thinking about AI implementation.
Instead of asking "How can we add AI to this feature?" the question becomes "How can AI make this experience feel magical?"
→ The former leads to bolted-on capabilities.
→ The latter drives fundamental reimagination of core product experiences.
2. 3-part framework to build AI-native products
I’ve distilled down three core principles you should use in building your AI products:
Democratize Experimentation
Bake AI Into the Cake
Build Invisible Infrastructure
Let’s break it down.
Facet 1 - Democratize Experimentation
AI capabilities evolve so rapidly that traditional product development approaches can't keep pace.
The most innovative AI implementations don't come from roadmaps.
They emerge from unexpected places — casual experiments that reveal surprising potential.
At Attio, this experimental mindset isn't confined to an "AI team" tucked away in some corner.
It's distributed across the entire organization.
Making it easily accessible to anyone on the team is the core.
Engineers, designers, and even non-technical team members have access to enterprise-level AI accounts and the freedom to explore possibilities.
This creates a culture where people eagerly build experiments in their free time and bring them to the office on Monday mornings.
Everyone feels a lot of ownership. Several engineers mentioned that they enjoyed coding on the weekends, building experiments, then coming in on Monday to show people.
This approach is the opposite of companies where AI implementation is:
Siloed within specialized teams
Dictated from the top down
Treated as a technical challenge rather than a creative opportunity
Small Experiments → Big Insights
The most valuable AI features often start as casual experiments rather than carefully planned roadmap items.
Here's the dirty secret about AI development most companies won't tell you: your first five ideas will probably fail. But your sixth might transform your product.
Attio runs numerous experiments but only ships productionized, hardened features that just work.
The key is experimentation. It's easy to sit in a meeting and say “let's build this AI-native call recorder.” But it really starts with an idea of the problem you want to solve.
From Experiment to Product
When promising experiments emerge, they undergo a collaborative refinement process.
This isn't your standard "pass it to engineering" workflow.
It's messier. More creative. More human.
The team engages deeply with customers throughout this process. In the early discovery phase for their call intelligence feature, they observed how different roles looked for completely different information in the same calls.
This insight led them to create customizable templates that allow sales and customer success teams to get insights in real-time from AI based on their own context, during the conversation.
Built-In Feedback Loops
Extensive dogfooding of AI features provides immediate feedback on their effectiveness.
For instance, they were dogfooding their new call recording for two months.
When Call Intelligence initially produced 20-line summaries in standup meetings, the team recognized that 2-3 lines would be more valuable.
By the next standup, the feature had already been adjusted.
This rapid iteration cycle creates a virtuous loop where each improvement leads to deeper insights and further refinement.
Facet 2 - Bake AI Into the Cake
The distinction between "AI sprinkles" and "AI cake" is very important.
Don't think about delight as a last-minute layer of polish. It's something that's there from the beginning.
AI sprinkles are flashy, attention-grabbing features sitting on top of existing functionality.
They're the technological equivalent of "Look, we have AI too!"
They come with bright animations and bold claims but deliver minimal substantive value.
AI cake represents intelligence baked into the product's architecture and experience.
It's less about showcasing AI and more about solving real problems in ways that feel intuitive and seamless.
Attio's Call Intelligence exemplifies this approach (look at the right, live).
Rather than simply adding transcription (a common "sprinkle"), they've built a system generating real-time insights during calls.
They’ve dedicated a lot more space for the insights rather than the call recorder because users can rely on the insights.
The product recognizes that value lies not in raw technology but in the actionable information it surfaces.
Other examples of this "cake" approach include:
AI variables in email sequences that dynamically generate content based on context
Out-of-office detection that automatically adjusts sequence timing
Workflows that name themselves based on their function
AI attributes that extend the data model without requiring technical expertise
What makes these features magical isn't their technical sophistication but their subtlety.
They don't announce themselves with neon signs.
They simply work, removing friction in ways users might not even consciously register.
It's a lot less flashy than other AI features. You don't need bright bold colors and animations.
Start with Problems, Not Technology
Instead of asking "Where can we add AI?" Attio starts with specific user problems.
Then they consider whether AI offers the most elegant solution.
Their team focuses on building AI natively into Attio rather than tacking on AI features. They care less about the transcription technology itself and more about the valuable data it can bring into their CRM.
This problem-first approach means AI is deployed only where it genuinely adds value — not wherever it could technically fit.
Revitalize Core Experiences
The most powerful AI implementations transform existing workflows rather than creating entirely new ones.
By enhancing fundamental activities users already perform, Attio creates value that's immediately accessible.
No need to learn new behaviors or interaction patterns.
Instead of building random features, they focus on generating insights from existing data like call recordings. They always return to first principles.
This approach delivers features that integrate naturally into existing workflows rather than demanding new ones.
Facet 3 - Build Invisible Infrastructure
Behind these seamless experiences sits robust infrastructure that most users will never see or think about.
Attio has invested heavily in a flexible AI framework that abstracts away complexity.
Model selection, prompt engineering, optimization — all happen behind the scenes.
This infrastructure allows teams to focus on solving user problems rather than wrestling with technical details.
The system pulls from different models based on performance needs and cost considerations.
It switches seamlessly between options to find the optimal balance.
The whole bit where you call the LLM is not really a problem, and you don't think about it. So you can focus on the product and solving a customer's need.
This approach has become increasingly viable as model capabilities have advanced and costs have declined.
What was once a stark tradeoff between speed and cost has evolved into a rich space for optimization.
Six months ago, it was either take too long or cost too much money. Now with the state of the models, there's huge optionality. They can dial in little tweaks.
The Human Element in Evaluation
Despite sophisticated technical infrastructure, Attio maintains that the most valuable testing is qualitative and human-centered.
Evaluation frameworks are nice. Human judgment is essential.
For things like, 'Is this summary actually useful?' they learned has to be a human in the loop while you're building and iterating on it.
A technically "accurate" summary that misses the point is worse than a slightly less precise one that captures the essential information.
The qualitative testing is way more helpful.
Evals are really good at testing quantitative results out of these LLMs. But a lot of the features Attio builds are qualitative.
The final polish requires human evaluation and adjustment.
Learning Through False Starts
Not every AI experiment at Attio makes it to production, but even "failures" provide valuable insights.
Sometimes features work technically but can't meet cost or performance requirements.
Rather than abandoning these ideas entirely, the team adds them to a backlog to revisit as model capabilities evolve.
There's failures that look like successes but can't get out for cost or speed.
The tech is moving so quickly. You go 'it doesn't work,' then you bump the model and you're like 'it's so great—let's get this into production tomorrow.'
3. Beyond Feature Factories: Building for the AI Era
To build AI well isn’t about any single feature.
It's in the methodology that consistently produces magical experiences.
5 Key Lessons
Democratize experimentation across your entire organization, not just among technical teams
Start with user problems, not technology capabilities — ask "what friction exists?" before "where can we add AI?"
Build for invisibility where AI features feel like natural extensions of the product, not bolted-on extras
Create flexible infrastructure that abstracts away complexity and optimizes for both performance and cost
Prioritize human evaluation over purely quantitative metrics when measuring AI feature success
The most powerful AI experiences don't come from adding intelligence to existing features.
They come from fundamentally reimagining product experiences with AI at their core.
When technology anticipates needs and solves problems before users even fully articulate them — that's when AI truly delivers on its transformative potential.
The companies that will thrive in the AI era aren't those that bolt the most features onto their products.
They're those that rebuild from first principles with AI natively integrated into their foundations.
Attio's approach offers inspiration for this transformation.
One that focuses not on showcasing technology but on creating moments of delight where AI capabilities disappear into the background of seamless, magical experiences.
I hope you liked the case study! I’ve been experimenting with ways to follow-up deep dives. Feel free to hit reply to share your thoughts.
Up Next
We have some great newsletters coming on:
How to Ace The Presentation Round of PM Interviews
Advanced Tactics: Product Sense Interview
European Market Deep Dive
I’m looking forward to sharing them with you,
Aakash
Great read - as a user, I'm OVER having poor AI slapped onto every tool I use. It doesn't help and in a lot of cases, makes things worse. Why must we always be in a rush to be the first, when we all just want it done right??
Interesting writeup and insights. Well 👍