One of the most lucid insights I’ve read recently comes from Adrian Levy’s article: “Perplexity and NotebookLM don’t use better AI — they use better intelligence flow architecture.” The core point is simple, yet often ignored: the success of products like Perplexity and NotebookLM doesn't depend on AI models that are "smarter" than the rest.
They use more or less the same technological ingredients available to everyone today: LLMs, RAG, retrieval systems, and prompts. Therefore, the real competitive advantage lies not in the model, but in the architecture that coordinates the work between human and artificial intelligence.
This is where what Levy calls Intelligence Flow Architecture comes into play: the design of how cognitive work is distributed and flowed between people and machines. In other words, the issue is not what AI is capable of in the abstract, but how the collaboration is organized. Who does what? Where does the intelligence live? When does the AI act autonomously, and when does it hand control back to the human?
These are design questions, not engineering ones. Yet, they are still surprisingly overlooked by product teams.
Perplexity and NotebookLM are excellent examples for understanding this. They start from similar technological foundations but generate radically different experiences because they adopt different architectural choices. It’s not just a matter of UI or "how the product looks": it’s a distinct philosophy on how humans and AI collaborate.
It is precisely from here that it’s worth deconstructing these cases, as they clearly show that today’s real leap in quality isn't in the AI itself, but in how we choose to use it.
The Difference Between an "AI Feature" and an "AI System"
Many products simply add a chatbot and call it innovation. Perplexity and NotebookLM do the opposite: they start with the most uncomfortable design question:
"Which part of the work should the AI do alone… and where should it stop?"
This is Intelligence Flow Architecture: you aren't just designing screens; you are designing:
Who decides
Who executes
When the "hand-off" occurs
How quality and risk are monitored
Levy demonstrates this well: in Perplexity, the value isn't "chat"; it’s the chain of search → source comparison → synthesis → citations already packaged within the flow. In NotebookLM, the value isn't "answering"; it’s transforming user-uploaded sources into a navigable knowledge base and generating insights on top of that.
The 4 Principles (And Why I Consider Them a UX Checklist)
Design Starting from Autonomy The mindset shift is: first define autonomous execution, then design the interface as a "dashboard" rather than an "engine." In successful cases, the user shouldn't have to micro-instruct; the system understands which work is repetitive/heavy and takes over.
Place Intelligence in a Precise Layer A brutally useful test: if I remove the AI, does the product still stand? If yes, it was often just an add-on. When AI is in the "right layer," the product cannot exist as an experience without it (it doesn't just "work worse").
Clear Human-AI Boundaries Reliable systems don't chase total autonomy; they design partial autonomy with legible hand-offs.
Optimized Cognitive Distribution It’s not about "delegating everything." It’s about delegating the right things:
Human: Goals, context, judgment, responsibility.
AI: Research, patterns, synthesis, scale, repetition.
Case Studies Beyond Perplexity/NotebookLM
1. Healthcare: Viz.ai and Time-Critical Diagnostics (Stroke)
The flow here is crystal clear:
The AI automatically analyzes CT scans as soon as they are available.
If it detects critical signs, it sends alerts to the team.
The doctor still reviews and decides.
UC Davis Health states it unambiguously: AI serves to prioritize, but clinical review remains human; time is a determining factor in outcomes. This is a perfect case of targeted autonomy (immediate triage) and strong boundaries (human clinical decision).
2. Education: Khanmigo as a Socratic Tutor
Education is full of risks (cheating, dependency, hallucinations). Khanmigo is interesting because it doesn’t play the game of "giving answers" but rather keeps the student in the cognitive loop through:
Monitorability: Teachers can access the full history of student conversations.
Socratic Style: It guides with questions and hints instead of delivering the solution. The AI is a "coach," not a "substitute."
3. Creativity: Adobe Firefly as a "Co-creator"
In the creative world, AI could do everything, but Adobe chooses to use it to free up time and multiply options, not to replace the creative's vision. A key takeaway from the "AI and the Creative Frontier" study: 90% of creators say generative AI tools help save time and money by offloading repetitive tasks. The value isn't "making the creative asset," but performing 20 useful iterations to reach the right choice faster.
4. Business/Enterprise: When AI is Truly a "Copilot"
In the enterprise world, it’s about governance. Microsoft Copilot Studio uses "gates": the process can be paused, the output is submitted to a human reviewer, and only after verification does the flow resume. This isn't a technical detail; it’s a design choice. The AI suggests and prepares the ground, but the final control—and responsibility—remains with the organization.
What I’ve Learned
The real lesson is this: with AI, my job is no longer just designing interfaces, but deciding how and when people and machines collaborate.
A good AI product doesn't work because "the AI is good." It works because someone decided who does what, when, and why. Today, when I look at an AI product, I ask four simple questions:
What does the AI actually do alone, without the user having to guide it step-by-step?
Is intelligence at the core of the product or just a surface-level addition?
When the AI finishes its work, how does it pass the ball to the human?
Are we using the capabilities of both well, or are we making people do work the machine could do better (or vice versa)?
If these points are clear, the experience is fluid. If not, even the most powerful AI becomes confusing, exhausting, or useless. Design today isn't about making AI "prettier"—it's about making it useful, understandable, and reliable.
