Most conversations about AI in a company start from the technology: what can the models do now? That is the wrong end of the telescope. The systems that pay for themselves start from a process that was ready — and after eight systems in production, the ready ones all share the same five traits.
1. The work repeats at volume
AI systems have a fixed cost to build and a near-zero cost to run one more time. That economics only works when the same shaped task happens over and over. A parts dealer typing twenty minutes of listing data per part, hundreds of times a month. A back office with a document backlog nobody could clear. If the task happens twice a quarter, a checklist beats a model.
A useful threshold: if you multiplied the time one instance takes by how often it happens and the yearly number does not embarrass anyone, stop here.
2. The judgment is describable
"It takes judgment" is where most automation conversations die — usually prematurely. The test is not whether the task needs judgment, but whether your best person can explain the judgment. When an engineering team could write down how they review a building model, we turned those rules into checks that run the same way every time. The judgment stayed theirs; the applying of it became software.
What does not work: taste, negotiation, anything where the rule changes depending on who is asking. Keep people on those.
3. The inputs already arrive digitally
A photo, a PDF, an email, a form submission — if the raw material of the process already lands somewhere digital, the system can start working the moment it arrives. Shelf photos from a store visit become a compliance report because the photo was already the input.
If the process starts with a phone call nobody records or a conversation in a hallway, fix the capture first. That is a workflow project, not an AI project — and it is usually the cheaper of the two.
4. Errors are cheap to catch
Every system we ship keeps a human on the edge cases, because models are confidently wrong at unpredictable moments. That only works when a wrong output is visible and reversible before it costs money. A mispriced draft listing gets caught at review; a wrong figure wired to a supplier does not.
Rank your candidate processes by the cost of an uncaught error. Automate from the cheap end.
5. Someone owns the outcome
The quietest predictor of success. A process with an owner has someone who notices when quality drifts, feeds corrections back, and cares whether the number moves. A process nobody owns stays broken — with or without AI.
What to do with a process that scores five
Do not write a strategy document. Pick the narrowest slice of the process that is still useful, put a working version in real hands within weeks, and decide on evidence. That is how we run engagements — and it is also how we would tell you to run one without us.
