In a manufacturing SME, the artificial intelligence examples that survive in production almost never touch the machine. They touch the paperwork around it. A customer quote read and priced on past jobs. A purchase order extracted and dropped into the ERP. A technical manual queried in plain language by a technician. An after-sales email sorted and answered before the shop floor opens.
Predictive maintenance takes up all the space in articles on the subject. It needs sensors, a failure history and a budget most SMEs will not commit this year. Tandem works with more than 50 companies on process automation, manufacturers among them. The projects that pay back fastest start in the back office before they reach the production line.
Which AI examples actually run in a manufacturing SME?
Five flows carry most of the value at a manufacturer. They share one trait. Their raw material is a document that already arrives every day, by email, through a form or on a portal, and that a person reads then retypes somewhere else.
- Incoming quotes. The AI reads the request, matches it to the catalogue and proposes a first price based on past jobs.
- Orders and sales administration. It extracts the lines of a purchase order or an invoice and files them into the management system.
- Technical documentation. It finds the right procedure across thousands of drawings and manuals, quoting the source document.
- Support and after-sales. It answers repetitive requests and prepares the file for technical cases.
- Tender monitoring. It watches public portals, spots relevant notices and summarises them for the decision.
Why manufacturers start in the office before the shop floor
Manufacturing is catching up fast. In the Insee ICT survey published in July 2026, the share of manufacturing firms using at least one AI technology rose from 7% to 17% in a year. The French average sits at 18% across all sectors. Manufacturing therefore more than doubled in twelve months while staying below the average, which means early deployments still differentiate.
Size matters as much as sector. Insee measures 15% adoption among firms of 10 to 49 employees and 31% among those of 50 to 249. Meanwhile the France Num barometer published in September 2026 by the Direction générale des Entreprises, across more than 9,000 respondents, reports 40% of very small firms and SMEs using AI, and 53% among SMEs alone. Only 19% pay for a solution. Many industrial companies entered through a free everyday tool, with no structured project behind it.
Quotes, the first workstream in a manufacturing SME
Quoting is the flow manufacturing leaders mention most often, for a simple reason. It is slow, it needs an expert, and a late answer loses the deal. Pricing a custom part takes hours of engineering time before anyone knows whether the customer will buy.
EIST, a Lyon company in boiler-making, sheet metal and custom industrial equipment, cut that time from two hours to ten minutes. In its France Num case published in June 2026, its managing director describes a six to eight month project run in-house with their assistant, supported by the Auvergne-Rhône-Alpes region. The tool analyses the 3D files of the parts and predicts manufacturing phases, production times, material costs and the final price from the company's own history.
We deployed the same logic in reverse at a window manufacturer. In the FPV Industries case, every incoming PDF quote is read, matched to the catalogue, then turned into an integration file for the in-house ERP. Thirty seconds per quote, for around 15 cents of processing cost, against a grid of close to a thousand business rules. One design choice deserves emphasis: the workflow recalculates no price at all. Pricing stays inside the manufacturer's ERP, which keeps control of it. We detailed the mechanics in our guide to automating quote processing.
Orders and sales admin, read the document then fill the ERP
A manufacturing SME receives purchase orders, order confirmations, supplier invoices and delivery notes every day. Formats change from one customer to the next. A person reads, understands and retypes. This is where copy-paste is most visible and easiest to put a number on.
The stack we deploy on this case has been stable for two years. A trigger sits on the mailbox or the shared folder. A reading engine turns the document into usable text. A small model extracts the useful fields, then pushes them to the target system. I described this architecture in my newsletter issue on the classic enterprise use cases. Unit processing cost falls below 5 cents per document on a simple extraction.
Finding an answer inside the technical documentation
A manufacturer accumulates drawings, routings, safety sheets, quality procedures and manuals for machines bought twenty years ago. A technician hunting for the right revision of a drawing loses a considerable amount of time. Studies put the share of working time spent searching for information at around 30%.
The answer is an internal assistant connected to the shared drives, the management system and the mailbox, answering with the source document and its date. Respecting existing access rights conditions acceptance. Nobody should see through the assistant what they cannot see in the folder. Budget around 29 euros per user per month for a market solution, observed in October 2026, and a tenfold cut in search time on everyday questions.
Customer support and inbound requests at a manufacturer
A manufacturer selling to distributors or installers receives a lot of repetitive questions. Availability of a reference, lead time, compatible spare part, order status. Across our support engagements, roughly one request in three falls into that simple register.
The calculation belongs before deployment, and it belongs in cost per request. A support team handling more than a thousand requests a month lands at a unit cost in the tens of euros, fully loaded. A resolution handled by an AI agent costs tens of cents. The ratio is obvious, but it only holds for requests whose answer already exists somewhere. A product application question on a building site goes back to a technician, always.
Do you need an AI agent or just a workflow?
This is the question that wastes the most time in manufacturing SMEs. The answer is almost always the workflow. I wrote it in my post on agents in October 2025: nearly every use case deployed in companies is a workflow, and that is a good thing. An autonomous agent is non-deterministic, hard to test without an exhaustive dataset and hard to control. In a factory, those three traits raise a liability question before they raise a technical one.
| Level | What decides | Example in the factory |
|---|---|---|
| Simple automation | Written rules, no model at all | An acknowledgement sent on every order |
| AI workflow | Rules, plus a model on one precise step | Reading a purchase order before entry |
| AI agent | The model picks its own steps | Preparing a tender response file |
Agents keep their place when the task is genuinely open-ended and the output gets reviewed. Responding to a tender is one of those, and we described how it works in our article on AI tender agents. For everything else, a controlled workflow costs less, can be tested and can be fixed.
What a manufacturer should prepare before starting
The preparatory work is about data, and it lasts longer than setting up the tool. At EIST, the team had to clean the data coming out of the ERP, structure the bills of materials, link 3D files to production history and consolidate the manufacturing routings. The managing director sums it up in a sentence every manufacturer should pin to the wall: « Si la donnée n'est pas propre, l'outil ne sera pas fiable. » Clean data, reliable tool. Funding exists to absorb that phase, and we covered the Diag Data IA grant from Bpifrance and its conditions.
The second reflex is accepting a change of target along the way. Setic Pourtier, a Roanne manufacturer of cable-industry machines employing 85 people, started its project on predictive maintenance. Customer feedback was lukewarm. The company pivoted towards monitoring the quality of the cables produced, and that version found its market. Sophie Rondepierre, a department head, tells the story in a video testimony produced with the regional authority DREETS Auvergne-Rhône-Alpes. The most heavily marketed topic in the sector served that company poorly.
Which AI examples in manufacturing should you start with?
Start with the document that arrives most often and leaves as manual data entry. In nine manufacturing SMEs out of ten, that is the quote or the purchase order. Volume is measurable, the validation rule is obvious, and results show up in weeks rather than quarters.
Keep predictive maintenance for later, once the failure history is usable. Keep the autonomous agent for genuinely open-ended tasks. And treat data preparation as the real project, because it decides the outcome. Tandem builds these flows as custom automation and works regularly on document chains close to those of the building trade, described on our construction sector page.
