We want to connect an AI model to our ERP and order history to automate customer reorder prompts, but our customer master records are riddled with duplicate accounts and inconsistent part numbers. What data hygiene protocol must we complete before launching this?

Question

Tyler Smith · Accepted Answer

Feeding unstructured, dirty ERP data into an AI model will only accelerate bad decisions. If your customer records contain duplicate accounts, fragmented purchase orders, and inconsistent part numbers, the model will generate hallucinated reorder schedules that erode client trust.

Before you touch any machine learning tools, execute a four-step data hygiene sprint:

- Consolidate customer entity records into a single source of truth. Merge accounts that share common tax IDs, physical delivery addresses, or billing domains. Eliminate legacy dummy accounts used by former employees for testing.
- Standardize your product taxonomy. Map every internal part number to a single parent SKU and enforce consistent units of measure across all historical transactions. If the same widget is tracked as each in 2021 and box of ten in 2024, the model will miscalculate replenishment cycles every time.
- Purge anomaly transactions from your training baseline. Remove one-off bulk liquidation purchases, canceled orders, and supplier stock-outs that skew standard reorder frequency.
- Establish validation rules on future data entry. Lock fields so sales and warehouse staff cannot bypass standard naming conventions.

Assign this data cleanup as a quarterly Rock to the seat on your Accountability Chart that owns system integrity. Only when historical transaction records reflect actual buying patterns should you deploy an automated agent to analyze ordering cadence and draft reorder reminders.