The price of an AI inference call has dropped by an order of magnitude in two years, and is still falling. This makes "smart" automation — workflows that previously needed human judgment, now handled by a model — economically obvious. Why pay a person to triage tickets when the model does it for 1/100th the cost? Why have humans summarize meetings when transcripts can be auto-summarized? The math, on the surface, is a slam dunk.
The hidden cost isn't tokens. It's the corner cases. The 5% of inputs the model gets wrong, the unusual cases that don't fit the pattern, the edge inputs that the model hallucinates a confident wrong answer for. A human, in the original workflow, handled all of these with their judgment. The smart automation handles 95% of them well and 5% of them confidently wrong, and somebody downstream has to catch the 5%.
This is fine — even great — if you've designed for the failure cases. The good automation patterns we see have explicit human review steps inserted at the moments where confident wrong answers are most expensive. Ticket triage: the model assigns priority, but the actual response is drafted by a person until the model's accuracy on response drafts hits a threshold. Meeting summaries: auto-generated, but the person who hosted the meeting reviews and edits before sending. The model accelerates the work; it doesn't replace the judgment.
Where this breaks down is the workflows that ship the model's output directly to a customer or counterparty without review. The customer service auto-reply that gives the wrong answer. The auto-generated contract clause that's subtly wrong. The auto-drafted email that goes to the CEO of a customer with a hallucinated commitment. The cost of catching these isn't in the token bill; it's in the trust the business has built and that the automation, by failing in a high-confidence way, quietly erodes. Smart automation isn't free. It's expensive in a different currency.