When to Replace expense reporting With an Agent — and When Not To
A field report on tool-first agents and what it changes for operators.
The interesting question is not whether ChatGPT Canvas works. It does. The interesting question is what teams do with it once the novelty wears off.
Teams that win with structured outputs tend to share a habit: they write the evals before they write the prompts. Everything else follows from that.
Inside Figma, the rollout looked less like a moonshot and more like a slow migration. A pilot, a champion, a quiet expansion, a budget line.
Snowflake has been quietly running incident response through Warp for months. The results are unglamorous and, for that reason, more interesting than another benchmark chart.
Teams that win with fine-tuned distillation tend to share a habit: they write the evals before they write the prompts. Everything else follows from that.
The cost curve matters here. Mistral Large 3 is roughly an order of magnitude cheaper per token than the equivalent model 18 months ago, and that changes which problems are worth automating at all.
Eval harnesses, once an afterthought, are becoming the most important piece of code in many AI projects. Databricks's team treats theirs the way an SRE team treats a runbook.
None of this guarantees a clean story. Anthropic could ship a model next month that rearranges the assumptions in this piece. But the direction of travel, for now, is clear enough to plan around.