AI Native Products

AI-native products that do things for their user. They are proactive, always available, self-improving, self-driving and accessible via a single action. In the large model era, we are deploying units of intelligence instead of units of compute.

“AI Native” applications and patterns are still too early to commit to, but we know tons from previous ML/DL cycles:

  1. Data is moat — so build evaluation datasets.
  2. Collect user feedback on AI output (e.g., thumbs up/down, ala ChatGPT).
  3. Compute will improve over time — just like CNNs can now train and run on $9 ASICs. But in the interim, give users a sense of progress: text streaming (ala ChatGPT), partial renders (ala Midjourney).
  4. Route between varying levels of intelligence based on use cases; make it all look seamless to the user. (e.g., Facebook deploys 20+ RL/other models per user).
  5. The product should work across modalities (text, voice), devices (mobile, desktop, watch, AR), and form factors (app, web, call, SMS).
  6. Big Tech + OAI et al. will build user behavior around the textbox and voice. Align with this affordance. Chat isn’t just for completions — agents will run jobs (some complex) in the background. Users often stare at a blank text box. They need nudges to guide their initial interaction.
  7. Agents working in the background need to show progress and (maybe) let the user see under the hood. Early analog: batch/cron jobs.
  8. Anthropomorphizing intelligence usually makes for poor UX — users don’t care for “Alice the CMO AI”; they want job X done. Do the job well and with traceability.
  9. Value-based pricing is more straightforward when you’re selling intelligence, but it still aligns cost and pricing models. Progress to outcome-based when high reliability is achieved and the value proposition is clearly understood.
  10. AI-native means not just the product, but also internal workflows and tooling. Facebook built its own servers. AI-native companies should rethink their tooling by deploying one agent per user. This agent watches analytics, journey, and feature use, and runs campaigns, replacing generic, rule-based workflows on Iterable/Customer.io.