
Your Database Is the Real Enterprise Asset: Why AI Needs ERP Data
Most companies do not have an AI problem. They have a fragmented-data problem. Sales lives in one tool, inventory in a spreadsheet, purchasing in email and finance in accounting software. Then someone asks an AI assistant, “Which customers will be affected if this shipment is delayed?” and the assistant, however clever, cannot answer reliably, because the answer is scattered across four places that disagree.
This is Part 3 of the series on why AI will not replace proper ERP. Part 2 showed that processes need control as well as intelligence. This article is about the asset underneath: the data.
What is the real enterprise asset in an AI era?
The real asset is the business’s own clean, connected, historical operating data: one customer, one product, one supplier, one employee and one asset, with every transaction linked to them. Models are rented and improve every quarter. Your data model is yours, and it is hard to rebuild.
Anyone can subscribe to a capable model this month. What no competitor can subscribe to is your twelve years of trusted pricing history, supplier performance, customer payment behaviour and stock movements, all tied to consistent records. That is what lets an AI feature say something specific and correct about your company instead of something generic and plausible.
What does a sound data foundation look like?
- One identity per thing. One customer record, not “ABC Pte Ltd”, “ABC Private Limited” and “A.B.C.” in three systems.
- Shared definitions. Sales, procurement, inventory, finance and service agree on what an active customer, a net price or an available unit means.
- Linked transactions. A sales order connects to the customer, price, stock, delivery, invoice, payment and margin without anyone stitching it together by hand.
- Clean history. Past transactions are complete and consistent enough to support reporting, forecasting and machine learning.
- Controlled change. When a record changes, the change flows to every process that uses it and is logged.
Why does AI struggle without this foundation?
Because an AI system can only reason over what it can reach, and it cannot tell whether what it reached is right. Poor quality, silos and inconsistent formats do not make AI cautious. They make it confidently wrong.
Gartner has put numbers on the risk. In February 2025 it predicted that, through 2026, organisations will abandon 60 percent of AI projects that are unsupported by AI-ready data. Its definition is useful: the data must be representative of the use case, including every pattern, error and outlier the model will meet. Data that has unresolved issues is not ready for AI, whatever the vendor demonstration showed.
This is the same lesson we described in The Resource Master Audit, where ERP rollouts fail at the data layer rather than the software layer. AI projects fail in the same place for the same reason.
How does ERP become the single source of operational truth?
An ERP stores master data once and makes every process refer to it. A sales order links to the customer, the price list, the stock position, the delivery, the invoice, the payment and the margin. A change in one record propagates to the processes that depend on it. This is unglamorous, and it is exactly what an AI feature needs underneath it.
| Question a manager asks | AI can | ERP must |
|---|---|---|
| Which customers are hit if this shipment is late? | Phrase the answer, rank by importance, draft the notifications | Hold the link from shipment to purchase order to sales orders to customers |
| What is our real margin on this product? | Explain the drivers in plain language | Provide landed cost, discounts, returns and rebates from posted transactions |
| Who are our best suppliers? | Summarise patterns and suggest actions | Record on-time delivery, quality rejections and price history per supplier |
| Will we run out of stock? | Forecast demand using history and seasonality | Maintain accurate on-hand, reserved and incoming quantities |
A scenario: the delayed shipment
A components importer learns that a container will arrive nine days late. In a fragmented business, this is a day of detective work. The buyer checks email for what was in the container. The warehouse checks a spreadsheet for what was reserved. Sales scrolls through chat messages to remember who was promised what. Finance has no idea which invoices depend on it.
In a business running on a connected ERP, the purchase order lists the container’s contents, the reservations show which sales orders depend on those items, and the sales orders show the customers and promised dates. An AI assistant on top can then do what AI is good at: draft a tailored delay notice for each affected customer, propose substitute stock, and rank the customers by revenue at risk. The AI is useful because the data is connected. Without the data, the same assistant would invent a plausible list.
What is master data, and who should own it?
Master data is the slowly changing information that every transaction refers to: who the customer is, what the product is, who supplies it, what it costs, which account it posts to. Transactions come and go every minute. Master data is the frame they hang on. If the frame is wrong, every transaction inherits the error.
| Master data | Typical content | Natural owner |
|---|---|---|
| Customers | Legal name, registration number, billing and delivery addresses, credit terms, tax status | Sales operations with finance |
| Products | Code, description, unit of measure, cost, price, tax class, category | Product or inventory manager |
| Suppliers | Legal name, bank details, payment terms, lead time, approval status | Procurement with finance |
| Employees | Role, department, approval limits, access rights | HR with IT |
| Chart of accounts | Accounts, tax codes, cost centres, analytic tags | Finance controller |
The point of naming an owner is accountability. When a supplier’s bank details change, one named person should verify the change and approve it. This is a classic fraud route, and an AI tool that can edit master data without that gate is a risk, not a convenience. AI can suggest that two customer records look like duplicates. A person who owns the data decides whether to merge them, and the ERP enforces the decision.
How can you tell whether your data is in trouble?
You do not need a consultant to get a first reading. Ask your team to check the following, using a sample of a hundred records.
- Duplicates: how many customers, suppliers or products appear more than once under slightly different names?
- Missing fields: how many records lack a tax status, a payment term, a unit of measure or a valid address?
- Inconsistent units: is the same item counted in cartons in one place and pieces in another?
- Orphan transactions: how many orders, bills or payments cannot be traced to a valid customer, supplier or product?
- Stale values: when were prices, lead times and stock counts last verified?
- Shadow copies: how many spreadsheets hold an unofficial version of the same data?
If more than a handful of records fail any one of these tests, an AI model trained on or reading from that data will reproduce the problem with great confidence. This is the practical meaning of Gartner’s warning that data must be representative of the use case, including its errors and outliers, before it is ready for AI.
How do you prepare your data for AI, step by step?
- Pick one process. Do not try to clean the whole company. Choose order to cash or procure to pay.
- List the master data it touches. Customers, products, prices, suppliers, accounts.
- Deduplicate and standardise. Agree one naming convention, one set of units and one set of categories.
- Fill the required fields. Make tax status, payment terms and addresses mandatory in the ERP so that new records cannot skip them.
- Load clean opening balances and history. Stock, receivables, payables and at least twelve to twenty-four months of transactions if you want trend analysis.
- Lock the doors. Give edit rights for master data to a small number of named people, and log every change.
Step four is the one most often missed. Cleaning data once is easy. Keeping it clean requires the system to refuse bad entries at the point of creation, which is exactly what an ERP’s required fields, validation rules and access rights do. In Odoo, for example, all apps read from one shared contact record and one product record, so the customer you update in sales is the same customer that accounting invoices.
A second scenario: the margin question
The managing director asks an AI assistant which product line has the best margin. In a fragmented company, the assistant is given a sales export from one tool and a cost list from another. The cost list holds the supplier’s invoice price but not freight, duty, or the returns that came back last quarter. The assistant produces a tidy ranking, and the company shifts its promotion budget towards a product that actually loses money after landed cost and returns.
In a connected ERP, landed cost is attached to each receipt, returns are linked to the original sale, discounts are recorded on the order line, and the margin report is built from posted transactions. The same AI assistant, asked the same question, now explains a correct ranking and can say why. The model did not get smarter. The data got truthful.
What breaks without ERP
- The same customer exists in several spellings, so history is split and credit is misjudged.
- Inventory in spreadsheets is stale by the time anyone reads it.
- Purchasing in email cannot be queried, so nothing links supply to demand.
- Finance numbers and operations numbers never reconcile, and every meeting begins with an argument about whose figure is right.
- AI answers inherit every one of these inconsistencies and present them fluently.
What AI cannot own
- Master-data ownership: someone must be accountable for each record being right.
- The definition of a customer, a product or a margin.
- The decision to merge or retire a record.
- The reconciliation between operational and financial truth.
Takeaway: AI is only as useful as the business data model it can access. In Part 4 we move from data to process design, and ask what actually causes low productivity.
Frequently asked questions
What does AI-ready data mean for an SME?
It means clean, consistent, connected records that represent your real operations, including the awkward cases. We cover the practical meaning in What Does AI-Ready Actually Mean for a Singapore SME?
Can AI clean up my messy data for me?
It can help find duplicates and suggest corrections, which is a good use. But a person must own the decision about which record is correct, and the ERP must enforce the result so the mess does not return.
Is a data warehouse enough instead of an ERP?
A warehouse can combine data for reporting, but it does not run the transactions or stop bad data entering at the source. Fixing quality where the data is created is cheaper than repairing it downstream.
How long does data cleanup take before an ERP rollout?
It depends on volume and discipline, but it is usually the longest part of the project and the one most often underestimated. Start with customers, suppliers, products and opening balances.
Who should own master data in an SME?
Assign each data domain to one named role, not to IT in general. Customers to sales operations with finance, suppliers to procurement with finance, products to the inventory or product manager. Owners approve changes and answer for quality.
How much historical data does AI need?
It depends on the use case. Trend and forecasting features usually need at least a year or two of consistent transactions, ideally covering a full seasonal cycle. Quality matters more than volume: a year of clean data beats five years of inconsistent data.
A question for you: if a customer called right now and asked what you owe them, what you have promised them and what they owe you, how many systems would you have to open?
Sources
By Willie Loo, Managing Director, Mxgsoft Pte Ltd and Moxogo Pte Ltd, Singapore. Part 3 of 9 in the series Why AI Will Not Replace Proper ERP.
Two companion articles extend this series: why agentic tools cannot replace enterprise data governance and audit trails, and how personal AI workflows differ from cross-department ERP control.
Next step: choose one entity, usually customers or products, and count how many places hold a version of it. That number is your fragmentation score.


