The burn
A mid-size advisory firm did what half its competitors were doing in the year of the AI gold rush: rolled out a general-purpose chatbot to "boost productivity". No workflow analysis, no governance, no definition of what good looked like. Six weeks in, a manager pasted half a client file into it to summarise. Eight weeks in, a partner caught a confidently invented case citation in a draft that was two reviews away from a client.
The rollout was quietly killed. And the partners drew the natural conclusion: AI is hype, and it's a data risk for a firm like ours. They weren't wrong about what happened. They were wrong about what caused it.
What changed
Eighteen months later the same firm came back at it — differently. Instead of a tool for everyone, one governed workflow for one bottleneck: first-draft engagement letters, a task eating six-plus senior hours a week. The workflow ran on the firm's own precedents, inside its own environment — no client data leaving the firm. Every draft landed in front of the responsible senior as a draft, never as an answer. The rule was written down: the AI proposes, the professional disposes.
The measure was one number — hours recovered at the seniors' billing rates. Within a month the workflow was live and the number was on a page a managing partner could defend. The same partners who banned the chatbot signed off the expansion to the second workflow.
The point
"AI is hype" and "AI works" are both true — they just describe different objects. An ungoverned tool dropped on a firm is hype, and the burn is deserved. A scoped, governed, measured workflow with partner judgement in the loop is just process engineering — the kind firms have always done, with a faster engine. The difference between the two isn't the model. It's the assurance around it.