Our ERP data is a mess. Do we have to clean it up first?

• 01Two different jobs get confused under the word "clean". Shape problems — inconsistent date formats, trailing whitespace, three spellings of the same unit of measure, codes that mean nothing outside the module that wrote them — are fixed in the pipeline, on every run, permanently. Missing facts and wrong facts are fixed at source, because no transformation step invents a value that was never recorded.
• 02Sort the mess into three buckets before anyone quotes the work. Formatting and type inconsistency: handled in transformation, no decisions needed from you. Duplicate and near-duplicate master records: handled by a matching rule, but someone of yours has to approve which record wins. Genuine disagreement between systems — the CRM says booked, the ERP says recognised — is not a data-quality defect at all, and no tool resolves it.
• 03The real blocker is definitional, not technical. Expect to settle, in writing: which date a sale belongs to, whether intercompany transactions are netted out, which cost basis margin uses, whether cancelled orders stay in the denominator, and which company codes or plants roll into "the business". That is a handful of two-hour sessions with people who can decide, not a cleansing programme — and it is the work that sets whether the first module lands in four to six weeks or drifts.
• 04Extraction surface determines the shape of the build, so establish it early. SAP ECC is normally read through its extractor layer or a replica rather than queried against production; Oracle EBS commonly through a read replica; Dynamics 365 Business Central typically through its OData/API surface or Dataverse; an on-premises SQL-backed ERP through an on-premises data gateway, which makes an outbound connection, so no inbound firewall port is opened. After the first full load, incremental refresh keys off a row-level watermark — a dependable modified-date or change-tracking column — so only changed records move.
• 05Reporting is a better data-quality instrument than an audit. Once the pipeline runs, record counts and control totals get reconciled against a report the ERP already produces, and the gaps become visible and specific: orders with no customer region, one plant that stopped posting scrap codes in March, a supplier that exists four times. Fixing those with named rows in front of you is a different task from cleaning a system in the abstract.
• 06The honest limit. If a table carries no dependable primary key and no modified-date or change-tracking column, there is nothing for watermark-based incremental refresh to key off, and the options narrow to a full reload each cycle or a change at source. And where the figure you want was never captured — blank scrap reason codes, promised dates typed into a free-text note — the fix is a process change in the ERP, and reporting is worth building after it, not before.

For the full explainer on this topic rather than this specific question, see the detailed guide on the blog.