Designing around what AI can't do
Hallucination, cost, and compliance aren't edge cases to patch later. They're constraints you design for from the start.
Lane Steiner
Engineering Manager
The honest list
Generative models hallucinate. They get expensive at scale. They carry compliance implications that legal will, rightly, want to talk about. None of this is a reason not to use them. It's a reason to design around them on purpose instead of discovering them in production.
Constraints, not surprises
We treat each limitation as a design input. Where a wrong answer is costly, we add a verification step or a human in the loop. Where cost is the risk, we cache, route to smaller models, and set budgets. Where data sensitivity is the risk, we decide up front what never leaves the boundary. The result is a system that fails safely, not one that fails surprisingly.
Designing around the limits is what lets you be aggressive everywhere else. Once the failure modes are contained, you can let the model do the work it's genuinely good at — and trust the system around it to catch the rest.
Related Articles
More notes from the team — on how we build, what we measure, and the calls we make along the way.
