Aparna Chennapragada on AI Product at Microsoft

Aparna Chennapragada on AI Product at Microsoft

transcript microsoft ai-product agents nlx enterprise product-strategy lenny-podcast

Aparna Chennapragada on AI Product at Microsoft

Guest: Aparna Chennapragada — Chief Product Officer at Microsoft (AI product strategy for productivity tools and agents). Previously CPO at Robinhood; VP at Google (Google Lens, Search, Shopping, AR, AI assistant); board member of eBay and Capital One; engineering leader at Akamai.
Host: Lenny Rachitsky
Source: Lenny’s Podcast. Recorded ~2025.

Overview

Aparna Chennapragada covers the enterprise AI product challenge (dual use case: product quality + governance), her framework for “living one year in the future,” the three-part agent definition (autonomy, complexity, asynchronous), and her concept of NLX — natural language interface design — as the emerging design discipline replacing traditional UX for AI products.

Key ideas

  1. NLX is the new UX. Natural language interfaces are not undesigned — they have grammars, structures, and UI elements that are invisible but must be explicitly designed. Prompts are new UI constructs; editable plans are new constructs; showing work/progress is a new construct; follow-up suggestions are a new construct. This is a design discipline in its infancy.
  2. Three properties of agents. Autonomy (delegation spectrum — higher intelligence enables higher-order task delegation); complexity (multi-step, not one-shot); asynchronous (works when you are not working).
  3. Living one year in the future. Chennapragada’s operating model: actively imagine and prototype what work looks like one year ahead. At Microsoft, operationalised as the “Frontier program” — a cohort of early adopters who can access cutting-edge experimental features without requiring the whole enterprise to adopt simultaneously.
  4. Enterprise dual use case. Every enterprise feature has two jobs: (a) product quality / usability and (b) governance, security, compliance. In consumer, the default is “make it work and make it delightful.” In enterprise, you must also ask “who has access, who audits it, what happens if it fails.” The Van Damme splits metaphor: simultaneously making the future happen and managing the adoption/governance side.
  5. PM role is not going away. But you earn the “editing function” through domain expertise and taste, not through title. AI unlocks latent good ideas from engineers and designers; the PM’s job is judgment, prioritisation, and synthesis — not gatekeeping.

NLX: natural language experience design

The shift from GUI to NLX:

  • GUI (graphical user interface): explicit, rigid, must be designed because it is not natural. Dropdown menus, buttons, forms — all designed artifacts.
  • NLX (natural language interface): elastic, conversational, but still requires explicit design. The design elements are invisible.

New NLX design constructs:

  • Prompt — the new equivalent of a text field or search box. Designing what to ask, how to frame, what to fill in by default matters.
  • Editable plan — when you give a high-level goal, showing the plan and allowing the user to edit before execution is a new UI pattern for agents.
  • Showing work / progress — the “thinking aloud” pattern (visible in ChatGPT reasoning, Copilot, DeepSeek). Too verbose = feels like a cron job; too terse = no confidence. Calibration is a design decision.
  • Proactive follow-ups — suggesting next actions after a task completes. Too many = annoying; too few = missed opportunity to guide users into the happy path.

The enterprise AI challenge

B2B dual use case. Every feature simultaneously serves:

  1. Product quality — does it work, is it delightful, does it help the user?
  2. Governance — is it secure, auditable, compliant, safely permissioned?

Enterprise product builders who come from consumer often default to one of two errors: (a) ignoring governance (“we’ll make it work first”) or (b) over-restricting UX in favour of safety. The art is holding both.

Frontier program. Chennapragada’s approach: do not require the whole company to move at once. Create a cohort of early adopters who can access experimental features. This respects the reality that productivity habits change slowly while enabling rapid iteration with willing users.

Three properties of agents

Chennapragada’s framework for thinking about what distinguishes agents:

  1. Autonomy. A spectrum — from tightly supervised copilot to fully delegated goal pursuit. More intelligence → higher-order tasks can be delegated.
  2. Complexity. Not one-shot (“summarise this document”), but multi-step (“help me knock this meeting out of the park by researching the attendees and generating the right persuasion pitch”).
  3. Asynchronous. Works when you are not watching. The shift from synchronous chatbot interaction to background autonomous operation.

Human-agent co-working

Chennapragada’s second “Roman empire”: designing human-agent collaboration. Not replacement — co-working. The question is: what tasks can be delegated, what can be inspected, how does information flow between people with agents as mediators?

Current limitation: all agent experiences are single-player. The future she is building toward: shared spaces where humans and agents collaborate, with outputs greater than what any individual or small team could produce alone.

See also