Skip to content
All articles
Industry 8 min read· by Lars Zimmermann

Order entry in the ERP: from twelve minutes to 26 seconds

A field report from my own company: what order entry used to cost, how it became 26 seconds and which control had to stay in place for that to work.

In short

In my precision engineering company, order entry took around twelve minutes and cost EUR 5.14 per transaction. Today it is 26 seconds and EUR 0.06, measured in live operation, not in a demo. The gain came from the process before the model: rules first, then automation. Human approval before posting has stayed, because liability stays in the company.

Auf Deutsch lesen: deutsche Fassung

Twelve minutes. That is how long it took us to enter an order. Customer data, line items, dates, prices, all by hand into the screen. Today it is 26 seconds. The process cost per transaction has fallen from EUR 5.14 to EUR 0.06. With around 3,100 order entries a year, the time saved comes to around 600 hours, or around 75 working days of a skilled employee, which used to disappear into typing.

I am not writing this as a consultant with a slide. I run a precision engineering company. It is my own business, my own money and my own risk. The figures were measured in live operation, not extrapolated in a demo. And the most interesting part of the story is not the technology. It is what I had to do before a single line of automation ran at all.

First I measured, and that was unpleasant

Before anything was automated, I stood alongside with the stopwatch. Not once, but over several weeks and across different types of order. The rush order on Friday afternoon behaves differently from the series order on Tuesday morning, and anyone who measures only one kind measures the wrong one.

The result was uncomfortable. Twelve minutes on average, but the range was the real problem. Some orders were in within four minutes, others took half an hour because a detail was missing and someone had to make a telephone call. Anyone who knows only the mean automates past their own company.

The EUR 5.14 needs an explanation, otherwise the figure is worthless. These are full costs per transaction, meaning personnel and system shares, not the hourly wage divided by five. I add this because such figures are readily dressed up in marketing. If you work out your own and stay honest in doing so, it will be higher than your first estimate. Mine was.

Without a before value, every improvement is a claim. With a before value, it is evidence.

Measuring brought something else that I had not expected. A good part of the twelve minutes was not down to typing at all. It was down to information being missing. No tool in the world speeds up a query to the customer. Anyone who does not record this share separately credits themselves afterwards with an improvement that is not one.

75 days a year, and who sits through them

75 working days a year. That is the figure that sticks with managing directors, and so it should. But the real question is not how many days it is. It is who sits through them.

With us it was not the cheapest person in the company. It was the one who can read an order. Who notices while capturing it that a tolerance does not match the drawing, that a delivery date is unrealistic for this material, that someone has ordered the wrong grade. Exactly this person sat there and typed it out.

In a labour market in which you search for months for someone like that, this is not only expensive. It is the worst use you can find for this knowledge. Added to that is a second effect that no cost accounting shows: anyone who fills in fields for half the morning eventually stops looking. Routine disguises itself as care.

BeforeToday
Time per order entryaround 12 minutes26 seconds
Process cost per transactionEUR 5.14EUR 0.06
Who typesskilled employee with manufacturing knowledgenobody; the skilled employee checks and approves
Time per year at around 3,100 ordersaround 620 hoursaround 22 hours; saved: around 600 hours, i.e. 75 working days
Measured in the live operation of a precision engineering company, not in a demo environment.

What actually does the work

By now 17 agents work in our system. That sounds like a lot and is the least spectacular part of the story. With us an agent is not a digital employee and not a new species. It is a piece of software with a narrowly defined task, fixed limits and a log.

The point I rarely hear on stages: these agents are not the cause of the improvement. They are its result. Before each one of them stood the question of how we actually do this step, and at the start that often could not be answered. Not because nobody knew, but because it ran a little differently at three workstations.

  • Which details does an order absolutely need before it may go into production, and which are a convenience? This list did not exist in writing before. It was in people’s heads, and slightly different in each head.
  • How do you recognise a special case? As long as the special case is only a gut feeling, no rule can catch it and no model can learn it. It has to become describable before it becomes automatable.
  • What happens when a detail is missing? Previously someone made a call. You have to answer this question afterwards too, otherwise the automation silently produces gaps that are only noticed in production.

First process, then tool, then AI. The sentence sounds like a calendar motto until you have tried it once in the wrong order. Anyone who automates an unclear rule does not get clarity. They get an unclear rule that now runs faster and is harder to correct.

AI does not fix bad processes. It makes them bad faster.

I worked on this for two years. Not two years of programming, but two years of operation, readjusting, collecting special cases and rebuilding again. That is part of the honest version of these figures. Anyone who sells you the same improvement as an installation over a weekend either meant a different company or calculates differently.

The part nobody talks about on stages

Order entry is not a harmless operation. Dates, prices, material scheduling and in the end accounting documents arise from it. As soon as an automation touches data relevant to the accounts, the same requirements apply as before. Only now nobody sits alongside any more who notices in passing.

That is why three things have stayed with us, although each of them costs seconds:

  • An approval before posting. A person sees the finished order and confirms it. Not for every field, but at the end. Anyone who removes this approval saves a few seconds and loses the only point at which responsibility becomes visible.
  • A log that can do more than “done”. For every transaction it is recorded which version of the rules ran, which data came in and what became of it. Without this state you cannot explain afterwards why an order looks the way it looks. For everything that runs into the bookkeeping, exactly this traceability is not optional.
  • Loud error behaviour. If something does not fit, the transaction stops and reports in. A run that does nothing when an error occurs and still reports success is more dangerous than no automation at all, because after that nobody looks any more.

From my audit practice I know the counter-model well. There an automation runs under the personal account of the colleague who built it, with full rights and no expiry date, and nobody can say which version of the rules applied three months ago. Technically that works splendidly. It is not auditable, and when something goes wrong it is still the managing director whose name is on the letterhead.

Human oversight is not a brake block you install so that a standard is satisfied. It is the point at which you notice that something is going wrong before the customer notices. And it has to be effective, not merely described. If the approving person confirms thousands of transactions a month and never rejects anything, the approval is a ritual and not a control. Incidentally, that is exactly the finding I write most often in audits.

What did not work

So that this does not turn into advertising copy, here is the other half.

The first attempt was too ambitious. I wanted the complete path from the incoming enquiry to release for production in one go. That failed because too many special cases came onto the table at the same time and we could no longer tell apart where it was getting stuck. The second attempt took a single step, let it run alongside for two weeks without being allowed to write, and only switched it live after that. These two weeks of idle running brought more than any list of requirements.

Second, the 26 seconds apply to the normal case. The special case still takes as long as before, sometimes longer, because now someone has to check why the automation did not like it. That is as it should be. Anyone who draws the average across everything and advertises with it is selling a figure that does not arrive in everyday work.

Third, the objection that always comes: my master data is too bad, I cannot start yet. That is not true. The data is not right today either, it is just that a person balances out the gaps in their head. The automation does not make this error bigger, it makes it visible. That is unpleasant and exactly the point. The clean-up is not the precondition, it is the first step of the work.

How to measure your own process

If you want a robust figure for your company, you need neither a project nor a consultant for it. You need two weeks of attention and the willingness to bear the result.

  • Separate the types of order before you measure. Standard, rush order, repeat part, special case. Four figures tell you more than one.
  • Measure instead of estimating. Two weeks, with the clock, at several workstations. Experience shows that estimates are too low, because nobody counts the queries.
  • Note the range, not just the mean. The fastest and the slowest transaction show you where the rules are missing.
  • Record waiting times separately. Time spent waiting for an answer is not automation potential. It is a process problem and needs to be solved in its own right.
  • Calculate full costs per transaction, not just personnel time. Systems, licences, error correction, rework. This figure is the one you will later calculate against.
  • Only then decide which single step comes first. And let it run alongside for two weeks before it is allowed to write.

At the end of these two weeks you have something most companies do not have: a baseline. Without it you can only claim every improvement. With it you can prove it, and to yourself as well, when the enthusiasm wanes after a few months.

Are you thinking about where automation should start in your company, and do you want not a pitch but an honest conversation with someone who has the same problems? That is exactly what the sparring format is for.

To the AI sparring session

Measure your order process before you automate it

  1. Separate the types of order

    Look at standard, rush order, repeat part and special case separately. Four figures say more than one mean across everything.

  2. Measure for two weeks instead of estimating

    With the clock, at several workstations, across different days of the week. Estimates are regularly too low because queries are not counted.

  3. Note the range

    Record the fastest and the slowest transaction. The range shows where rules are missing. The mean hides it.

  4. Record waiting times separately

    Time spent waiting for an answer is not automation potential, but a process problem in its own right.

  5. Calculate full costs per transaction

    Add together the personnel share, system share, error correction and rework. This figure is the later benchmark.

  6. Select a single step

    Take the step with the largest share and the fewest special cases first, not the whole operation at once.

  7. Let it run alongside for two weeks without write access

    Observe the automation first before it writes data. The idle running shows the special cases nobody thought of in conversation.

Share: LinkedIn E-Mail

Frequently asked questions

Are the figures from a demo or from real operation?+

From real operation. Twelve minutes to 26 seconds and EUR 5.14 to EUR 0.06 per transaction were measured in my own precision engineering company, over several weeks and across different types of order. The EUR 5.14 are full costs including personnel and system shares, not an hourly wage divided by five.

Do you need 17 agents for this?+

No. The 17 grew over two years, one after the other, and each has a narrowly defined task. The benefit does not come from their number, but from the fact that the process behind each one was clarified beforehand. Anyone who starts with one step and gets it clean gets further than someone who starts ten at once.

Does a person have to approve every automated order?+

With us yes, at the end of the transaction, not for every field. The approval costs seconds and is the point at which responsibility stays visible. What matters is that it is effective: if nothing is ever rejected, it is a ritual and not a control. That is exactly the point where I probe most often in audits.

My master data is unreliable. Should I start anyway?+

Yes, but knowing that the clean-up is part of the work and not its precondition. The data is not right today either, it is just that a person balances out the gaps in their head. Automation does not make this error bigger, it makes it visible.

What has to be observed with data relevant to the accounts?+

That a business transaction must remain traceable in how it came about. For every transaction it should be recorded which version of the rules ran, which data came in and what became of it. This includes a dedicated account for the automated service instead of the personal access of the colleague who built it, and error behaviour that reports in rather than carrying on silently.

Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)

Last updated: 06 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.

Questions about your own case?

In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.

Continue reading