ERP Data Quality

Analytics, automation, and AI inherit ERP data quality. When item master, units of measure, customer hierarchy, or open transaction hygiene are weak, downstream systems produce confident wrong answers.

:::info Portfolio architecture example

This page describes sanitized data-quality patterns for mid-market ERP transformation. Examples use representative distribution and inventory scenarios—not confidential client or employer records.

:::

:::note Research-informed

Data-quality and adoption discipline here are informed by doctoral ERP research and later practice—not claimed as direct dissertation excerpts. See Research Overview.

:::

Definition of quality (operational)

ERP data is fit for use when a downstream consumer can answer:

1. What entity does this key refer to?
2. What event happened, when, and in what quantity?
3. What evidence exists if the record is challenged?
4. What is known vs. assumed vs. missing?

Data domains and typical failure modes

Domain	What breaks	Downstream impact
Item master	Duplicate IDs, wrong UOM, inactive items sold	Pick errors, bad margin, bad forecasts
Customer / ship-to	Stale addresses, merged accounts	Revenue attribution, credit, service levels
Vendor	Lead time defaults, bad payment terms	Planning noise, accrual mismatch
BOM / kit (if used)	Phantom components, rounding	Allocation and cost explosion errors
Open transactions	Unposted batches, stuck integrations	False ATP, month-end surprises

Pattern 1: Master data ownership matrix

Every critical field has a named owner, source of truth, and change workflow.

Field group → Owner role → Authoritative system → Change ticket / approval → Effective date

Without ownership, "data cleanup" becomes a recurring hero project.

Pattern 2: Reason codes as quality instrumentation

Treat reason codes on adjustments, overrides, and manual entries as quality telemetry, not annoyance fields.

Portfolio examples:

• CYCLE-COUNT-ADJ — physical vs. system variance
• UOM-CONVERSION-ERR — caught at receipt or pick
• INTEGRATION-REPLAY — idempotent repost from WMS/shipping
• MASTER-DATA-FIX — corrected item attribute with approver

Trend reason codes before blaming operators or integrations.

Pattern 3: Key integrity rules before integration scale

Minimum public-safe checklist:

• Unique item identifiers with documented alias policy (UPC, vendor SKU, legacy ID)
• UOM conversion table tested with edge quantities
• Status flags enforced (inactive, discontinued, non-stock)
• Customer and ship-to merge/archive policy
• Duplicate detection report on a schedule

Pattern 4: Evidence bundle for exceptions

When a number is disputed, an evidence bundle should exist:

• Source transaction IDs and timestamps
• Operator or integration message ID
• Related document (PO, receipt, pick, shipment)
• Approval record for manual override
• Reconciliation note if timing difference explains variance

This discipline supports finance, operations, and future AI governance—not just audit.

Pattern 5: Data-quality gates for AI readiness

Before adding models or agents on ERP data, assess:

Gate	Question
Decision clarity	What decision improves if the signal is right?
Label truth	Can outcomes be verified against source transactions?
Coverage	What percentage of volume is excluded by bad keys or missing fields?
Latency	Is data fresh enough for the decision window?
Human oversight	Who intervenes when the signal conflicts with operations?

The AI Readiness Diagnostic formalizes several of these checks in a runnable public tool.

Measurement examples (not production claims)

Portfolio diagnostic metrics teams often track during transformation:

• Master data exception rate by domain
• Percent of shipments with complete scan chain
• Open integration error queue age
• Repeat adjustment rate by item / location
• Percent of revenue through items passing quality rules

Anti-patterns

• "Clean data later" after AI pilot launch
• Spreadsheet shadow systems without reconciliation path
• One-time purge without ownership or prevention rules
• Reporting KPIs on uncorrected historical keys

Related work

• Supply Chain and Warehouse
• FIFO and Inventory Costing
• ERP AI Delay Risk — ML requires explicit label and feature hygiene