The best AI system is not always the one that uses AI at every step.
When Calab.ai built a system to inspect photographs of utility meter installations, testing revealed an uncomfortable problem: the same photograph could produce different answers, even when the model sounded certain.
Muhammad Abdullah, Forward Deployed Engineer at Calab.ai, explores how that finding shaped the design: choosing predictable checks where possible, using AI where judgement is needed, and bringing in a person when the evidence is unclear.


