Twelve questions to answer before you spend a dollar on AI. Most AI projects fail on process and adoption, not on the model. This is the checklist I run on my own operations and on client floors. If you cannot check a box, that is the work to do first, and it is usually cheaper than the software.
Matthew Rochaine / industrial engineer, author of TIME. FLOW. PROFIT.
Phase 1: Before you buy anything
AI applied to an unmeasured process automates your guesses.
You have timed the process yourself, recentlyNot the SOP version. The one that happens on a normal Tuesday, measured with a stopwatch or a timestamp log.
You can name the constraintThe one step that sets the pace of everything. If three people give three answers, nobody knows yet.
You know what an hour at the constraint is worthIn dollars. This single number is how every AI proposal gets ranked, including the ones from vendors.
The obvious waste is already goneIf a step can simply be removed, remove it. Automating a wasteful step buys a faster wasteful step and a bill.
Phase 2: Picking the first project
The first project buys credibility for every one after it. Pick it to win.
The data already existsIn a system, a spreadsheet, or a folder of documents. A project that starts with a year of data collection is two projects, and the first one is boring.
The work being removed is retyping, reading, or routingThe strongest early wins take clerical load off people. Judgment stays human; the machine drafts, extracts, and flags.
A wrong answer is visible and cheapSomeone reviews the output before it touches a customer or a ledger. If an error would be silent or expensive, this is not the first project.
You can measure the payback in weeksHours saved, errors caught, or throughput gained, against the number from Phase 1. If it cannot be measured, it cannot be defended at budget time.
Phase 3: Making it stick
The tools are never the hard part. Adoption is.
One named person owns itNot a committee, not the vendor. A person on your payroll whose job now includes this system.
The people whose work changes were in the roomBefore the build, not after. The operator who was consulted will make it work; the one who was surprised will prove it does not.
There is a standard, written way to use itOne page. What goes in, what comes out, what to do when it looks wrong. If it only lives in someone's head, it retires when they do.
You can turn it offA rollback plan means the pilot is safe to try, which means people will actually try it.
The one rule
Never let the tool pick the problem. Find the constraint first, price it, and then ask whether AI is the cheapest way to relieve it. Sometimes it is. Sometimes the answer is a printed fixture, a layout change, or a second person on one station for two hours a day.
Matthew Rochaine helps small and mid-size operations put AI to work where it pays. Wellington, Florida.