ERP data quality decides whether an AI connected to your management system produces sound analysis or very convincing mistakes. In food and drink SMEs, six flaws come up almost every time. You do not need a perfect database: fix the flaws that affect the data you open to AI, then keep to a data entry rule.
This article is for business leaders considering connecting an AI agent to their ERP. It covers the six most common flaws in the sector, what each one distorts, and a four-step method to clean up your data without an IT department.
Key takeaway
six flaws distort most calculations in the food industry: duplicate customers, mixed units, misclassified items, outdated cost prices, rebates outside the ERP and purchase prices in free text. You do not clean the whole database, only the data blocks opened to AI, often three to five. The method has four steps: list, sort, fix, set a data entry rule. Each block has a named owner, or the database degrades again within a few months.
ERP data quality: why AI makes it visible
Your ERP records, and each team reads the part that concerns it. A data entry error often stays invisible: sales administration knows the customer recorded under two codes, the financial controller removes year-end rebates by hand.
An AI calculates and cross-checks, without your teams' unwritten knowledge. An inconsistency that went unnoticed in a spreadsheet becomes a false alert or, worse, a missed one.
| Situation | ERP used on its own | AI connected to the ERP |
|---|---|---|
| Customer recorded under two codes | Sales administration makes the link mentally | Two separate customers, each with a distorted ordering pattern |
| Sales in kilos and in cases for the same item | The sales rep converts mentally | Volumes added up without conversion, misleading trends |
| Last year's cost price | Corrected in a separate Excel file | Margin calculated on an outdated cost |
| Year-end rebates outside the ERP | Removed by hand at year end | Margin per retailer overstated all year |
This is also why an AI agent does not replace a solid management system, a point we develop in our article on why ERP vendors lag behind on AI.
The 6 flaws that stop you from cleaning ERP data in the food industry
These flaws are not specific to one vendor. They appear in EURAGRO, Sage, VIF or Divalto, because they come from data entry habits, not from the software.
1. Duplicate customer records: one retailer under several codes
The same retailer appears under several codes: one per delivery platform, an old code kept after a move, an account created in a rush. Your teams know it is the same customer. The ERP does not.
To an AI, these are two or three different customers. The retailer's revenue is split, its ordering pattern looks irregular, and an agent tasked with spotting customers who are drifting away may flag a drop that does not exist.
The check: sort the customer list by name and company registration number, and attach each secondary code to a main account.
2. Mixed units: kilos, cases, pieces and variable weight
In the food industry, the same item is often sold by the kilo to one customer, by the case to another, by the piece in self-service. Variable-weight products add another layer: the invoiced weight differs from the theoretical weight on the item record.
If conversion factors are missing or wrong, the AI adds kilos to cases. Volumes per item can no longer be compared from one month to the next, and the average price per kilo becomes meaningless.
The check: make sure every item sold in several units has a conversion factor filled in and consistent with the actual average weight.
3. Item master data: families and ranges wrongly assigned
Item master data (family, range, brand, dedicated customer) is used to group sales. Over time, items are created by copying another one and keep the wrong family, or end up in a "miscellaneous" category that grows every year.
An AI analysing profitability by range then compares the wrong sets.
The check: export active items with their family, and look first at the "miscellaneous" category and at items created in the last two years.
4. Cost prices and bills of materials that no longer keep up
The cost price of an item depends on its bill of materials (raw materials, packaging, yield) and on purchase prices. In many SMEs, it was calculated once, then updated in a separate Excel file, not in the ERP.
Yet it is the data an analysis agent uses most: it is what tells you whether a raw material increase pushes an item below its margin threshold. A cost price six months out of date makes the alert useless.
The check: date the last update of the cost price of your main items, and check that the yields in the bills of materials match what the production floor actually sees.
5. Year-end rebates and retailer terms managed outside the ERP
Year-end rebates, listing fees and some terms negotiated with large retailers often live in a spreadsheet in the sales or accounts department. The ERP only knows the invoiced price.
As a result, the margin per customer calculated by the AI is overstated all year, and the gap is concentrated on the retailers that matter most.
The check: list the terms that do not go through the invoice, and decide either to enter them in the ERP or to give the AI access to the file that holds them.
6. Purchase prices and suppliers entered as free text
On the purchasing side, flaws are less visible: the same supplier created twice, raw materials entered under different names, purchase prices recorded on receipt with varying units (kilo, carcass, crate).
An agent comparing your purchase prices with market quotations needs a comparable unit price per raw material. Without it, the comparison is impossible. We cover these uses in our article on AI applied to raw material purchasing.
The check: reconcile the list of active suppliers, and make sure each strategic raw material has a single purchasing unit.
Clean up ERP data before an AI project: the four-step method
Cleaning the whole database would be long, costly and pointless. The rule is the one Salesforce recommends to prepare your data for an AI project: start with the data the project actually needs.
Step one: list the known flaws. Bring sales administration, accounts and purchasing together for an hour. Each person names the corrections they make mentally or in a separate file: that is your starting list.
Step two: sort by the blocks opened to AI. If the agent will analyse sales, duplicate customers and units come first. If it will monitor margins, cost prices and rebates come first. The rest can wait.
Step three: fix, block by block. One person per block, a short deadline. Most corrections are done by exporting, sorting in a spreadsheet, then re-importing or entering the data in the ERP.
Step four: set a data entry rule. One simple rule per block, for example: "no item created without a family and a conversion factor". This is what stops you starting over in a year.
This work assumes you know where your data sits and who is responsible for it, which steps 2 and 3 of our data and AI governance roadmap cover.
Prepare your data for AI: summary by flaw
| Flaw | What the AI gets wrong | Simple check | Owner |
|---|---|---|---|
| Duplicate customers | Ordering pattern, revenue per retailer | Sort by name and registration number | Sales administration manager |
| Mixed units | Volumes, average price per kilo | Conversion factors filled in | Sales administration or logistics manager |
| Misclassified items | Profitability by range | Review of the "miscellaneous" family | Sales manager |
| Outdated cost prices | Margin per item, threshold alerts | Date of last update | Financial controller |
| Rebates outside the ERP | Margin per customer | List of terms outside the invoice | Sales director |
| Purchase prices in free text | Gap with the market | Single purchasing unit per raw material | Purchasing manager |
What cleaning up your data does not do
It does not replace a change of ERP. If your software cannot handle variable weight or batches, no clean-up will fill that gap. But in the vast majority of cases, the software can do it: it is the data entry habits that have drifted.
It does not require an MDM tool. Master data management software is for groups consolidating several ERPs. For an SME, a spreadsheet, a written rule and an owner per block are enough.
It is not done once and for all. Without a data entry rule and an owner, duplicates come back. This is why data entry at source matters as much as the clean-up: in our meat-cutting traceability case study, nothing goes into the ERP without a manager's approval.
Clean your ERP data: where to start
Choose the first use of AI. Sales, margins or purchasing: this choice decides which blocks to clean first.
Hold the one-hour review. Sales administration, accounts and purchasing around a table, with a single question: "What do you correct by hand?"
Name an owner and a rule per block. A name, a written data entry rule, a review date in three months.
Frequently asked questions about ERP data quality
Do you need a perfect database to connect an AI?
No. No SME has a perfect database, and waiting for one means never starting. What matters is that the data used by the first use case is reliable: sales and customers for a sales agent, cost prices for a margin agent. The rest can be corrected as you go, once the agent is in place.
How long does an ERP data clean-up take in an SME?
For three to five data blocks, allow a few weeks, alongside day-to-day work. The time mainly depends on the number of duplicates and the state of your cost prices. It is not an IT project, it is an organisational one.
Who should handle it when there is no IT department?
The people who use the data every day: sales administration for customers and orders, the financial controller for cost prices, the purchasing manager for suppliers. They already know the flaws. The business leader settles the trade-offs, for example which customer code to keep.
Should you change ERP if the data is poor?
Rarely. Poor data almost always comes from data entry habits, and it would follow the company into a new ERP. A change of ERP is justified when the software cannot handle a business need, such as variable weight or batch traceability, not when the database is untidy.
Can an AI agent spot data flaws itself?
Partly. An agent that calculates margins or ordering patterns brings inconsistencies to light: an implausible negative margin, a purchase price outside any range, a customer who seems to order twice. It can flag them. However, deciding which customer code to keep or which cost price is right remains a human decision.
How do you stop the database degrading again?
With three things: a data entry rule per block, a named owner and a regular review. The rule must be simple and verifiable, for example no item created without a family and a unit. Without an owner, the best rule is forgotten within a few months.
Going further
Data clean-up is the fourth step of our data and AI governance roadmap for food and drink SMEs. The next step is to write your usage rules, with our AI usage policy template for food and drink SMEs. To see how an agent then combines your internal data with market data, read our article on AI agents connected to several data sources.
Want to know whether your data is ready? Book 30 minutes with us: together we look at which blocks to open first and which flaws to fix before connecting an agent.
Ready to take control of your data?
Book a 30-minute demo and see what Agrolytics can do for you.
Book a demo
