Manufacturers in Michigan get sold AI as a plant-wide program. The pitch covers the office and the floor in the same breath: quoting, scheduling, inspection, maintenance, customer email, document intake, and a dashboard that somehow ties it together. The slide looks complete. The first ninety days usually look like three half-started pilots and a team that has already decided the technology is noise.
The useful move is narrower. Pick shop floor AI or office AI first. Finish one battle. Then decide whether the second one is worth the next quarter.
Those two domains use similar models and very different operating realities. Treating them as one program is how good intentions turn into a thin layer of demos.
What office AI actually is in a manufacturer
Office AI in this context is the work that happens away from the spindle: quoting and estimating support, RFQ and drawing intake, change-order summarization, customer email drafts, AP and invoice matching, and the internal hunt through procedures and old job packets.
The success signal shows up in turnaround time, quote consistency, fewer lost documents, and fewer hours burned on administrative chase. The failure mode is also familiar. Someone pastes a customer drawing into a tool that should not have it. A draft email goes out with the wrong revision. An estimator accepts a price recommendation without opening the history that supposedly justified it.
Office AI is usually easier to start because the workflow is already on screens. The people involved already live in email, ERP screens, shared drives, and PDF viewers. You can pilot with a small group without stopping a cell. That is why vendors love leading with it.
A stamping supplier in West Michigan cut RFQ triage time by putting an assistant in front of inbound customer packets. The tool did not win the jobs. It sorted the packet, flagged missing pages, and drafted a checklist the estimator used before touching price. That is office AI doing honest work. It is also the kind of win that disappears if the same team is simultaneously being asked to retrain visual inspection and rewrite the scheduling board.
What shop floor AI actually is
Shop floor AI is closer to metal, people, and takt. Visual inspection assists, scheduling recommendations, quality photo review, scrap coding helpers, and tools that surface constraints a supervisor used to hold in their head.
The success signal is throughput, scrap, first-pass yield, overtime, and whether the board still makes sense at 10 a.m. when a machine goes down. The failure mode is sharper. A false accept ships. A schedule recommendation ignores a tooling change that only the lead knows. Operators stop trusting the screen and go back to the whiteboard while management keeps citing the pilot in meetings.
Shop floor AI is harder to start cleanly because the environment fights you. Lighting changes. Parts vary. Tribal knowledge is not in the ERP. The people who must use the tool are measured on output, not on patience with software. If the pilot adds friction on a Tuesday when a rush order from a Detroit OEM just moved up, the tool loses the room fast.
A machining operation north of Ann Arbor got real value from a first-pass vision check on a high-volume turned part. They kept humans on the ambiguous cases and measured escapes weekly. That project worked because it was the only AI project in flight. The same plant tried to bolt on a scheduling assistant three weeks later and both efforts slowed. The quality lead and the production scheduler were pulling from the same thin pool of process engineers to clean data and sit in vendor calls.
Why running both at once usually fails
The constraint is not GPU capacity. The constraint is attention from the few people who understand both the process and the systems.
In a mid-sized shop, that group is small: an operations manager, a lead estimator or scheduler, someone from quality, and whoever keeps the ERP from catching fire. Every AI pilot needs those people for requirements, exception handling, training, and the awkward week when the tool is wrong in public. Two pilots double the meetings and half the follow-through.
There is also a trust budget. Operators and office staff will give a new tool a short window. If the company announces a transformation and then ships two shaky assistants, the window closes for everything branded AI. People do not separate "office chatbot" from "inspection camera" in their heads as cleanly as the org chart does. They hear one story: we bought computers to tell us how to work, and the computers are uneven.
Vendors rarely help you sequence. Their incentive is a bigger footprint. Your incentive is one measurable improvement that survives contact with a real month of orders.
How to choose the first battle
Choose based on the bottleneck that is already hurting, not based on which demo looked cleaner.
If quotes are late, inconsistent across estimators, or winning work that loses money on the floor, start in the office with quoting support. Keep the human review heavy. Measure quote cycle time and margin after the job closes.
If the floor is the constraint, be specific about which floor problem. Visual inspection on a stable, high-volume part family is a different project from scheduling across a job shop with constant changeovers. Pick the one where success is visible inside a week or two of running, not a year of "learning the model."
A useful test question: if this pilot fails, can we shut it off on Friday without rewriting how the company takes orders or runs the board? Office pilots are often easier to pause. Floor pilots that get wired into release gates are not. That alone is a reason for some shops to prove the muscle in the office before they put AI near a ship decision.
Another test: who owns the outcome. If no named person will stand up in the Friday meeting and say whether the pilot helped, do not start it. "Innovation team" is not an owner in a 120-person manufacturer.
Regional context matters too. A Tier supplier living on OEM schedule changes may need office speed on change orders before they need another screen at the cell. A high-volume plant with stable parts and chronic inspection overtime may need the floor project first. The answer is local. The mistake is pretending the answer is "both, now."
What finishing looks like
Finishing a battle means more than a go-live photo.
For an office pilot, finishing means the tool is in the daily path for a defined queue, the exception cases are documented, the data handling rules are clear, and you have eight to twelve weeks of before-and-after numbers you would show a skeptical plant manager. For a floor pilot, finishing means operators use it without a babysitter on shift, false accepts and false rejects are tracked, and the override path is muscle memory.
Only then do you open the second front. The second project benefits from the first one in boring ways: you already know how your IT vendor will host it, how you will train people, how you will measure, and how you will kill a tool that is not earning its keep.
The manufacturers getting value from AI in this region are not the ones with the longest initiative list. They are the ones who can point to one workflow that got cheaper, faster, or less error-prone, and who still have the team's patience left for the next workflow. Shop floor AI and office AI can both pay. They pay more when they are queued like jobs on a finite machine, not launched like a rebrand.