AI for manufacturing: 6 jobs, and which one to start with
AI lands at six desks inside a small manufacturer: quoting, costing, scheduling, purchasing, order entry and forecasting. Every one of them reads the same thing, one record of what the plant quoted, built, paid and shipped. Each desk below gets a section naming the job as it runs today and what a person still decides after AI drafts it. The record that feeds all six comes after them.
What AI does at each desk
AI for manufacturing at a 20 to 200 person shop is software that reads the plant's own records and drafts the work at six desks. Those desks are quoting, costing, scheduling, purchasing, order entry and forecasting, and a person approves every output. When no packaged product fits a desk, the module is custom manufacturing software built beside the ERP. The other answers on this query come from the plant floor: digital twins, predictive maintenance, machine vision, cobots and generative design. Those are real projects, and a manufacturer ten times this size buys them. What a robot or a machine that sees can actually do this year is separated from the rest in physical AI, and what a plant needs on record before it applies. A shop this size rarely has the staff to run them, and none of them touches the desks where margin is decided. The money in a custom shop is made between the request arriving and the invoice going out.
| Desk | What arrives today | What AI drafts | What the person decides |
|---|---|---|---|
| Quoting | An email with a drawing attached | A price anchored on comparable parts the shop has built | The number that goes out |
| Costing | A standard somebody set years ago | Cost per part as a recipe, line by line | Which line to trust when two disagree |
| Scheduling | A date made from optimism | The measured span for work of that shape | The date the shop commits to |
| Purchasing | Supplier quotes as PDFs | Each line matched to the price last paid | Whether the order goes |
| Order entry | A customer purchase order as an attachment | A draft ERP order with the mismatches called out | Accept the draft or correct it |
| Forecasting | One person's read on what is coming | What the sales history supports, and where it refuses | What to buy against it |
Quoting
The request arrives as an email with a drawing attached. Somebody opens it, recognizes the shape, and prices it from memory of the last part like it. That somebody is usually the owner or the one estimator, and every other request queues behind him. A plant that sells configured products quotes through a CPQ instead, as what is CPQ explains.
Top Shops 2025, reported by Modern Machine Shop, covers the top fifth of North American machining operations in the industry's annual benchmarking survey. Those shops turn a quote around inside a day, and book 19 percent more of what they quote. Most shops do not get near that. Across manufacturers and distributors more broadly, 71 percent say a quote takes at least a day to produce. That is the 2025 Built to Sell Report, 200 respondents at firms of 100 or more staff, fielded July 2025. The number is a floor rather than an average.
AI reads the drawing and the request. It then pulls the parts this shop has already quoted and already built that match on material, size, process and tolerance. A suggested price goes in front of the estimator with those comparables beside it. He changes the number or he sends it. The rule that sets that number is the whole argument, and it reads the shop's own won and lost quotes rather than a market rate.
A machine shop priced every quote from experience, with the drawings buried in an email thread. A plant shopping for software instead can compare the CPQ tools by what each one does. The suggested price now arrives anchored on what the shop charged for comparable parts. The dimensions come off the title block, and the metal cost comes from recent invoices. The system writes a draft estimate and a draft reply and stops there. The end-to-end run-through has not happened yet.
A plant that sells configured products quotes from rules instead, usually kept in an Excel builder, and the quote often needs a drawing. CPQ for manufacturing covers that case, and engineer to order covers the orders the rules do not reach. AI CAD covers what the new drawing tools produce.
Costing
The price quoted and the cost the job actually ran rarely meet in the same place. The quote lives in a spreadsheet or in the estimator's head. The cost lives in the ERP as a standard somebody set years ago, and in time records nobody reads back.
AI assembles cost per part as a recipe the shop can read. Material enters at the price actually paid on recent invoices for that grade and that size. Time comes from the shop's own job records for work of that shape. The ERP's standard sits beside it, so the difference between the system's number and the floor's number is visible.
A recipe beats a number, because any line can be opened to see where it came from. When a line has no basis, the recipe says so instead of filling the gap with a zero. That is cost per part as a recipe the shop can read.
Scheduling
Promise dates get made from optimism and from whoever is standing in the room. The date goes on the acknowledgement, and the floor finds out afterwards.
Metal fabricators run about 84 percent on-time delivery industry wide, against better than 90 percent in the top quartile. That is the Fabricators and Manufacturers Association benchmarking survey, as cited by Infor in August 2026. Roughly one order in six lands after the date somebody promised.
AI reads the span this shop measured between release and ship for work of that shape, and hands the desk a delay it can defend. The contract term for that customer gets checked against the suggestion, and a conflict is shown rather than resolved. Nothing auto-fills. The desk types the date it is willing to commit to.
Purchasing
Supplier quotes arrive as PDFs, and somebody retypes them into a purchase order. The price last paid for the same part sits in the ERP, and checking it takes longer than trusting the quote.
AI reads each supplier PDF, matches every line to the part it is for, and sets the quoted price against the price last paid. Lines that moved get flagged with the size of the move and the date of the last purchase. Lines it cannot match are named rather than dropped.
The buyer approves the order. Nothing reaches a supplier without a person putting their name on it.
Order entry
A customer purchase order arrives as an email attachment, in whatever layout that customer uses. Somebody types it into the ERP line by line, and a mistyped digit becomes a shipment.
AI reads the attachment, matches every line against the price file and the catalogue, and drafts the order for the desk. Mismatches are called out on the line they belong to. Examples are a price that disagrees with the contract, or a part number that resolves to two items. The desk accepts the draft or corrects it.
That is one emailed purchase order, followed from the mailbox to the ERP line.
Forecasting
What will be needed, and when, is held in one person's head. The sales history that would answer it sits in the ERP, split across codes nobody has reconciled.
AI reads the sales history and the open orders and says what the record supports. Where the history is thin, it refuses and names what is missing instead of extrapolating from a handful of orders. Stock levels are the same question asked earlier, and the answer is only as good as the sales history underneath it. This is the least proven of the six, and it is the one we would build last.
Which desk queues behind one person
Tell me the desk and what the plant runs today, and I will tell you what it would take to read it.
Start a conversationThe record underneath all six
None of the six runs on a chat window pointed at the ERP. Each one needs one record of the business, assembled out of the systems the plant already owns.
Five sources hold it: the ERP, the mailbox, the drawing vault, the shared drive and the spreadsheets. Those spreadsheets exist because the ERP never had a field for what they hold. All five get copied into one data warehouse. Nothing is replaced and nothing is migrated.
Then the same part number has to mean the same thing in all of them. A code in the ERP, a code on the drawing and a code in the price spreadsheet often differ. The difference is a suffix, a space or a revision letter. Until that reconciliation is done, a number assembled across two systems is a guess with a decimal point on it.
Every value keeps a line back to the record it came from. A disputed price opens to the invoice behind it, so the argument ends at a document. When the record cannot answer, the system says so and names what is missing. A silent zero is the failure that costs money, because a zero gets quoted.
A named person approves every quote, every message and every write-back. Nothing sends on its own.
At a Quebec manufacturer, the costing engine reconciles 1,603 figures exactly against the client's own cached numbers. It reads 5,261 of the 5,262 distinct sold codes in that sales history and names the one it cannot. Those are counts of work done, and not a claim about what the shop earned.
Once the record exists, a question can cross it. You can ask one question that spans the ERP, the mailbox and the drawing, and read the source named beside the answer.
Derik at Thrive was instrumental in taking incredibly messy data we inherited in a business we acquired and, through using AI, organized it in record time in a way that made it reviewable by our team for final review and approval. He proposed solutions that were effective to implement and everyone on our team was impressed and happy to work with him.
Rios-Karim Mercier, Belmont Capital.
Where the data sits, and where the model runs
Your data stays at rest on your own server in Canada, one dedicated machine per client. Inference runs one of two ways and you pick. It runs on that same server with an open-weight model, so nothing leaves the box. Or it runs through a frontier model under a written zero-data-retention control, with the model tier named in the contract. Zero retention is not available for every tier.
This is not an audited certification and it is not a legal conclusion. There is no SOC 2 here and no compliance opinion. The control is described so your counsel can decide. The same choice, and who signs the output that comes back, is worked through in industrial AI in a plant of 20 to 200 people.
The same pattern at a distributor
A distributor has the same problem with a shorter list of desks: order entry, purchasing and supplier documents. Orders arrive as email and PDF, one person carries the retyping, and that person's vacation is an operational risk.
Underneath it sits the same record problem in a different shape. The same product carries different specifications depending on which file you open. That could be the supplier's sheet, the ERP or the spreadsheet sales works from. Nobody owns which one is right, so quotes go out from whichever file was open.
A distributor held 39 GB across more than 14,000 files that disagreed with each other. One catalogue now feeds four systems, and most values carry a recorded origin. A verification queue is deployed for the values that still need a human decision. The client team has not started working its backlog through it.
Order entry is the desk that can go early, because the work is legible and the volume is daily. It reads the same way at a distributor, where buying and order entry usually sit with the same two people. That shape has its own page: AI for supply chain when the function is two people and a spreadsheet.
What has to be true before any of it works
Four conditions decide whether a plant is ready, and all four are cheap to check.
- The ERP can export. If nothing leaves it, nothing reads it, and no amount of model quality fixes that.
- Time is recorded against the job, not against the day. A day total is a payroll number. Cost per part needs the hours booked against the job.
- Somebody owns the answer when two systems disagree. Reconciliation is a decision, and a decision needs a name attached to it.
- The data is allowed to leave the building, or the deployment is shaped so that it does not have to. Both are workable, and the second takes longer to stand up.
Check those four before anyone writes code.
Which module to build first
Build the desk where one person is the bottleneck and where the record already exists. The second half is the one plants fail.
Quoting looks like the obvious first build, because the estimator is visibly the queue. Pricing from history needs costs the shop can defend, and those live in job records nobody has reconciled yet. That is why costing has to come first when the cost side is not built.
Costing is the less visible module, and it is what makes the quoting suggestion worth an estimator's attention. Once cost per part has a basis, the quote has an anchor.
Order entry is the exception. It needs the price file and the catalogue and little else, so it can go early while the cost side is still being assembled.
A forward deployed engineer works on site, inside the plant, and builds each module against the real data rather than a sample of it. Questions get answered in the room where the work happens, by the person writing the code.