Building Ari, note #1: less is more
Spent the last few days teaching Ari, our AI support agent, how to reply less. A few examples: don't reply when a teammate is explicitly tagged. Spot messages that aren't support requests: well-wishes or FYIs, the kind that are natural on Slack. And my favorite: it used to talk indefinitely with other companies' support bots. Two AIs, politely replying to each other forever. We built a detector to spot the loop and break it.
Notably, none of this needed a better model or advanced AI techniques. Some of it was plain deterministic code: if a teammate is tagged, don't even call the model. The rest was small, targeted classifiers, validated with evals.
And why teach it to reply less? Because a wrong answer delivered with full confidence erodes trust faster than a skipped one.
Most of the time, less is more: fewer replies, less code, more trust.