AI-native, not bolted on
Three product categories — bolted-on, enhanced, native — and why the label determines architecture, margins, and what you must own.
Spin-out from AI-Native Product Building. Informed by Thinking.inc — AI-Native Build Guide 2026.
Claim
"AI-powered" on the landing page is not a category. Three levels matter:
| Level | If you remove AI | Architecture center |
|---|---|---|
| Bolted-on | Product still works | Monolith + API call |
| Enhanced | Experience degrades | ML beside app logic |
| Native | Product breaks | Model + data flywheel |
The label sets reversibility: bolted-on features are copyable with the same API key; native products bet on data loops, eval layers, and orchestration you cannot swap in a weekend.
The decision
Architect as AI-native when inference is the value loop — remove the model and nothing useful remains.
Architect as AI-enhanced when traditional software owns the loop and models improve a legible subsystem (recommendations, search, drafting).
Reject bolted-on as strategy — a summarize button is not a moat. Use it for learning, not for category positioning.
Five layers worth naming early
Even before you pick LangGraph vs custom glue, write which layers you own:
- Orchestration — prompts, tools, routing, validation (not one API call)
- Context — what the model sees, when, and why
- Data — retrieval, embeddings, freshness (existential for native)
- Evaluation — probabilistic quality, drift, human review — not only unit tests
- Agents (if used) — boundaries, escalation, forbidden side effects
If you cannot draw these on a whiteboard, you have implementation — not architecture.
Pressure-test
- Would the product still earn trust if the model API doubled price overnight?
- Is your "native" claim true if a screenshot demo hides the non-AI fallback path?
- Do you have an eval layer or only green CI?
- Are gross margins modeled honestly (native often lands lower than classic SaaS)?
- Can a second engineer change orchestration without reading chat history?
Related
- Handbook: Architecture decisions
- Handbook: Agentic development
- Handbook: Scaling
- Thinking.inc build guide — full practitioner stack