Amazon's 6 Rules for Building With AI

Amazon's retail engineering organization ("Stores") published a short list of AI-native engineering tenets, and this module frames how to use them. The goal is not to admire Amazon's culture — it's to treat the tenets as a lens on tradeoffs every team shipping AI will face, while staying honest about where internal corporate policy and engineering judgment blur together.

Amazon's retail engineering organization ("Stores") published a short list of AI-native engineering tenets—internal rules for how teams should adopt AI. The list was reported by Business Insider from an internal document; Amazon confirmed the broad direction in a statement to the outlet. You do not need to admire Amazon's culture to treat the tenets as a useful lens: they are explicit tradeoffs—speed vs. cost, novelty vs. control, scale vs. niceties—that every team shipping AI will face.

This guide walks through all six tenets in an order that matches how most production systems break: auditability first, then execution speed, tool choice, upgrade discipline, ownership, and organizational scaling. Each section states the rule plainly, what Amazon says it is optimizing for, and what it means for your codebase or platform—not slogans, but decisions you can implement this week.

One caveat: internal policy at a trillion-dollar retailer is not scripture. The tenets bundle engineering judgment with organizational politics—central platforms need leverage, and "not all preferences are requirements" can double as cover for slow internal roadmaps. Read them as pressure-tested tradeoffs, then steal the mechanics that fit your size and risk profile.

The order matters: Amazon puts auditability ahead of raw performance, and so does this guide. This module covers Tenet 6 — the rule that every deployed AI system must be understandable and traceable, even at the cost of efficiency — because it's the foundation that makes every other tenet tractable.

• Structured decision logs for anything that picks a model, a prompt variant, a tool, or a policy path. Free-form chat logs are not enough; you need machine-queryable fields.
• Versioned prompts and configs checked into source control or a config service with history—so "what did we deploy Tuesday?" is answerable in minutes.
• Explicit failure and escalation paths when confidence is low—instead of silently averaging away uncertainty.

{
  "event": "model_route_decided",
  "request_id": "req_8f3a…",
  "chosen_model": "claude-sonnet-4",
  "reason": "input_complexity_score=0.82 > threshold_0.55",
  "candidates_evaluated": ["haiku", "sonnet"],
  "prompt_template_version": "support-agent@v37",
  "policy_flags": { "pii_redaction": true }
}

If you cannot produce that artifact during a five-alarm incident, you do not yet meet the spirit of this tenet.

With auditability as the foundation, the next three tenets govern how you actually ship: deliver working solutions before optimizing cost, reach for AI only when it's genuinely the best tool, and upgrade models deliberately rather than chasing every release.

The last two tenets are about people and platforms, not models. Central AI teams can't own every domain, and shared platforms can't honor every bespoke request. This module covers who must stay accountable (Tenet 4) and how to say no without losing customers (Tenet 5).

Tenets are only useful if you measure yourself against them. This module turns the six rules into a self-audit you can run on your own stack today, plus the original source if you want to go deeper.

Use this as a blunt self-audit. For each line, pick Strong / Partial / Missing.

• [ ] Decisions (routing, policy, tool choice) emit structured, queryable logs tied to request IDs.
• [ ] Prompts/configs are versioned and deploys are reproducible.
• [ ] On-call can trace an incident end-to-end without raw prompt dumps alone.

• [ ] You have shipped at least one production path end-to-end; cost tuning is scheduled, not blocking launch.
• [ ] Spend visibility exists—even if optimization lags.

• [ ] Non-LLM baselines exist for core workflows; LLMs justify their use with measurable lift.

• [ ] Model upgrades require an explicit evaluation gate and rollback plan.

• [ ] Domain owners co-own metrics, datasets, and acceptance criteria—not an AI center of excellence acting alone.

• [ ] Roadmap reflects scaled impact; bespoke asks are triaged and mostly routed into platforms or extensions.

If your stack scores mostly Strong on Tenet 6 and Partial elsewhere, you are directionally aligned with what Amazon encoded: trustworthy systems first, throughput second, clever models third.

• Eugene Kim, Amazon Creates 6 Rules for Engineers Building With AI, Business Insider, 2026. (Summarizes Amazon Stores engineering tenets from an internal document; includes Amazon spokesperson comment.)