AI for data quality rules

Two kinds of AI help

Workshops find obvious rules (domestic suppliers need a tax number) and miss implicit ones (in plant 1000, FERT always has MRP type PD). ML rule mining finds IF-THEN patterns in existing records. LLM rule drafting reads a data profile plus policy text. One is grounded but blind to intent; the other understands intent but can hallucinate. With stewards approving, they cover each other.

Rule mining in MDG

Rule mining in MDG data quality management supports Product, Business Partner and custom objects. You pick a focus area and the IF and THEN fields; a background job proposes rules (default cap 100). Setup needs Fiori apps such as Manage Rule Mining, their OData services and authorizations, so involve Basis and security early.

An LLM rule-drafting pipeline

Principle: send the model a profile, not the data. A profile (fill rate, distinct count, top patterns per field) plus a few masked samples is usually enough. Your own script then runs each proposal against the full data and computes the real pass rate. Translate proposals into a reviewed rule library; never eval() model-written code.

You help a master data team design data quality rules for SAP MDG
(Business Partner, supplier role).
Inputs: FIELD PROFILE (JSON), POLICY EXTRACT, FIVE MASKED SAMPLE RECORDS.

Propose up to 10 rules. For each return:
- id, rule_text (one business sentence), fields_involved
- condition (IF) and check (THEN), precise and testable
- source: "profile", "policy" or "both"
- expected_pass_rate (0-1), estimated from the profile
- implementation_hint: "BRFplus decision table", "BRFplus formula" or "needs ABAP"
- risk: which legitimate records this rule could block

Use only fields in the profile. Do not invent field names.
Return valid JSON only: {"rules": [...]}

Try it: complete the principles

We send the model a ______hard errorwarningour script on the full datafull supplier tableprofileeval() of the model's codereviewed rule librarythe model's own estimate and five masked samples. The real pass rate is computed by ______hard errorwarningour script on the full datafull supplier tableprofileeval() of the model's codereviewed rule librarythe model's own estimate. Model proposals are mapped to a ______hard errorwarningour script on the full datafull supplier tableprofileeval() of the model's codereviewed rule librarythe model's own estimate before they run. Every new rule goes live as a ______hard errorwarningour script on the full datafull supplier tableprofileeval() of the model's codereviewed rule librarythe model's own estimate first.

What steward review looks like

Watch the gap between expected and actual pass rate: a big gap means the model misread the profile or there is a hidden segment.

Try it: the 18% rule

Where the rule is enforced

Three places: DQ evaluation measures the existing stock, mass processing fixes violations as governed CRs, and CR validation blocks new bad data. Order matters: evaluate, remediate, then warn, then error. Otherwise a new rule blocks every change to old, uncleaned records.

Measuring value

Track acceptance rate (below about a third, stewards stop reading), false-block rate (valid records failed, seen through overrides) and time to rule. Break acceptance down by category to see where workshops are still needed.

Try it: the rule-discovery loop

Check yourself