"Conversations have grammars, they have structures, they have UI elements, they're invisible. NLX is the new UX. That doesn't mean it's not designed."
Natural language interfaces are elastic, not rigid like GUIs
Every conversation has invisible structure: grammar, rules, UI elements
This requires intentional design, not just "chat with AI"
Prompts, plans, editable outputs are emerging NLX constructs
Framework
What Are Agents? Three Product Dimensions
Autonomy: Not just summarize a doc. Tell the agent: "Go analyze these competitive threats and brief me on the pitch."
Complexity: Build a prototype, write code, prepare presentations, multi-step tasks with loops
Natural interaction: Chat + voice + gesture + async. Works when you're not sitting at your laptop
The agent example"I told my research agent: go look at who's in my meeting, understand their views, come back with persuasion tactics for me." Not just saving time, giving me synapses I didn't have. Superpowers.
The Frontier
Building a Time-Travel Product Lab
The concept: Institutionalize living one year in the future. What would a company look like with all cutting-edge AI tools and agents at hand?
The experiment: Frontier team within Microsoft: real engineers building real products with the latest models, reasoning engines, agents
The unlock: How does a 3-person team with tons of compute and AI assistants change the work? What emerges?
The speed: Compressed tech cycles (weeks vs. years) require parallel rollout: both careful enterprise change management AND Frontier experimental features
Why Frontier matters
It's not gatekeeping early adopters. Enterprise needs both: long-term governance AND experimental programs where smart people can discover new ways of working before the whole company changes.
The duality
Van Damme doing the splits: one leg in rapid AI cycles (weeks), one leg in slow habit change (years). Both true at once, even inside Microsoft.
Prototype Your Taste
The New PM Skill: Taste-Making
If you're a process/TPS report PM, AI makes you less relevant, it can track meetings and emails
But the editing function becomes critical: supply of prototypes is 10x, so curation matters more
You don't get gatekeeper authority just from title, you have to earn it with judgment
The flip side: smart engineers now have experts in their pocket, unlocking ideas that would never surface
The taste-making bar risesWith AI democratizing building, the value of a PM shifts from "can you build it?" to "should we build it? Is this the right solution? How do we get adoption?" These questions matter more, not less.
Contrarian
The AI Adoption Myths
✗Models are getting matureINSTEAD →✓ The baby just grew up to a 15-year-old in a month. Your old priors about what AI can't do are outdated. Update them.
✗Product managers are obsoleteINSTEAD →✓ PM role changes, not disappears. Taste-making and editing become the core skill. The bar is higher, not lower.
✗Just let the model drive the productINSTEAD →✓ Natural language interfaces are still designed. Prompts, plans, editable outputs, these are new UI constructs. You have to build them.
✗Code is dead, coding is overINSTEAD →✓ We're just moving up the abstraction ladder. Software Operators (SOs) instead of Cs. Same computer science, higher-level thinking.