Customer story · Mechanical engineering

How Microtechniek stopped order entry waiting on one person

20 to 30 emails a day landed in one shared inbox, and one person keyed every purchase order in them into an on-premise ERP that had no door open for outside systems. Orders that arrived on a Friday afternoon sat until Monday. Back office automation on Lleverage now reads each order, works out who sent it, matches it against Ridder IQ and writes it in, at any hour.

A five-axis machining centre with an automated pallet loader on the Microtechniek shop floorPhoto from microtechniek.nl
Around 60 orders a week land in Ridder IQ on their own

Measured in August 2026: between 45 and 71 purchase orders a week written into the ERP without anyone typing them, including the ones matched to an existing quote. The admin steps in only where the workflow is unsure.

Weekend orders are already waiting on Monday

Order entry used to happen when one person had time for it, in office hours. It now runs through evenings, weekends and holidays, so a Friday-afternoon order is in the system before anyone is back at a desk.

More than 500 hours a year back for the admin desk

Each order took about 10 minutes by hand: open the PDF, identify the customer, check the contact, build the order, key more than 20 fields, file the email. That is now a review of the few orders the workflow flags.

Precision work, and the unglamorous 10 minutes that come before it

Microtechniek is a 130-person mechanical engineering firm in Velsen-Noord, serving heavy industry and aerospace manufacturers. Across 4 semi-autonomous business units it turns and mills components, welds and maintains installations on customer sites, and makes high-precision parts for aerospace, nuclear, optical and semiconductor work. It is the kind of business where a tolerance is measured in hundredths of a millimetre and a wrong article number is a scrapped week.

Work like that is exacting on the shop floor and surprisingly ordinary before it. Nothing gets cut until an order exists in the ERP with the right customer, the right contact, the right delivery date and the right shipping address on it. Somebody has to put it there. In a jobbing engineering firm that somebody reads a purchase order, recognises the company behind it, finds them in the system, and types.

Microtechniek had already done the hard work on its back office. Over the preceding years, workflow tools and Power BI replaced manual processes one after another, and the admin team shrank from 4 people to 1. One category held out, and it was the category that mattered most. Anything that had to read from or write to the ERP stayed manual, because the ERP had no way in.

20 to 30 emails a day, and one person to read them

Every working day, 20 to 30 emails arrived in a shared inbox: purchase orders, but also timesheets, quote requests, service notifications and general correspondence. Each purchase order needed more than 20 fields extracted and resolved by hand: customer identification, order category, delivery dates, contact persons, shipping addresses. Many of them had to be matched against entities already in Ridder IQ, the company's on-premise ERP.

Some of those orders were easy. Others took judgement that is not written down anywhere. One of Microtechniek's largest customers uses inconsistent abbreviations across several sub-locations, so a single order can reference different entities inside the same group, and telling them apart is knowledge that takes months on the desk to build. Get it wrong and the order is booked against the wrong site.

One admin did all of it. Open the PDF. Work out who sent it. Check whether the contact already exists in Ridder IQ. Build the order container. Key the fields. File the email in the right folder. About 10 minutes an order, over an hour a day, every working day.

The part that hurt was not the hour. It was that the hour could only happen when one person was at their desk. Nothing was processed in the evening, nothing at the weekend. An order that arrived at 16:00 on a Friday afternoon waited until Monday morning before it existed anywhere the factory could see it.

Photo from microtechniek.nl

The back office automation Microtechniek runs on now

The reason this bottleneck outlived every other one was the ERP itself. Ridder IQ ran on-premise, and the on-premise installation did not expose the API module the cloud version has. In manufacturing this is a familiar dead end, and it is usually where the conversation about automation stops. 2 other automation projects at similar companies had already failed at exactly this point.

Microtechniek's director, Sebastiaan Röben, framed the project as more than an efficiency exercise. The hours saved on order entry were real, but the bigger prize was the door. If an agent could write into Ridder IQ, the same route would open up invoice processing, procurement and everything else that touches the ERP. So the door came first. In December 2025 a Ridder specialist installed the API module on-site, and the Lleverage team reverse-engineered the cloud API to work with the on-premise environment. Not a standard integration path. It worked.

Joost Verdoorn led the build from the Lleverage side, working with Martin Barnas, Microtechniek's ERP administrator, and Janneke from the improvement team. The design priority was accuracy over speed. Before the agent touched a live order, it had to prove it could handle the full complexity of Microtechniek's real order data, inconsistent customer abbreviations and multi-location orders included.

With a way in, the workflow follows the shape of any good order desk. It first classifies what arrives, because roughly half of what lands in the inbox is not an order: timesheets, quote requests, service entry notifications, general correspondence. Those are routed for separate handling. A real purchase order has its fields extracted, its sender matched against the customer records in Ridder IQ, and where it follows a quotation already in the ERP, the two are linked so the order arrives priced and specified rather than re-entered. A sender the system has never seen gets a contact created rather than stalling the order. Then the order is written and the email filed in the right Outlook folder.

What the workflow will not do is push ahead when it is unsure. Every extraction and every match carries a confidence, and anything below the line goes to a person instead of being guessed at. The admin's day changes from typing everything to looking at the few that need a human.

What happened to each step

Before 6 steps · about 10 minutes per order Now Straight through Review only the orders the workflow flags Exceptions only · runs day and night Open each order PDF AUTOMATED Work out which customer it is AUTOMATED Look the contact up in the ERP AUTOMATED Build the order in the ERP AUTOMATED Key more than 20 fields AUTOMATED File the email in the right folder AUTOMATED

100 real orders before a single live one

Microtechniek's team had been through a hard ERP migration to Ridder IQ a few years earlier, and trust in new systems was not automatic. So the validation was built to earn it rather than ask for it.

Instead of a small pilot and optimism, the agent was fed more than 100 historical purchase orders, the same documents the admin had already processed by hand over previous months, and every extracted field was checked against what had actually been entered into Ridder IQ. The first sample of 40 came back at 99.3% field-level accuracy. The full set of 100-plus reached 99.5%. Corrections made along the way did not stay corrections: each of Janneke's corrections was fed back into the agent as an automated test case, so the same mistake could not come back later.

Production went live on 8 March 2026, a weekend day chosen on purpose, so the orders arriving over the weekend would be the agent's first real workload. By Monday they were sitting in Ridder IQ.

What changed

Since go-live the agent has handled around 120 incoming emails a week. It filters out the half that are not orders and writes the rest into Ridder IQ: 45 to 71 purchase orders a week through August 2026, 60 in a typical week, with a few each week linked to the quote they came from. Processed mail files itself, and new contacts appear in the ERP as they turn up rather than as a blocking question. The admin who spent an hour a day keying now reviews exceptions, and internally the team talks about the agent the way they would talk about a new colleague.

At roughly 10 minutes an order, that is more than 500 hours a year of administrative processing returned across all 4 business units, without anyone added to the admin desk. The hours are the easy number to quote and the least interesting thing that changed. What actually changed is that order entry stopped being a function of one person's availability. The key-person dependency that sat quietly behind the whole process is gone, and because corrections feed back into the agent, accuracy improves as volume grows instead of degrading under it.

From an hour a day to exceptions only

Purchase orders typed into Ridder IQ by hand, per week

Before
~60
After
0
99.5%field-level accuracy across more than 100 real orders, checked against what had been entered by hand
500+ hrsreturned a year: about 10 minutes an order, 60 orders a week, all 4 business units
Any hourevenings, weekends and holidays included, so a Friday order does not wait for Monday
"The order is in the system immediately. We're no longer dependent on when someone has time to process that email."— Martin Barnas, ERP Administrator, Microtechniek

The pattern: when the ERP has no door

Microtechniek's shape is one of the most common in back office automation, and one of the most commonly given up on. An on-premise ERP holds the record. It was never built to be written into by anything other than a person with a keyboard. So every automation idea that touches it dies at the same step, and the business concludes that its systems are the problem.

The tell is easy to spot from the inside: a company that has automated its reporting, its dashboards and its internal admin, and still keys orders by hand. That gap says nobody has opened the door yet.

Opening it is a one-off cost every later project inherits. Order intake usually goes first, because it is high volume and the rules are knowable. Supplier invoices, purchase confirmations and goods receipts follow the same route with different documents. The expensive part was never the reading. It was the writing, and that only has to be solved once.

What comes next

The order agent covers the shared inbox that serves all 4 of Microtechniek's business units. Everything below builds on the route into the ERP rather than starting over.

  • Orders that already know which quote they came from: live. Incoming purchase orders are linked automatically to the existing quote in Ridder IQ, so the quote stops being a document someone has to remember to look up.
  • AI-assisted production planning: a feasibility study is underway. Capacity conflicts and delivery risks currently depend on individual experience rather than systematic oversight.
  • Invoices and procurement through the same door: the reason the ERP was opened first. With a write path into Ridder IQ proven, they become a question of which rules to apply rather than whether the system can be reached at all.

The project that started as a way to build practical AI experience has become the foundation for a broader automation roadmap.

Give your back office an AI workforce