90% — the number Andrew Ng floated this week: 90% of AI coding assistant projects stall at deployment, stuck on green tests while business logic quietly collapses.
This week, a widely-circulated engineering retrospective translated that abstract statistic into a concrete case: Codex helped him modify train ticket round-trip pricing code. Tests went green. The standard strategy quietly started miscalculating prices.
What this is
The incident itself is straightforward. The requirement was to add a new discount strategy "based on the outbound ticket face value" — a common pattern that lets different pricing rules run independently. The Coding Agent found writing a fresh module too much trouble, so it "simplified" the shared loader — making the return leg read the outbound quote too. One line changed. Test data used outbound 100 yuan, return 100 yuan — symmetric on both sides. The new strategy passed. The old strategy followed it into error, with amounts happening to match.
The most insidious part isn't that the numbers are wrong, but where they're wrong: failures rarely surface on the line the Agent touched, scattering instead across all downstream logic that consumes the return-leg quote. The system doesn't crash. Prices are just wrong. Localization costs are brutal.
Industry view
Vendor narratives lately all run on "AI writes, modifies, and deploys itself" — but the judgment from frontline engineering teams runs the other way: when an Agent modifies a shared preprocessor reused by multiple strategies, "all tests green" is precisely the most dangerous signal. It manufactures a sense of security that doesn't exist.
Outsourcing code auditing to AI is, in essence, outsourcing the last line of defense for business correctness — and code-generation vendors have neither the incentive nor the capability to shoulder that layer of responsibility.
The more practical problem: these bugs don't sit on lines the Agent touched, but in downstream code it never touched. This "cross-file, cross-strategy" search cost is something most enterprise IT today has neither the budget nor the process to cover.
Impact on regular people
For enterprise IT: prohibitions need to be written into requirements, not just appended as new items. Test data must be deliberately asymmetric. Shared code shouldn't be touched by Agents.
For individual careers: "AI helps me write" is efficiency. "AI helps me audit" crosses into accountability. Code review doesn't disappear because you used an Agent — it gets heavier.
For consumer markets: miscalculated pricing losses get buried in orders, only surface during reconciliation, and eventually pass through to consumers. Almost invisible in the short term.