Essay
AI does not repair a broken process
A model can draft, classify and summarize. It cannot supply the missing owner, the missing metric or the missing stop.
Generative AI is genuinely useful on work the company already understands. It speeds drafting, pulls patterns out of a pile of tickets, and turns a messy interview into a first version of a procedure. Those are real gains. They are not a strategy.
The failure mode is using the model to skip the operating argument. If two departments disagree on the definition of a customer, a summary of their emails will sound fluent and still be wrong. Fluency is not agreement.
Three uses, three rules
Drafting can be loose, as long as a person publishes. Classification needs a labeled sample, an error rate the business will tolerate, and someone who reviews the misses. Decisions — money, a customer commitment, a production change — stay with a named human. The model may prepare the brief. It does not cast the vote.
Governance here is not a committee and a policy PDF. It is a one-page allowance: what the company will try this quarter, what it refuses to automate, and which number will tell it to continue or stop.
Companies that start there get adoption. Companies that start with a pile of licenses get a demo, a slide, and the same broken process running a little faster.