Generative AI is moving quickly into the systems that companies use every day.
That includes ERP.
This is not surprising. An ERP system already knows a lot about a business. It knows what was sold, what was purchased, what is in stock, what is being produced, which customers pay late, which suppliers deliver on time, where costs are rising and where something does not look quite right.
Now imagine adding AI on top of all that.
Suddenly, the ERP does not only store information. It can begin to interpret it.
It can notice patterns.
It can warn people.
It can compare today with last year.
It can explain why a margin may be falling.
It can tell a production manager that scrap is increasing.
It can tell purchasing that a supplier has become more expensive.
It can warn finance that a customer who normally pays in 15 days is now at 42 days.
This is where AI becomes very interesting.
But it is also where we need to be careful.
Because there is a very big difference between AI helping a company make better decisions and AI making the decisions for the company.
That difference matters more than ever.
Imagine this happening on a normal Monday morning
A production manager opens SIX ERP at 8:10 in the morning.
Normally, he would look through a dashboard. Maybe he checks yesterday’s output, compares some numbers, speaks with a shift supervisor and then slowly understands what happened.
With AI, the system could greet him differently.
It could say:
Production output yesterday was normal, but material consumption was 8.7% higher than expected. Most of the difference came from Production Line 3. Scrap on that line increased during the late shift. Supplier prices were unchanged.
That is useful.
Very useful.
The manager does not have to search through ten reports.
He already knows where to look.
But now imagine the AI continues:
Recommended action: stop Production Line 3 immediately.
That is a different matter.
Maybe stopping the line is correct.
Maybe it is completely wrong.
Perhaps the scrap was caused by a one-time test batch.
Perhaps a new worker entered the wrong quantity.
Perhaps material was booked late.
Perhaps maintenance already fixed the problem at 6:00 in the morning.
The AI can see the data that is available to it.
The experienced production manager may know the story behind the data.
And that is the point.
AI can be incredibly good at finding the signal. Humans are still extremely important when deciding what the signal really means.
AI may become one of the best controllers a company has ever had
There is a lot of talk about AI replacing jobs.
In an ERP environment, one of the most valuable uses may be something different.
AI can become a permanent controller.
Think about how businesses often work today.
At the end of the month, somebody creates a report.
Management looks at the report.
Someone notices that costs went up.
Then questions begin.
Why did costs go up?
Was it purchasing?
Was it production?
Was it energy?
Was it overtime?
Was it transportation?
People start searching.
A few days later, the company understands what happened three or four weeks ago.
AI can change that.
Instead of waiting until the end of the month, the system can watch continuously.
It can compare thousands of transactions as they happen.
It can notice that material prices are slowly rising.
It can notice that one warehouse has unusual stock movements.
It can notice that one sales person is giving discounts that are far above average.
It can notice that overtime is increasing at the same time that production output is falling.
It can even notice small changes before a human would normally consider them important.
That is one of the strongest cases for AI inside ERP.
Not because AI should become the boss.
But because AI can become the colleague who never gets tired of checking the numbers.
The machine sees everything. The human sees the situation.
Let us take a simple supplier example.
The ERP shows two suppliers.
Supplier A sells a component for €9.80.
Supplier B sells the same component for €9.25.
Supplier B is cheaper.
Delivery performance is almost the same.
Quality is almost the same.
An AI system may look at the numbers and say:
Move more purchases to Supplier B. Estimated annual savings: €74,000.
That recommendation may be perfectly logical.
But then the purchasing manager says:
“No.”
Why?
Because five years ago, when another supplier failed during the busiest month of the year, Supplier A opened its warehouse on a Saturday night and sent a truck.
Because Supplier A knows the company’s production schedule.
Because when the factory suddenly needs an extra 5,000 pieces, Supplier A usually finds a way.
Because the purchasing manager knows that Supplier B looks excellent on paper but becomes unreliable every December.
Where is that information?
Maybe nowhere.
It may not be in the ERP.
It may not be in a contract.
It may live inside the head of a person who has worked in purchasing for twenty years.
This kind of knowledge is often called tacit knowledge.
It is the knowledge people collect by doing the job.
They learn which customer always says an order is urgent even when it is not.
They learn which machine makes a strange noise two days before something breaks.
They learn which supplier can be trusted when everybody else says no.
They learn which employee can solve a difficult production problem without making a big drama about it.
A database is full of facts.
Experience is often full of context.
A good company needs both.
This is also a generation question, but probably not in the way people think
It would be easy to say:
“Older employees understand business. Younger employees understand technology.”
That would be far too simple.
There are brilliant young employees and terrible experienced ones.
There are older employees who understand AI extremely well.
And Millennials are no longer really a new generation in the workforce. Many are now experienced managers, specialists and company leaders.
The more interesting point is about experience itself.
Gen Z is entering companies at a time when AI is becoming normal.
For many younger employees, asking an AI system for help may feel as natural as searching Google felt to the generation before them.
That can be a huge advantage.
A 22-year-old employee may be able to use AI faster and more naturally than somebody who has worked in the company for 30 years.
But there is something AI cannot simply give that employee overnight.
It cannot give them the memory of having seen ten different crises.
It cannot give them the feeling of knowing when a customer is really angry and when they are simply negotiating.
It cannot give them the memory of a supplier failure from 2018.
It cannot give them the experience of seeing a production plan that looks perfect in Excel and then falls apart on the factory floor.
Those things come from time.
The smartest company will not choose between young digital talent and experienced employees.
It will connect them.
Imagine a young buyer who is comfortable working with AI sitting next to a purchasing manager with 25 years of experience.
The AI finds the unusual price movement.
The younger employee understands how to explore the data.
The experienced manager explains why the number matters.
That is far stronger than any of them working alone.
The real danger is not always that AI is wrong
Sometimes the bigger danger is that people stop questioning it.
Imagine an AI recommendation appears on a beautiful dashboard.
There are charts.
There is a confidence score.
There is a paragraph explaining the conclusion.
It looks professional.
The system says:
Reduce finished-goods inventory by 18%.
A manager thinks:
“The AI looked at millions of records. It probably knows.”
Click.
Approve.
The next suggestion appears.
Approve.
Another one.
Approve.
After some time, the human technically still approves the decisions.
But they are no longer really deciding.
They are simply clicking.
This is one of the most important risks around AI.
People naturally trust systems that appear confident.
And modern AI is extremely good at sounding confident.
That can be dangerous.
A person who says, “I am not sure,” makes us careful.
A machine that gives a three-paragraph explanation with numbers and a neat conclusion can make us relax too much.
But the explanation can still be wrong.
A confident answer can still be a bad answer
Generative AI does not think like a calculator.
A calculator either gives you 2 + 2 = 4 or it is broken.
AI is different.
It creates answers based on patterns.
That means it can sometimes produce something that looks completely believable even when part of it is incorrect.
This becomes particularly dangerous inside business software.
Imagine an AI agent tells management:
Profitability of Product Group C declined because freight costs increased 14%.
Everybody starts discussing logistics.
But what if freight costs were entered twice in one period?
What if a currency conversion was wrong?
What if one large invoice was booked in the wrong month?
The AI may have analysed the numbers correctly.
The numbers themselves were wrong.
This is why data quality becomes even more important when AI enters ERP.
Bad data was always a problem.
AI can make bad data more dangerous because it can turn it into a very convincing story.
AI can also misunderstand why something happened
This is another important problem.
Suppose an ERP notices that whenever overtime goes up, quality goes down.
That is a real pattern.
The AI may report:
High overtime appears to reduce production quality.
That could be true.
Workers may be tired.
But maybe the real story is different.
Perhaps overtime increases because the company is producing unusually difficult products.
Those products also create more quality problems.
In that case, overtime did not cause the problem.
Both things were caused by something else.
Humans make this mistake too.
But when AI produces an answer quickly and confidently, people may stop looking for other explanations.
That is why AI should often say:
"Here is the most likely explanation."
Not:
"This is definitely the cause."
There is a big difference.
The past can also teach the wrong lesson
AI learns from data.
Businesses usually have a lot of historical data.
That sounds perfect.
But history is not always something we should repeat.
Imagine that a company has always kept too much stock.
If AI learns from the company’s old behavior, it may decide that high inventory is normal.
Imagine that management historically gave the best projects only to a small group of people.
If an AI system learns from those old decisions, it may continue the same pattern.
Or imagine the company has always accepted poor payment behavior from one major customer.
AI may learn that this is simply how business is done.
Historical data tells us what happened.
It does not automatically tell us what should happen next.
Sometimes the whole reason management introduces a new system is because the old way was wrong.
And sometimes the world simply changes
This is where experience becomes very important.
AI loves patterns.
Business loves to break patterns.
A war starts.
A border closes.
A supplier goes bankrupt.
A new tax rule appears.
Energy prices jump.
A factory floods.
A major customer announces that it is closing a plant.
Suddenly, the previous five years are not a very good guide to next month.
People who have been through unusual situations before may recognize that something has changed.
They may say:
“Stop using normal assumptions. This is not a normal year.”
An AI system can also detect change, but it needs to be designed to know when its old patterns are becoming less useful.
A very good AI should sometimes say:
"I do not have enough reliable historical information for this situation."
That is a good answer.
One of the biggest mistakes is giving AI the wrong goal
Imagine telling an AI agent:
Reduce inventory by 20%.
It may succeed.
Inventory goes down.
Great.
But customer deliveries become slower.
Production starts waiting for components.
Emergency freight increases.
The factory saves €200,000 in inventory but spends €300,000 fixing the problems.
The AI achieved the target.
The company lost money.
Or tell an AI:
Reduce purchasing costs.
It switches to cheaper suppliers.
Six months later, quality claims have doubled.
Or say:
Improve short-term cash flow.
It delays supplier payments.
Cash flow improves this month.
Then two important suppliers reduce the company’s credit terms.
The AI may have done exactly what it was asked to do.
The problem was the question.
Businesses are connected systems.
You cannot normally optimize one number without affecting something else.
Experienced managers know this because they have seen it happen.
AI needs rules, boundaries and multiple goals.
Business is not only numbers
This may sound obvious, but it is easy to forget.
Companies are also relationships.
A customer may be unprofitable this year but strategically important.
A supplier may be slightly expensive but incredibly reliable.
An employee may have average productivity figures but be the person everybody calls when something goes wrong.
A sales manager may decide not to increase a customer’s price because they know a much larger project is coming.
An AI system may not know these things unless somebody has put them into the system.
This is why fully self-running companies are such an attractive idea and such a dangerous one at the same time.
The numbers can look clear.
Real life is often messy.
The faster AI works, the faster it can also make mistakes
There is another thing people sometimes forget.
Automation increases speed.
That is usually considered a benefit.
But mistakes also become faster.
A human buyer might make one bad purchase order.
An automated AI agent could make 500 before lunch.
A human manager might approve one poor discount.
An AI pricing agent could change prices for 20,000 customers.
This does not mean we should avoid automation.
It means automation needs brakes.
Cars became faster over time.
They also got better brakes, seat belts, sensors and safety systems.
AI in ERP needs the same thinking.
The more authority we give the system, the stronger the controls around it need to become.
Several smart AI agents can still create one stupid result
This may become a real ERP challenge.
Imagine a future company with several AI agents.
One controls inventory.
One manages purchasing.
One watches cash flow.
One optimizes pricing.
One monitors suppliers.
Each one has its own job.
Now something strange happens.
The inventory AI decides stock is too high and reduces safety stock.
The purchasing AI sees shortages coming and starts making urgent purchases.
The finance AI sees cash requirements going up and delays supplier payments.
The supplier-risk AI notices complaints from suppliers and lowers their scores.
The purchasing AI then changes suppliers.
Production receives slightly different material.
Quality drops.
The sales AI sees customer complaints and starts giving discounts.
Every AI agent followed its own logic.
Together they created a mess.
This is why a company cannot simply add more and more agents and assume it becomes smarter.
Somebody has to look at the whole system.
That may be another AI layer.
But it should also involve humans.
There is another danger we should talk about: people may stop learning
Imagine a junior financial controller joining a company.
In the past, that person might spend years learning.
They would investigate strange numbers.
They would talk to purchasing.
They would argue with production.
They would find booking errors.
They would learn why cash flow is different from profit.
They would slowly build experience.
Five years later, they become very good.
Now imagine AI does all the difficult investigation.
Every morning, the employee gets a message:
Variance identified. Cause: increased material usage. Recommended correction: €47,320.
The employee presses Approve.
The work is faster.
But what did they learn?
This is a serious long-term question.
If AI performs all junior analytical work, where will the future senior experts come from?
Companies should be careful not to automate away the learning process.
AI should help younger employees understand the business.
It should not make understanding unnecessary.
This is where good ERP design can make a difference
Imagine instead that the AI says:
Material usage is 8.4% above standard. The biggest difference comes from Line 2 during the night shift. Three possible causes are unusually high scrap, incorrect booking of material, or a change in the production recipe.
Then the system asks:
Would you like to see the underlying transactions?
Now the employee learns.
They can open the material movements.
They can inspect the production order.
They can see which shift was affected.
They can compare it with previous orders.
The AI is not simply giving the answer.
It is helping the employee investigate.
This is what we mean when we say AI should make people stronger.
AI should be allowed to do more when the risk is small
Not every decision needs a committee.
There is no reason a company should require three signatures because AI wants to order €35 worth of office paper.
Some tasks are repetitive, low-risk and easy to reverse.
AI should be allowed to automate many of those.
For example, automatically categorizing incoming emails is usually low risk.
Creating a draft purchase order is low risk.
Reordering cheap packaging material within a known limit may also be reasonable.
But the rules should change when consequences become bigger.
An AI system recommending a €400,000 purchase should probably not execute it by itself.
An AI system evaluating whether an employee should be dismissed should certainly not be the final decision-maker.
An AI system can support those decisions.
It should not own them.
A useful rule is simple:
The harder a mistake is to reverse, the more human control should remain.
Think of AI as climbing an autonomy ladder
A company does not need to decide between completely manual and completely autonomous.
There are many steps in between.
At the first level, AI simply informs.
It says:
Stock will probably run out in nine days.
At the next level, AI recommends.
We suggest ordering another 3,000 pieces.
Then AI can prepare the action.
Purchase order created and ready for approval.
Later, within certain limits, AI can execute.
Packaging material below safety stock. Order value €280. Automatically reordered.
That can make sense.
But the higher AI climbs, the more carefully the company should ask:
Who is responsible?
What can the AI change?
How much money can it spend?
Which customers can it affect?
What happens if it is wrong?
Can we undo the action?
Can we stop the system immediately?
Those are management questions, not only technical questions.
Human control must be real control
There is a popular phrase in AI:
Human in the loop.
It sounds safe.
But it is not enough by itself.
Imagine the AI creates 400 recommendations every day.
A manager has thirty minutes to approve them.
Click.
Click.
Click.
Click.
There is technically a human in the process.
But that person is not really controlling anything.
They are simply becoming the person whose name appears in the audit trail.
Real human control means the person has enough time, knowledge and information to disagree.
That means the ERP should show why the AI made the recommendation.
The manager should be able to see the source data.
The system should show uncertainty.
And the manager should be able to say:
“No. I know something the system does not.”
That is real control.
The most useful button might be “Challenge AI”
Imagine a purchasing manager receives this recommendation:
Change preferred supplier from A to B. Expected saving: €74,000 per year.
The manager disagrees.
Instead of simply clicking Reject, they write:
Supplier A provides guaranteed emergency capacity during peak season. Supplier B does not.
Now something interesting happens.
The human has just added knowledge to the system.
Next time, the AI can take that information into account.
This is where human experience and AI start learning from each other.
The ERP no longer contains only transactions.
It begins to contain the reasons behind decisions.
That can become extremely valuable.
Especially when experienced employees retire.
We should use AI to preserve experience, not throw it away
Many companies have people who know things nobody else knows.
The warehouse manager who knows why a certain product should never be stacked three pallets high.
The production manager who knows that one machine behaves badly in very hot weather.
The salesperson who knows that one customer always negotiates aggressively in December but always buys in January.
The purchasing manager who knows which supplier will actually answer the phone at 2:00 in the morning.
Companies lose this knowledge when people leave.
AI creates an interesting opportunity.
Instead of asking:
“How can AI replace our experienced employees?”
we should ask:
“How can AI help us capture what our experienced employees know?”
That is a much better question.
If ERP systems begin storing not only decisions but also explanations, exceptions and lessons learned, AI can help future employees use that experience.
AI can become a teacher for the next generation
This may be one of the most positive uses of AI.
Imagine a young employee sees a supplier delivery problem.
Instead of simply saying what to do, the system explains:
Before changing supplier, check current contracts, open purchase orders, quality history, alternative capacity, seasonality and any strategic agreements.
That employee begins to understand how an experienced buyer thinks.
Imagine a junior production planner is about to move an order.
The AI says:
This change solves today's capacity problem but will create a material shortage on Thursday.
Now the employee learns to think two steps ahead.
This is much more valuable than AI simply doing the work silently.
A great AI system should sometimes answer a question.
Sometimes it should ask one.
What AI is really good at
There are areas where humans should happily let the machine do a lot of the work.
AI is very good at looking through huge amounts of data.
It can compare thousands of invoices.
It can notice unusual discounts.
It can detect strange stock movements.
It can find suppliers whose delivery performance is getting worse.
It can compare real material consumption with planned consumption.
It can forecast cash flow.
It can estimate demand.
It can explain what changed in the business.
This is especially powerful because most companies already have the information.
The problem is that the information is spread across too many tables, reports and screens.
AI can turn that data into a conversation.
A CEO may simply ask:
Why did profit fall in Germany last month?
The system can begin investigating.
Then:
Which customers caused most of the difference?
Then:
Was it lower sales prices or higher costs?
Then:
Show me the five largest unusual transactions.
That changes how people interact with ERP.
The system no longer only reports.
It helps people think.
AI can also become the colleague who keeps asking uncomfortable questions
This may be even more valuable.
Good control is not only about finding mistakes.
It is about challenging assumptions.
Imagine a salesperson wants to give a 17% discount.
The AI says:
This customer's average discount is 6%. At 17%, the order falls below the company's normal gross margin target. Please explain the reason.
That does not stop the salesperson.
Maybe there is a very good reason.
Perhaps this is the first order of a much larger contract.
But now the decision is visible.
Or imagine purchasing continues buying from an expensive supplier.
AI asks:
Comparable suppliers are 8% cheaper. Why is this supplier still preferred?
Maybe the answer is reliability.
Good.
Write it down.
AI becomes a second pair of eyes.
That is a very strong role.
A good AI system should also know when to say “I don’t know”
This may sound strange.
People usually expect AI to answer.
But one of the signs of a mature business system is knowing when not to act.
Imagine the AI finds that two databases disagree.
One says inventory is 4,800 pieces.
The other says 5,600.
The worst possible behavior would be to quietly choose one and continue.
The better response is:
I cannot make a reliable recommendation because two important data sources disagree.
Or:
This transaction is outside my authority.
Or:
Market conditions are very different from the period used for this forecast.
Or:
Human decision required.
That is not weakness.
That is good control.
The future company should not be built around one kind of intelligence
The strongest businesses may eventually combine three kinds of intelligence.
The first is machine intelligence.
Fast.
Patient.
Very good at comparing, calculating and finding patterns.
The second is organizational intelligence.
This lives inside the ERP.
The company’s rules.
Its workflows.
Its transactions.
Its responsibilities.
Its permissions.
Its history.
The third is human intelligence.
Experience.
Judgment.
Relationships.
Ethics.
Creativity.
Common sense.
Responsibility.
The winning company is not the company that removes one of these layers.
It is the company that gets all three working together.
So what should SIX ERP become in this world?
For us, the future of ERP is not a system that quietly takes over the company.
It is a system that makes the company easier to understand.
The ERP should still be the reliable source of business information.
Business Intelligence should help people see what is happening.
AI should make it easier to ask questions, find problems, understand patterns and discover risks.
Agents can go further and perform controlled actions.
But those actions should have boundaries.
The system should know:
what it may do,
what it may recommend,
what it must ask permission for,
and when it must stop.
That is the kind of autonomy we believe makes sense.
Not maximum autonomy.
Controlled autonomy.
The real question is not “Can AI do it?”
For almost every software demonstration today, somebody asks:
“Can AI automate this?”
More and more often, the answer is yes.
That is no longer the most interesting question.
The better question is:
Should it?
And if yes:
Under which conditions?
With which limits?
Who checks it?
Who is responsible?
What happens when it is wrong?
Can it be stopped?
Can the action be reversed?
Will employees still understand the process in five years?
These questions are more important than a flashy AI demo.
One day, companies may regret building systems nobody understands
Imagine a company ten years from now.
AI plans production.
AI orders material.
AI sets prices.
AI manages inventory.
AI evaluates suppliers.
AI creates financial forecasts.
Everybody is impressed.
Then something unusual happens.
The system begins making bad decisions.
Management asks why.
Nobody knows.
The people who understood the old processes have retired.
The younger employees never learned them because the AI always handled everything.
The company has built a very advanced system.
But nobody really understands the company anymore.
That is not digital transformation.
That is dependency.
We should avoid building that future.
A better future is possible
Now imagine something else.
The AI watches the entire company.
It notices strange things immediately.
It warns people before problems become expensive.
It explains where costs are moving.
It prepares orders.
It checks invoices.
It helps managers understand the business.
It teaches younger employees how experienced people think.
It records the knowledge of employees who have spent twenty years solving problems.
It challenges decisions when numbers look strange.
And when a decision becomes important, complicated or dangerous, it says:
This needs a human.
That is a much stronger vision.
It uses the best part of AI without pretending that software has suddenly gained twenty years of business experience.
AI should run the numbers — not run the company
Generative AI will become part of ERP.
That is not really a question anymore.
The opportunity is too large.
AI can make companies faster.
It can make them more transparent.
It can find risks earlier.
It can make business intelligence easier to use.
It can remove a huge amount of repetitive work.
It can help younger employees learn faster.
And it can give experienced employees analytical power they never had before.
But companies should be careful with one idea:
autopilot.
A business is not an airplane flying through an empty sky on a fixed route.
It changes constantly.
Customers change.
People change.
Markets change.
Suppliers change.
Regulations change.
Unexpected things happen.
Experience still matters.
Responsibility still matters.
Judgment still matters.
The future should therefore not be:
Humans or AI.
It should be:
Humans with AI.
Let AI search.
Let it calculate.
Let it monitor.
Let it compare.
Let it warn.
Let it recommend.
Let it perform safe actions when the rules are clear.
But when the consequence is serious, keep a responsible person in control.
Because the best ERP of the future should not be an autopilot for the enterprise.
It should be a much better cockpit for the people responsible for flying it.
Sources and further reading
This article is based on research and guidance from organizations including the European Commission on human oversight under the EU AI Act, NIST’s AI Risk Management Framework, OECD research on generative AI and workplace learning, International Labor Organization research on AI and jobs, Microsoft Research on AI and critical thinking, Stanford’s AI Index, and research published in Nature on human reliance on AI systems.
