Industrial AI in a plant of 20 to 200 people
Industrial AI covers two different projects that share one name. One half reads signals from machines that report their own state, so it needs sensors, storage for the history and someone to keep both running. The other half reads the records a plant already keeps in its ERP, its mailbox and its drawing files. At 20 to 200 people, the second half is the one you can start on systems you have already bought.
What industrial AI means
IBM describes industrial AI as artificial intelligence applied to real-world industrial operations, manufacturing systems and physical infrastructure. It separates the term from enterprise AI because it touches physical operations under real-time and safety constraints. IFS describes it as AI applied in industrial settings such as manufacturing, energy, aerospace and construction. NVIDIA folds physical AI into the same term and names simulation, robotic fleets and digital twins among the things it covers.
Those definitions were written for a company that already runs instrumented machines and a historian to store what they report. They leave out the half of the term that needs no instrumentation at all. The same phrase covers software that reads the records a plant keeps anyway: past quotes, job times, invoices, purchase orders, price files and drawings.
Two owners read the same definition and come away with different numbers. One is pricing a hardware project with an integrator attached to it. The other is pricing software against data his company already owns and pays to store. The section below sorts the field along that line, so you can tell which project a vendor page describes before you ask what it costs. The suppliers sort along the same line, and a separate page groups the Canadian ones by what they do.
Physical AI is the narrower robotics term inside this one, and it has a page of its own.
Two kinds of industrial AI
Sort every use of the term by what it has to read before it can say anything.
Uses that need a sensor network
Predictive maintenance reads vibration, temperature and current from a machine over months. Vision inspection reads camera frames at a station under controlled lighting. In-process optimization reads a live line through a control layer. Robot and cobot cells read their own position and the part in front of them. A digital twin reads all of it and holds a model in step with the equipment.
Every one of them needs machines that report their own state, somewhere to keep that history, and a person who maintains both. That is a capital project, and it arrives with a hardware vendor and an integrator. That project is what most vendor pages on this term describe, which is why a reader arrives expecting robots.
ThriveAI does not build robots, robot cells, machine vision rigs, PLC or motion control, autonomous vehicles, or any other hardware. We work on the second half of the term, and the rest of this page is about that half.
Uses that run on records the plant already keeps
Quoting, cost per part, promise dates, purchasing, order entry and forecasting run on records that exist before anyone starts. The inputs sit in the ERP, the mailbox, the drawing vault, the shared drive and the spreadsheets. Nothing gets installed on a machine, and no line stops while the work happens.
The work is reading those systems and reconciling them, so that a part number means the same thing in each one. What comes out is a draft, and a named person approves it before it leaves the building.
| Use | What it reads | What it needs first | Who buys it at this size |
|---|---|---|---|
| Predictive maintenance | Machine signals over time | Sensors, a historian, someone to keep them running | A plant with a maintenance department |
| Vision inspection | Camera frames at the station | A rig, lighting, a labelled defect set | A plant running one high-volume part |
| Process optimization | Live line data | Instrumented equipment and a control layer | A continuous process rather than a job shop |
| Quoting and costing | Past quotes, job times, invoices | An ERP that exports and time booked per job | A shop where one person prices everything |
| Purchasing and order entry | Supplier PDFs, customer orders, price files | A mailbox and a catalogue somebody owns | A shop or a distributor retyping documents |
| Forecasting | Sales history and open orders | Sales codes somebody has reconciled | Anyone buying against a forecast today |
Read the table as two projects with one name. The top three rows start with a purchase order for hardware. The bottom three start with an export from a system the plant already pays for. Only one of them is cheap enough to test against your own numbers this quarter.
Which half of the term applies to you
Tell me what your ERP holds and where the quoting work queues up.
Start a conversationWhat a plant this size can run this year
Quoting is where the queue is visible from the front office. A request arrives as an email with a drawing attached, and one person prices it from memory of the last part like it. The 2025 Built to Sell report surveyed 200 mid-sized business-to-business manufacturers in the United States in July 2025. Quote speed is the lever it measures. Modern Machine Shop's Top Shops 2025 reports that its Top Shops quote in one day and record a quote-to-book ratio 19 percent higher. That number belongs to the survey that produced it.
Cost per part is the second desk, and it reads as a recipe rather than as a standard somebody set years ago. Material sits at the price last paid. Time comes from comparable jobs, and the outside operations and the overhead each carry their own line. Every line keeps a basis, so an estimator can see where a number came from and overrule it. A cost assembled that way survives a customer asking why the price moved.
Promise dates are the third. A Fabricators and Manufacturers Association benchmarking survey cited by Infor puts metal fabricators at 84 percent on-time delivery. The top quartile runs better than 90 percent, so roughly one order in six lands late. The measured span between release and ship for work of that shape already sits in the job records. A date can be read from it rather than guessed.
Purchasing and order entry are document work. Supplier quotes arrive as PDFs, customer orders arrive as attachments, and somebody retypes both into the ERP by hand. The hub page walks through what AI drafts at each desk in a plant in more detail than this page has room for.
At a machine shop, a quote request arrives in the mailbox with a drawing attached. The suggested price comes back anchored on what the shop charged for comparable parts. The dimensions come off the title block and the metal cost from recent invoices. It writes a draft estimate into the accounting system and a draft reply into the mailbox, and stops there. Its costed lines reconcile exactly against the shop's own cached figures 1,603 times, and the code grammar reads 5,261 of the 5,262 distinct sold codes. The end-to-end run-through has not happened yet.
What has to be true in your own systems
Five checks decide whether the record side of the term is available to you this year. Each one takes an afternoon to answer, and none of them needs a vendor in the room.
- The ERP can export. Look for a CSV or Excel export on the quote and job screens, or ask the reseller for a read-only ODBC user.
- Time is booked against the job rather than against the day. Pull last month's time tickets and check whether the job number is filled in on every row.
- The documents are readable as text rather than as photographs of paper. Try selecting the text in three scanned orders with the cursor, because what fails needs getting text out of a scanned order or drawing first.
- Somebody owns the answer when two systems disagree. Name the person who decides today when the ERP quantity and the shelf disagree.
- The data is allowed to leave the building, or the deployment is shaped so that it does not have to. Read the confidentiality clause in your largest customer contract, because the answer is often already written there.
A no on any of the first three moves the starting point rather than closing the door. The fourth check is the one owners underestimate, because reconciling two systems means telling somebody their number was wrong.
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 who signs the output
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.
A named person approves every quote, every message and every write-back. Nothing sends on its own. When the record cannot answer a question, the system says so and names what is missing. A silent zero gets quoted and then it gets built.
That concern outranks money for Canadian buyers. Statistics Canada reports that cybersecurity or privacy concerns are the barrier businesses most often name as limiting their use of AI. That barrier stands at 13.4 percent, ahead of cost at 10.6 percent.
There is no audited certification behind any of this and no legal conclusion: no SOC 2, and no compliance opinion about any statute. The control gets described in writing so your own counsel can decide what it means for you. ThriveAI builds one record of the business on a dedicated server for plants of this size.
The same term on the buying side
Vendor pages file supply chain under industrial AI, and at 20 to 200 people that function is two people and a spreadsheet. Demand sensing, multi-echelon inventory optimization and network design all assume a planning team, a planning system and a data feed from customers. None of that exists at this size, and buying the software does not create it.
What survives sits next to the desk. A supplier quote gets read and each line matched against the price last paid. An emailed order becomes a draft that somebody checks before it enters the ERP. Parts whose lead time moved since the last purchase order get flagged while there is still time to react. The supply chain page covers what survives when the whole function is two people.
Forecasting sits on the same side of the split. It reads sales history and open orders, and it depends on sales codes somebody has already reconciled. A code that means two things produces a number that means neither. What comes back is a quantity an owner reads before he commits to a purchase order, with the history behind it in view.
At a distributor, 39 GB across more than 14,000 files disagreed with each other about what the products were. That became one catalogue feeding four systems, with most values carrying a recorded origin. The verification queue is deployed and the client team has not started working its backlog through it.
Where to start
Pick the desk where one person is the bottleneck and where the record already exists in a system that exports. If both sides of the term qualify, the record side goes first, because it pays for itself out of systems the plant has already bought.
Size predicts less than the vendor pages imply. Statistics Canada reports 19.9 percent AI use among businesses with 1 to 4 employees, against 27.8 percent among businesses with 100 or more. Under eight points separate businesses with 1 to 4 employees from businesses with 100 or more.
A first build reads the plant's real data rather than a sample, because a sample produces a demo that fails on contact with the shop. The longer write-up covers what building one of these involves.
A forward-deployed engineer works on site and builds each module against the plant's real data rather than against a sample of it. The people who will approve the output watch it being built, and they say so while it is still wrong.