Sit in enough meetings about AI and you start noticing the same argument on repeat. One person is asking whether the company should automate invoice processing. Another is asking whether it’s okay to draft a campaign brief with Claude. The meeting treats these as one topic, with one risk profile, needing one decision.
It isn’t one decision. It’s two, and they have almost nothing in common.
The first kind of AI use is a thinking partner. Someone opens Claude or ChatGPT and works through a task that is theirs: drafting a memo, pressure-testing a plan, summarizing a forty-page report, untangling a spreadsheet formula. The person reads the output, keeps what’s useful, and discards the rest. Nothing reaches anyone else without passing through their judgment first.
The second kind is infrastructure. A model wired into a process: routing support tickets, extracting fields from invoices, drafting the weekly report. It runs whether or not anyone is watching, and its output lands somewhere real (a customer inbox, a ledger, a decision) without a person necessarily reading it on the way.
The first kind is judged the way you judge advice. The second kind has to be judged the way you judge software.
Advice comes with a person attached
What makes the thinking-partner kind safe isn’t the model. It’s the person sitting between the model and the consequence. A bad output costs a few wasted minutes and gets deleted. A good one saves an afternoon. The quality control is built into the shape of the work.
That’s why this kind needs so little ceremony. It doesn’t need a roadmap or a platform decision. It needs a clear rule about what may be pasted into a chat and what may never be, accounts the company controls, and permission.
The expensive mistake here is delay. Fold it into the AI strategy, route it through a committee, revisit next quarter. Meanwhile the people who would benefit either use it quietly on personal accounts, which is the worst version of every risk involved, or don’t use it at all, which is a quieter kind of loss.
Software doesn’t get judgment for free
The infrastructure kind removes the person from the loop, and everything that made the first kind casual stops applying. The judgment a person was providing for free now has to be engineered in: which inputs qualify, what gets checked before the output counts, who owns the system when its behavior drifts.
That’s why this kind has a filter in front of it. A workflow earns automation when it has a well-defined input, a checkable output, and enough volume to evaluate against, and it earns an agent shape far more rarely than the buzz suggests. Done well, it takes weeks, involves real engineering, and ends up boring: an eval suite, monitoring, a runbook, an owner.
The question that sorts them
Treat the first kind like the second and you get the committee failure: months of policy work to govern what is, functionally, a better way to think. Treat the second kind like the first and you get the quiet failure: a model wired into a process with chat-level casualness, no eval, no owner, and nobody assigned to notice when its behavior shifts.
One question sorts almost every case: does a person review every output before it matters? If yes, you’re in the first kind, and the right move is to enable it this week. If no, you’re in the second kind, and the right move is to scope it like software.
Next time AI comes up in a leadership meeting, make everyone say which kind they mean. The first kind is a data rule and a permission slip. The second is an engineering project. Neither goes well while it’s mistaken for the other.