AI adoption in manufacturing works best in the back office, one process at a time. A manufacturer scans where the hours go, lets an AI agent prepare the work while people check it, turns every correction into a rule and lets steps run on their own once they have earned it. The second process then starts from what the first one learned.
At Lleverage we think nobody should let AI run their back office on day one. Autonomy is earned, one step at a time. This guide follows the path we use with manufacturers, wholesalers and distributors: 5 steps, 3 questions to ask before the first meeting and the first 90 days. We build AI agents for the whole back office inside the ERP a company already runs, and a demo shows one working on your own documents.
Why start AI in the back office now?
Four shifts point to the same place to start.
- Hiring will not close the gap. In Koninklijke Metaalunie's second-quarter 2026 barometer, labour shortage is the obstacle Dutch metal companies name most, at 46%. The UK's Office for National Statistics counts about 48,000 open vacancies in manufacturing for July to September 2026. That makes the next 150 orders a week a process to automate.
- At this size, AI is already normal. Statistics Netherlands reported that 66% of Dutch companies with 250 or more employees used AI in 2025, up from 42% in 2023 (the 2025 figure is provisional). Among industrial companies of all sizes the share rose from 7% to 16% over the same two years. The open question is where AI pays off first in the back office.
- People want to keep control. In a survey of manufacturers by Rootstock, reported on 12 November 2026, 73% prefer either human-directed AI or AI that works on its own within defined limits and escalates when needed, and only 13% prefer fully autonomous AI. Rootstock sells manufacturing ERP, so read it as vendor research. That is why the path below starts with a person checking every case.
- The invoice itself is changing. Belgium has required structured e-invoices between companies since 1 January 2026. The EU adopted its VAT in the Digital Age package on 11 March 2025, and from 1 July 2030 cross-border B2B transactions between member states fall under digital reporting requirements based on e-invoicing. Fewer invoices will need reading. The matching, the coding and the exceptions will remain.
All four point to the same start: where people lose hours retyping, with the check they ask for built in.
What are the 5 steps to an autonomous back office?
Order intake, supplier invoices, quotes and customer questions all follow the same path. Trust is built first, and autonomy follows from it.
| Step | What happens | What the company decides |
|---|---|---|
| 1. Scan | The work is mapped to find where the hours go | The first process, its owner and how the ERP can be reached |
| 2. Supervise | The agent does the work and a person checks every case | What the review screen shows |
| 3. Learn | Every correction becomes a rule the company owns | Which corrections become rules |
| 4. Automate | Steps that have proven themselves run on their own and only exceptions reach a person | Which steps have earned it |
| 5. Compound | The next process starts from what the first one learned | What the second process is |
How do you scan the back office and pick the first process?
A scan maps which processes take the hours, who owns them and how the ERP can be reached. It starts with the inboxes, the ERP and the people who do the work. For every process it records how many documents arrive, how long each one takes, who owns the result and how often the rules change. The output is a short list of processes, ranked by the hours they take and by how alike their cases are.
The right first process has 3 things: volume, an owner who feels the pain and a majority of cases that look alike. In manufacturing that is most often one of four.
- Customer order intake: orders arrive by email and PDF in the customer's own article names. The largest accounts may use EDI, so the work sits in the long tail: the PDF, the Excel sheet and the email that says "same as last time".
- Supplier invoices: invoices arrive before the goods, one invoice cites three purchase orders and approval limits are set on the total instead of the variance. E-invoicing will remove the reading, and the matching stays.
- Customer questions: in many service teams most questions are variations of the same few, and the answers sit in manuals, order history and the heads of two senior people.
- Quotes and specifications: only a few senior people know the pricing logic, so this process usually comes later, as an agent next to the engineer.
The order management agent handles the first and the last of these. Our guide to AI order processing for manufacturers covers the order case in depth.
What 3 questions should you ask before the first meeting?
Three answers decide the start date more than anything else.
- Can the ERP be reached? Find out whether there is an API and who controls it: your own IT, the reseller or the ERP vendor. Where there is no API at all, the agent works the screen with the access a person has.
- Does it run on your own servers? An ERP on your own servers needs a service account and network access, and sometimes an extra module. Lleverage reaches on-premise systems without opening them to the internet. Find this out in the first week.
- What does it cost to let a system create documents? Ask your ERP vendor whether it charges for documents created by an external system, or licenses a non-human user separately. Ask before you write the business case, not after.
A difficult answer tells you what to plan for. Microtechniek, a precision engineering company with 130 people, ran Ridder IQ on its own servers without the API module the cloud version has. In December 2025 a Ridder specialist installed the module on site, and the ERP became reachable. In August 2026, between 45 and 71 purchase orders a week were written into the ERP without anyone typing them, which gives the admin desk more than 500 hours a year back.
Two people belong in the room from the start: one who feels the pain every day and one who can answer the API question. Projects led by IT alone tend to stall. If your company has a works council, inform it during the scan, because a system that changes how work is done may need its advice.
What does supervision look like?
In the supervise step the agent does the work and shows how it got there. A document arrives by email, as a PDF, an Excel sheet or plain text. The agent reads it, matches it against the customers, articles and prices in the ERP and prepares the result with a confidence score for every field. A person sees the prepared result next to the original document and confirms, corrects or rejects it.
The confidence score is the agent's own measure of how sure it is of each piece of work. During supervision a person still checks every case, and the score shows where to look first. Later your company sets a line: below it a person decides, above it the work goes through on its own. Only the customer moves the line.
Four kinds of order line matter most in practice:
- a unit the ERP does not sell, such as boxes ordered where the ERP sells pieces
- a repeat order with no article number
- a price that is not the contract price
- a delivery date that cannot be met, such as a Sunday
Supervision also writes down the rules that today live in one or two heads: which of a customer's three accounts an order goes to, which price comes first (the framework contract, then the price list, then the order history), how units convert, and who handles partial deliveries and unknown articles.
Xpol, a wholesaler of fresh flowers, wrote down 25 customer rulesets in working sessions with its specialists, before one of them retired. Its agent prepares a draft in Business Central that a person signs off with one click, and it leaves ambiguous fields blank instead of guessing. Every blank field points at a framework contract or a customer code that needs cleaning up.
"It's a matter of building trust in the organisation with these kinds of initiatives. You can't just throw something like this over the fence." Cees Maaskant, General Manager, Xpol
How does a step earn autonomy?
A process is not one switch. Order intake is five steps, and in an example setup they sit at four different levels.
| Step | Example level |
|---|---|
| Receive and read the order | Runs alone |
| Match customer and article | Supervised |
| Check commercial terms | Assisted |
| Approve an exception | By hand |
| Create the order in the ERP | Runs alone |
A person moves a step up on evidence: the success rate, the corrections and what a mistake would cost. One setting moves it back. Some steps we recommend keeping with a person for good: changed bank details, master data changes and orders above a limit. Matching and approving should never be the same hand, so name the owner of the exception queue before go-live.
SPL Treatments, an aerospace surface engineering company, shows the learning loop. Its agent reads a purchase order, follows the specifications it refers to, checks them against a library of more than 800 standards and drafts the route sheet, with every recommendation cited to the section it came from. A first-time part took about an hour and now takes 2 to 2.5 minutes. The technical lead rates every result with a thumbs up or down, and what he marks down is what the agent learns from next.
A process that runs on its own still needs an owner. Formats change when a supplier sends a new invoice layout or a customer a new order template. Watch accuracy and corrections per step, and when they drop, move the step back to supervised until the new format is learned. Ask every vendor what happens to open cases when the AI model cannot be reached. They should wait for a person and never be created unchecked.
Which process comes second, and when should you not start yet?
The second process takes far less effort than the first. The ERP connection, the master data clean-up and the written rules are already in place, and the company brain already knows the customers, articles, suppliers and exceptions. It pays to know the second process before the first one starts. Supplier invoices are often followed by the supplier confirmations that explain their variances. Order intake is often followed by quotes for the same customers.
Place your own processes on two axes: how many documents arrive each month, and how clear the rules are.
| Rules unclear | Rules clear | |
|---|---|---|
| High volume | Customer order intake: start here. It stays supervised for longer | Supplier invoices with a goods receipt: start here. If finance has an owner, it reaches autonomy fastest |
| Low volume | Quotes and make-or-buy: later. A supervised agent works next to the person who knows | Confirmations against the purchase order: next. The rules are clear and it reuses the supplier master |
Four situations call for waiting:
- Fewer than about 500 documents a month. Below that, the first agent rarely pays back the ERP connection, unless the process is planning or quoting.
- Nobody owns the process. A project without an owner stalls in supervision.
- An ERP migration in the coming months. Plan the agent on the new ERP.
- Master data in the middle of a big clean-up. Finish it first, or let the agent's exception list drive it.
What do the first 90 days look like?
Our playbook plans a first process in 90 days, on the condition that the ERP can be reached. It has 3 phases, and each one states what it asks of your team. The first process typically runs supervised within a few weeks, built by our forward deployed engineers with your team. SIG Benelux, a building materials distributor, ran its first process as a four-week Kickstart with our partner SPAIK on one mailbox. 73% of its order emails now arrive in the ERP ready to review, and 88% of lines match the right article first time.
| Phase | What happens | From your team |
|---|---|---|
| Days 1 to 30: scan | One process is chosen, the 3 questions are answered and the review screen is agreed | A process owner for about 2 hours a week, a technical lead and a test environment |
| Days 31 to 60: supervise | The agent runs while a person watches, and every correction becomes a rule | The process owner, for about 30 minutes a day |
| Days 61 to 90: learn and automate | Manual hours are measured before and after, and the first steps earn autonomy | A decision per step on which ones move up |
What does AI adoption in manufacturing return?
Take an example desk: 150 orders a week at 15 minutes each is 37.5 hours of typing and checking. If 80% of the orders go through untouched, 30 hours come back, about 0.8 FTE. The 7.5 hours that stay are the exceptions: the odd price, the unknown article, the customer on the phone. Redo the sum with your own numbers.
Customers show the same pattern at their own scale:
- Topa Bathroom Products processes over 90% of incoming orders into Business Central with no manual input. Nearly 4 FTE (3.8) moved to after-sales and service planning, and customers get an order confirmation within 30 seconds.
- Xpol saves about 20 minutes per large order across 150 orders a week, and has avoided 1 to 2 planned hires.
- Microtechniek gets more than 500 hours a year back at the admin desk, and its order entry runs in the evening and at weekends.
When you compare vendors, ask whether "touchless" means no human touch at all or only no correction, and whether accuracy is counted per field or per document. Numbers only compare when the definitions are the same. Our survey of 11 CEOs found that only 4 of them measure what AI delivers, so set the measurement up before the first process starts.
What do the projects that succeed do differently?
| What they do | What stalls the rest |
|---|---|
| One process, the simple 80%. They automate the majority that looks the same every time, and people keep the exceptions | Assuming the ERP is reachable. Write access, posting rights, licences and a test environment that matches production have each delayed a project |
| Owners in week one. They name who is responsible for ERP access, the exception queue and the conversation with the works council | Skipping the master data. Several accounts per customer, units that do not match and articles without a bill of materials. The agent finds them faster than people did |
| The team owns the rules. Correction logs improve the agent and are never used to judge individual people | Switching steps on silently. A step should run on its own only because someone decided it should |
| Hours before and after. They measure manual hours per step and decide in advance where the time goes | Expecting all the hours back during supervision. Checking is quicker than typing, but most of the hours come back when steps run on their own |
Start with one process and let the agent earn more. The free AI adoption playbook for manufacturing goes through the 5 steps in 28 pages, and our guide to back office automation for manufacturers shows what runs in production today. For the finance side, see accounts payable automation for manufacturers and the accounts payable agent. To see an agent work on your own documents, book a demo and bring 5 orders, invoices or quote requests, including one you find difficult.
Frequently asked questions
Where should a manufacturer start with AI?
Start in the back office with one process that has volume, an owner who feels the pain and mostly similar cases. In manufacturing that is most often customer order intake or supplier invoices. A scan ranks your processes by the hours they take, so the first process follows from your own numbers.
How long does AI adoption in a back office take?
The first process typically runs supervised within a few weeks, built by our forward deployed engineers with your team. Our playbook plans 90 days for the first process to reach its first steps on their own, once the ERP can be reached. Each process after the first starts faster.
Does AI adoption mean changing the ERP?
No. The agent fills the same fields a person fills, through the system's own interfaces, so the ERP checks the work the same way. Nothing is migrated. On-premise systems are reached without opening them to the internet, and where there is no API the agent works the screen with the access a person has.
Who decides what the AI agent does on its own?
Your company does. The agent scores how sure it is of each piece of work. Below the line a person decides, above it the work goes through on its own, and only the customer moves the line. A person moves a step up on evidence and one setting moves it back.
How many documents a month justify an AI agent?
As a rule of thumb, about 500 documents a month for one process. Below that, the first agent rarely pays back the ERP connection, unless the process is planning or quoting. Volume matters less when the work is scarce knowledge, as with first-time parts at SPL Treatments.
What should an AI agent never do alone?
We recommend keeping changed bank details, master data changes and orders above a limit with a person. The agent in Lleverage's accounts payable work never releases a payment. Matching and approving should never be the same hand, and the exception queue needs a named owner before go-live.



