Seven signs your ERP data is not ready for AI

The most common thing we hear from companies that tried AI on their own ERP is some version of “the model answers beautifully, but the number is wrong”. The model is not wrong: it read what was there. The problem is that nobody had checked what was there.

A language model does not fix data; it amplifies data. If master data is duplicated, the answer comes out duplicated with convincing prose around it. That is why the first job of any AI project on company data is the most boring one: looking at the data.

These are the seven signs that show up most often in our mappings. If you recognize three or more, your company is still at the data mapping stage, not the implementation stage.

1. Free text doing the job of a structured field

The project, the salesperson or the reason for an adjustment goes into “notes”, “history” or “remarks”. An experienced human reads it and understands. A report cannot group it. An AI agent groups it wrong, with confidence.

Fix: promote the information to its own field, with a closed list of values, and migrate the history with explicit rules.

2. The same thing under different names

“Customer”, “payer” and “account” are the same entity in three modules. “Project” in finance is “job” in procurement and “cost center” in accounting. Without a single vocabulary, every question to the AI needs three translations.

Fix: a data dictionary with the canonical name, the synonyms and the system where each one appears.

3. Duplicate records

The same supplier under two tax IDs typed differently, the same product under three codes, the same employee in two cost centers. We covered the cost of this in Why your ERP master data does not match inventory.

Fix: uniqueness rules (tax ID, barcode, employee number), a deduplication routine and separation between who creates and who approves.

4. Dates that do not represent the fact

The posting date is the date someone typed it, not the date the thing happened. Accrual, issue, due, payment and posting dates get mixed. Any trend analysis, cash forecast or seasonality comes out shifted.

Fix: define, per document, which date answers which question, and require the date of the fact as a mandatory field.

5. Recurring adjustments

Every month there is a “stock adjustment”, a “balance correction” or a “reversal” of the same size. A recurring adjustment is a symptom of a missing rule: the system cannot do something the operation does every month, and someone fixes it by hand.

Fix: map the process behind the adjustment and bring it into the system, or at least record it as a written rule.

6. Nobody can explain a column

There is a field called flag_tp2 that every report uses and nobody knows what it means. Or a status “7” the team learned to ignore. Data without an owner and without a definition is what confuses a model most, because it will invent a plausible explanation.

Fix: an inventory of fields with owner, definition and examples, starting with the ones used in management reports.

7. The report that depends on one person

The monthly close only happens because one analyst knows which lines to exclude, which spreadsheet to cross-check and which account to ignore. That knowledge is in no system. When the AI tries to reproduce the report, it cannot, and when that analyst goes on vacation, neither can the company.

Fix: document the process step by step, with the exceptions, and turn it into an executable rule. It is exactly the kind of context an agent needs to operate.

The order matters

All seven signs have a fix, and none of them requires changing your ERP. What they require is order: map (know what is wrong and what it costs), design the architecture (decide what to fix in the system, what to solve with a rule and what to hand to an agent), and only then implement AI in production. Skipping stages is what produces “answers beautifully, but it is wrong”.

If you want an outside reading of your data before investing in AI, a diagnostic call with Koltren is the first step, with no strings attached.

Book a call