The AI conversation split into two this year, and most firms are still following the wrong half.

The AI conversation split into two this year, and most firms are still following the wrong half.

The loud half is about the models: which one reasons better, which one is fastest, which one to switch to next quarter. It changes every few weeks, and it barely touches how work actually gets done inside a firm.

The quiet half is about workflow design, and it’s the one actually reshaping operations. Redesigning a process so software can carry a task from start to finish — instead of a person doing it in pieces across five different tools — is a much harder problem than swapping a model. It means deciding what a human still needs to check, what can run unsupervised, and who is accountable when the software gets something wrong.

This is the problem LEA works on every day with RIAs. The firms treating document and data work as a model-selection problem are optimizing the wrong variable. The ones treating it as a workflow-redesign problem are the ones actually removing hours of work, because they rebuilt the process instead of bolting a chatbot onto the old one.

Model quality was the constraint two years ago. Workflow design is the constraint now.