Search for AI in manufacturing and you will be shown predictive maintenance, computer vision inspection and demand forecasting. Those are real technologies and they work. They are also written for plants with hundreds of machines, a maintenance department and somebody whose whole job is data. If you run a workshop with thirty people, two of whom spend their mornings answering quote requests, that material answers a question you did not ask.
The narrower question is the useful one. Which jobs in a small manufacturing business can software take over safely, which should you buy ready made instead of commissioning, which should nobody automate at all, and what do European and Estonian rules require before any of it goes near a machine. This is the same shape we used for logistics companies and for accounting firms, because the answer keeps landing in the same place.
Where a small manufacturer actually leaks money
Scrap rate and machine downtime get watched. They are visible, they sit in the budget, and people argue about them in production meetings. The expensive leaks are the ones nobody counts, and in a small manufacturing business they are almost all on the office side of the wall.
- Quote requests that age. An enquiry arrives as an email with a drawing attached and a vague quantity. It sits until somebody has a clear hour. By the time the price goes out, the customer has ordered from a shop that answered the same day.
- The same order typed three times. Once into the quote, once into the production plan, once into the invoice. Every retype is a chance to transpose a number, and a transposed number becomes a remake.
- Paper that has to become data. Delivery notes, material certificates, shift logs, subcontractor dockets. Somebody types these up later, and the invoice waits for them.
- Customers ringing to ask where their order is. Every one of those calls is a status update that nobody sent, taking a person off something else.
None of that is a machine problem. The machines are usually the part of the business that already works. We wrote the general version of this argument in business process automation for companies: automate the handling, leave the deciding with people.
The jobs software takes over well
Reading enquiries into a form
A quote request is unstructured on arrival and structured by the time it is useful. Material, quantity, tolerance, finish, delivery date. Language models are good at pulling those fields out of an email and a spreadsheet into a form a person can check in fifteen seconds instead of transcribing for ten minutes. The person still prices it. The typing goes away.
Getting an order into the system once
The plumbing between your quoting, planning and accounting systems is dull work. It is also where most of the recoverable time sits. We build these on Make.com when the cloud is fine, and on self-hosted n8n when the data should not leave the building. In manufacturing it often should not, because the drawings belong to the customer. Neither of these is AI in the interesting sense. Both pay back faster than anything that is.
Turning paperwork into records
Delivery notes and certificates photographed on the floor can be read and filed against the job automatically. This is the same document flow problem a haulier has with consignment notes, and it behaves the same way: the win is not the reading, it is that the invoice stops waiting for somebody to find a quiet moment.
Telling customers where their order is
If a job slips, the customer finds out either from you on Tuesday or from their own goods-in on Friday. The first costs a message. The second costs the relationship. A status update that fires off the production plan without anybody remembering to send it removes a whole category of incoming call.
The phone
A voice agent answers on the first ring, identifies the caller, finds the order and either answers or takes a proper message into the system. We build these on Vapi, with speech from ElevenLabs and the line from Twilio. The thing worth testing before you buy is the language mix. An Estonian workshop takes calls in Estonian, runs a floor where Russian is common, and quotes Finnish and Swedish customers in English. Speech models handle small languages unevenly, so ask to hear a live call in each language you actually take rather than a recorded reel in English. We compare the options in our voice agent work.
What to buy rather than commission
Some of this problem space is solved, and solved better than anything you could commission. Buying the solved parts is not a compromise, it is the correct call.
- Vision inspection. Vendors have trained on volumes of defect imagery you will never assemble. Buy it.
- Predictive maintenance. If your machines already emit sensor data, the tooling to act on it exists. If they do not, adding sensors is a capital question, not an AI question.
- Production scheduling and MES. Mature market. A commissioned scheduler is a research project wearing a delivery date.
- ERP and quality management. Established, audited, and boring, which for the system of record is exactly the specification.
What is worth commissioning is the joinery: the phone line, the intake of enquiries, the document flow, and the links between systems that were never built to speak to each other. No vendor sells that, because it differs in every business. It is also why so many factory pilots stall, a pattern we pulled apart in why AI pilots fail to reach production.
What nobody should automate
Three things belong to people, and the reason is the same each time: somebody has to be accountable, and accountability does not survive being handed to a workflow.
- Machine safety functions. This has a legal conformity route of its own, covered below. It is not a place to be clever.
- Final quality release. Signing that a part is fit to ship is a statement to a customer and sometimes to a regulator. Software can gather every measurement that informs the decision. A person makes it.
- Material and supplier substitution. Swapping a grade because the usual one is out of stock changes what the customer contracted for. That is a phone call, not a rule.
The rules that apply before software touches a machine
Manufacturing carries a rulebook that office automation does not, and it is worth knowing which of your projects falls inside it.
Regulation (EU) 2023/1230 on machinery was adopted on 14 June 2023. It applies from 14 January 2027, and it repeals Directive 2006/42/EC from the same date. Its Annex I names a category that matters here: safety components with fully or partially self-evolving behaviour that use machine learning. Machinery with such systems built in is named too. Both need third-party conformity assessment.
The recitals then draw a clear line. Those rules do not apply to software that cannot learn or evolve and only runs fixed automated functions. In plain terms: a simple interlock is one thing. A learning safety function is another, and the second one brings a notified body with it.
Two frameworks sit alongside that. The EU AI Act sorts systems into four risk tiers. It also puts a transparency duty on systems that talk to people, which catches your voice agent even though it never touches a machine. The NIST AI Risk Management Framework gives you four functions: govern, map, measure and manage. It is the usual way to write this down so an auditor accepts it. Workplace safety duties run in parallel, and in Estonia the Labour Inspectorate supervises them.
The practical read is short. Office automation is light touch. Anything that talks to customers picks up a transparency duty. Anything inside a safety function is a conformity project with a timetable.
What Estonian manufacturers have that most do not
AIRE, the AI and Robotics Estonia digital innovation hub, exists specifically to help Estonian industrial companies adopt AI and robotics, and talking to them costs nothing before you talk to anybody who sells software. The state plumbing here also helps: company data, e-invoicing and tax filing are machine readable by default, which removes a class of integration pain manufacturers in larger markets pay consultants to solve.
Measure one normal month before you commission anything
The step that decides whether any of this works is the cheapest one, and it is the one most often skipped. Before you buy or build, count one ordinary month.
How many enquiries arrived. How long each waited before a price went out. How many orders were retyped, and how many remakes traced back to a transcription error. How many calls asked where an order was. How many invoices went out late because paperwork had not been entered. You do not need a system for this. You need a tally sheet and four weeks of somebody being honest.
That count is what turns "we should do something with AI" into a ranked list with euros beside each line, which is the only form in which this decision can be made properly. It is also the entire content of our AI audit, and the reason we give it away. If you want the document version of the same discipline, how to write an AI strategy covers the one-page format we use.
What it costs
Our audit is free, takes about 30 minutes, and ends with a written list of where your business is losing money, ranked by what each leak costs per month. If you decide to build afterwards, build-and-run starts from 600 euros per month plus VAT, and the number comes after the audit rather than before it. Published market rates for the wider consulting sector run 150 to 350 dollars an hour with scoped builds between 20,000 and 150,000 dollars, which we set out in full in what an AI consultant costs.
One point of honesty about our own evidence. The receipts we can show are from other kinds of operation, not from a shop floor. On a Tartu transport business we took structured data coverage from 4 items to 75, cut a hero image from 60KB to 14KB, published an llms.txt with a real AI crawler allowlist, and built the Make and n8n workflows and the Estonian-language voice agents behind them. That is website and machine-readability work and integration plumbing. It is not a claim about scrap rates, and we will not invent one. What transfers is the method, which the consultancy page describes.
Questions manufacturers ask about AI
What should a manufacturing company automate with AI first?
The office around the floor, not the floor. The jobs that pay back soonest are reading quote requests and drawings out of email into a structured form, getting order details into the system once instead of three times, turning delivery notes and certificates into records, and telling customers where their order is before they ring to ask. None of these touch a machine, so none of them carry machine safety risk.
Should a small manufacturer build its own AI quality inspection?
No. Vision inspection, predictive maintenance and production scheduling are mature markets with vendors who have solved them on far more data than you have. A commissioned version costs more and does less. What is worth commissioning is the connective tissue between systems that were never designed to speak to each other, which is the part no vendor sells you.
What should a manufacturing company never automate?
Machine safety functions, final quality release, and material or supplier substitution. Safety functions carry a legal conformity route of their own. Quality release is a signature that says a part is fit to ship, and a person owns that. Substituting a material changes what the customer contracted for, so it is a conversation, not a workflow.
Do EU rules apply if I put AI on a machine?
They can. Regulation (EU) 2023/1230 on machinery applies from 14 January 2027 and replaces Directive 2006/42/EC. It lists safety components with fully or partially self-evolving behaviour using machine learning as a category needing third-party conformity assessment. Software that cannot learn or evolve, and only runs fixed automated functions, sits outside that provision. Office automation that never touches a safety function is a different question again.
What does it cost to start with AI in a manufacturing business?
Our audit is free, takes about 30 minutes, and ends with your leaks ranked by monthly cost in writing. Build-and-run starts from 600 euros per month plus VAT, quoted after the audit rather than before it. Published market rates for the wider consulting sector run 150 to 350 dollars an hour, with scoped builds between 20,000 and 150,000 dollars.
The machines are the part of this business that already works. The inbox, the phone and the pile of dockets by the door are the parts quietly costing you orders, and those are the ones worth pointing software at first. If you want that ranked for your own operation, the shape of the engagement is on the automation service page and you can book the audit from there.