Architecture
AI-Native Product Architecture
Designing a system around a component that can be wrong
A product with a probabilistic component needs bounded uncertainty, reversible actions, visible confidence, honest escalation and a per-interaction economic model, and the interaction shape decides all of them.
Most AI features fail for reasons that have nothing to do with the model. They fail because nobody decided whether the system suggests, drafts or acts, and so nobody designed the undo, the abstention, the escalation or the cost ceiling that each of those shapes implies. This book makes the interaction shape the first decision and derives the rest from it: reversibility, confidence display, routing, unit economics, latency budgets, caching that cannot serve a wrong answer faster, and the evidence that tells you to cut the feature.
What it makes operable
- 01Choose between suggest, draft and act before designing anything else
- 02Design reversibility, abstention and escalation into the product surface
- 03Model the unit economics of one interaction, including its failures
- 04Decide on evidence when to cut the feature
Contents
12 of 12 published- 01A Component That Can Be WrongWhat a probabilistic dependency does to everything downstream of it.11 min
- 02Deciding Where AI Belongs and Where It Does NotA test for whether the model is load-bearing or decorative.12 min
- 03Suggest, Draft, or Act: Choosing the Interaction ShapeThe one decision that fixes reversibility, latency and cost for everything after.11 min
- 04Designing Reversibility Into the ProductUndo, preview and staging as product features rather than engineering afterthoughts.12 min
- 05Confidence, Abstention, and EscalationLetting the product decline, and making a decline useful to whoever reads it.13 min
- 06Routing: Capability, Cost, FallbackA routing table derived from per-step eval results rather than from instinct.13 min
- 07The Unit Economics of a FeatureCost per successful interaction, counting retries, review and failure.13 min
- 08Latency Budgets and Perceived SpeedStreaming, optimistic state, and the gap between fast and feeling fast.12 min
- 09Caching Without Serving the Wrong Answer FasterCache identity that includes the principal, the policy and the corpus revision.13 min
- 10Data Flywheels That Do Not Poison ThemselvesFeedback loops that improve the system instead of laundering its errors.13 min
- 11Failure UX: What the User Sees When It Is WrongDesigning the wrong answer, because you are going to ship one.13 min
- 12Deciding to Cut the FeatureThe evidence and the threshold for stopping without a political fight.13 min
Evidence
3 sourcesThird-party sources behind the book's premise. Every figure in them belongs to the party that published it and is attributed to them in the text.
- AI Project Failure Statistics 2026: The Complete Picture
Pertama Partners · 2026-02-08 · Industry report · reported
- LLM inference prices have fallen rapidly but unequally across tasks
Epoch AI · 2025-03-12 · Industry report · reported
- Small Language Models are the Future of Agentic AI
arXiv · 2025-06-02 · Research paper · reported
Continue through the library
