How to Pick Your First AI Project (Most Companies Get This Wrong)
How to choose your first AI project in a manufacturing or operations business: four tests that predict success, three tempting project types to refuse, and why "clean the data first" is a trap.
Most companies that try AI and quit didn't fail because AI can't do the work. They failed because they picked the wrong first project.
The first project isn't really about the hours it saves. It's the proof. If it works, you get believers, and the second and third projects get easy. If it flops, "we tried AI and it didn't work" becomes company gospel for the next three years.
So here is a filter: four tests for a good first project, and three project types to refuse even when they sound great.
The four tests
It's repetitive. The same shaped work, over and over — quotes, checklists, validations, reports, lookups. If your people groan when it lands on their desk, that's the sound of a candidate.
It's document-driven. PDFs, spreadsheets, drawings, specs, emails. Reading things, pulling data out of them, checking that data against other things. This is the sweet spot of what AI agents do, and most operations are full of it.
Skilled people are doing unskilled parts of it. The expensive failure mode in most companies isn't people being lazy — it's your best engineer spending four hours re-keying part numbers out of a PDF. You want projects where the judgment stays human and the drudgery leaves.
You can measure it before and after. Hours per quote. Items cleared per week. Days of turnaround. If you can't put a number on the before, you will never prove the after — and provability is the entire point of a first project.
Notice what's not on the list: impressive, customer-facing, strategic. Your first AI project should be boring. Boring is where the money is.
The three refusals
| Refuse first | Why |
|---|---|
| Anything customer-facing | Internal mistakes get caught by your engineer; customer-facing mistakes get caught by customers. Take that risk deliberately, later. |
| The everything platform | One rollout that transforms quoting, scheduling, inventory and HR means eighteen months of licenses and change management with nothing provably better. Slow proof kills belief. |
| The data-perfection prerequisite | "First we need to clean up our data" sounds responsible and produces a two-year detour. Modern agents are good at messy data — reading mess is much of what they're for. |
One process, proven, beats ten processes promised. And on data: pick a contained process and let the AI meet your data where it is. You'll clean what matters as you go, because you'll finally know which data matters.
The shortlist for manufacturers
For most custom manufacturers, the four tests produce the same candidates:
- Quoting and estimating
- Order entry
- Compliance and quality checklists
- Validating bills of materials against prints
- Turning tribal knowledge into work instructions
Any of these can be the one — if you measured the before.
The part that matters more than the project
Whoever does this work today has to be in the room when it gets automated. Not informed afterward — in the room, building the checks, deciding where the human review stays. That's the difference between a tool people trust and a tool people quietly stop using.
Most AI failures aren't technology failures. They're trust failures.
Common questions
What makes a good first AI project?
Four tests. It is repetitive — the same shaped work arriving over and over. It is document-driven — PDFs, spreadsheets, drawings, emails. It has skilled people doing unskilled parts of it, like an engineer re-keying part numbers from a PDF. And it is measurable — a number exists for the "before," so the "after" can be proven.
Should our first AI project be a customer-facing chatbot?
No. When an internal tool makes a mistake, your team catches it and the system gets tuned. When a customer-facing tool makes a mistake, a customer catches it. Customer-facing AI is a risk to take deliberately later — never as the first move.
Do we need to clean up our data before starting with AI?
Usually not, and treating data cleanup as a prerequisite is a common trap that produces a multi-year detour and no results. Modern AI agents are good at messy data — inconsistent formats and mismatched documents are much of what they are for. Pick one contained process, let the AI meet your data where it is, and you will learn which data actually matters.