Inbox agent tests its errors

An inbox agent can look accurate simply because it’s checking its work against labels it created itself. A preprint about Foyer, a self-hosted personal inbox agent, argues that this makes it hard to know how often an agent is actually wrong. Its proposed fix is to measure performance against labels supplied by a human before training the systems being compared.

The study used 739 Italian and English emails labelled by the owner, then compared nine systems, including keyword rules, a traditional machine-learning model and generative language models. The paper reports that the existing rules scored 28% on message type, below the 58% majority-class baseline. With around 200 labels, logistic regression had the highest accuracy on each field, though it missed the two rarest classes. The authors also found that accuracy alone didn’t capture the practical risk: the most accurate system most often ignored emails that needed action. The paper is a single-user case study, so its results shouldn’t be treated as a general verdict on inbox agents.

For an operator considering AI for email or other routine decisions, the useful lesson is to set aside examples a person has checked, then test the system against those before relying on its own classifications. That kind of evaluation discipline is worth considering before putting AI to work on routine decisions. Read the preprint.