BlogAI

AI in ERP: what it really does and where it stops.

AI in ERP is useful for three things: turning documents into drafts, answering plain-language questions about your data and spotting exceptions. It does not replace people’s decisions, and it is only as good as your data.

What AI means in an ERP system

Today “AI” usually means language models: programs that read and write text, understand questions and pull information out of documents. In ERP they work as an assistant on top of your data, not as a separate system.

Keep two things apart. The ERP holds the data and the rules: orders, stock, prices, permissions. AI helps people work with them faster. If the data is messy, AI will not tidy it up on its own.

So “do you have AI?” is a less useful question than “what exactly does AI do with my orders?”. The answer should be a concrete example, not a promise. Ideally with one of your own documents, in the demo.

Does well: turns documents into drafts

Much of the data entry in a small company is retyping: an order from an email, a supplier invoice, a PDF specification. AI can read the document and fill in a draft in the system.

For example: a customer emails “Please send 200 cartons 40×30×20 and 50 of the big ones, by Friday.” AI recognises the customer and suggests both items and the date. The employee checks what “the big ones” are, corrects and confirms. Instead of typing, they review.

The same goes for supplier invoices. AI extracts the supplier, lines and amounts, and the system compares them with the purchase order. Differences show up before anyone approves the payment.

The key word is draft. The document enters the system only after a person confirms it.

Does well: answers questions about your data

Instead of hunting for a report or waiting for someone to build one, you ask in plain words. “Which invoices are more than 30 days overdue?” “How much board do we need until the end of the month?”

A good answer is not just text. It shows a table and links to the documents, so you can check where the number comes from. If an answer cannot be checked, do not base decisions on it.

The biggest gain is for people who do not work with reports every day. The manager asks on their phone instead of waiting for Friday’s report.

Questions need to be specific. “How are we doing?” gets a vague answer. “How many October orders have not shipped yet, and why?” gets a table you can act on.

Does well: catches exceptions

Many problems show up in the data before they show up on the floor: an order whose stage is falling behind, a material that will run out before the next delivery, a customer who pays later and later.

  • A warning about orders at risk of running late.
  • An alert for material below its reorder point.
  • A price or margin that is out of the ordinary.
  • A short morning summary of what needs action today.

A plain rule can do part of this. AI helps when the exception needs explaining in words or has to be pieced together from several places.

For example: October’s margin is lower than usual. A rule only says “margin dropped”. AI can add that the cause is three orders from one customer with a bigger discount, plus pricier material from a new supplier. That is the difference between an alert and an explanation.

Beware of too many alerts. If people get twenty a day, they stop reading them. Three that genuinely need action are better.

The limits: where AI gets it wrong

  • It can be wrong. Language models sometimes give a confident but wrong answer. That is why important actions go through a person.
  • It depends on the data. If items are duplicated or stock is wrong, the answers will be wrong too.
  • It does not know what was not recorded. An agreement made over the phone and never entered does not exist for AI.
  • It is not accountable. The decision to send, pay or issue stock stays with a person.
  • Permissions and data. It must be clear what AI can see and where the data is processed.

None of this is a reason to give up. It is a reason for AI to work as an assistant with clear limits, not as an autopilot.

The practical rule is simple. The more expensive the mistake, the more careful the review. A draft order takes seconds to check. A supplier payment deserves more attention, whoever prepared it.

Where to start

  1. Tidy your master data. Customers, items and stock without duplicates. Without this, AI only speeds up mistakes.
  2. Pick one tedious task. For example, entering orders from email or supplier invoices.
  3. Run it with confirmation. AI drafts, a person reviews each one. See how many corrections there are.
  4. Add questions and alerts. Once people are used to it, extend to reports and warnings.

Measure the simple things: how long one order takes before and after, how many drafts get corrected and what the most common corrections are. They show where the data or the rules need tidying.

If you still run on Excel, AI is not the first step. The first step is orders and stock in one place.

What to ask the vendor

If all you hear in the demo is “AI will optimise everything”, ask for a concrete example. More questions in the demo questions guide.

How we approach it at Leanovix

AI is coming to Leanovix, and we are building it with the first clients. It will follow the rules in this article: nothing is sent, paid or issued without confirmation, and AI sees only what the user may see. See AI in Leanovix.

Frequently asked questions

Can AI replace an office employee?

Rarely completely. AI removes some of the retyping and searching, while people review, decide and talk to customers.

Do I need AI in my ERP system?

Not necessarily. Sort out orders, stock and invoices first. AI makes sense once the data is right and you handle many documents or questions.

Is my data safe if the ERP uses AI?

It depends on the system. Ask what AI can see, where data is processed, whether it is used for training and whether there is an audit log of actions.

Show us how you work today.

In 30 minutes we’ll find the process that costs you the most and see how it would look in Leanovix.

Book a demo