⌘K
Accelerator LibraryERP Master Data Quality Agent

ERP Master Data Quality Agent

Version v1.4 · Production · Deployed at 12 organisations — continuously scans ERP master data for completeness gaps, behavioural anomalies and duplicate entities, and routes every finding to an accountable owner with a full audit trail. Read-only against your ERP.

ProductionCross-cutting · ERP and Master DataDPDP-aligned Read-only against source ERP Featured

What this pack does

Enterprise ERP master data decays quietly. Vendors are created twice under slightly different legal names, statutory fields are left blank during month-end pressure, and dealer classifications drift from the way the business actually operates. None of it breaks a system — it just quietly distorts spend analytics, tax reconciliation and channel reporting.

This accelerator deploys a scoped Master Data Quality workspace that scans customer, vendor and material master data on a schedule, produces findings from three independent detection methods, and routes each one to a named owner with a validation workflow. Findings close only when the correction is confirmed in the source system's change log — not when someone clicks "done".

Typical first-scan outcome at a mid-size ERP estate: 200–400 rule failures, 15–30 behavioural anomalies and 5–15 duplicate clusters, with 60–70% of findings owned by three or fewer people.

How it works — six agents, three detection methods

1
Extraction Agent
Reads master data from the lake copy or a read-only ERP OData service. No writes, ever.
2
Rules Engine
84 deterministic completeness, format, referential and statutory rules. Zero token cost.
3
Anomaly Agent
Profiles each entity against its own 24-month history; flags behavioural deviations with a confidence score.
4
Entity Resolution Agent
Blocks, embeds and clusters candidate duplicates; a frontier model adjudicates only borderline pairs.
5
Summarisation Agent
Writes a plain-language narrative and business impact for each finding, for non-technical owners.
6
Routing Agent
Assigns each finding to the accountable owner and forum defined in the ownership routing matrix.
RULE
Deterministic. No model, no tokens, no ambiguity.
ML ANOMALY
Behavioural deviation vs the entity's own history, with confidence.
ENTITY MATCH
Similarity clustering with LLM adjudication on borderline pairs only.

What this pack deliberately does not do

  • It does not write to, correct or delete anything in your ERP. Corrections are made by your team in the system of record.
  • It does not auto-merge duplicate records. Clusters are recommendations that a steward confirms or rejects.
  • It does not replace your MDM tool. It tells you what is wrong, who owns it, and whether it was fixed.
  • It does not act on anomalies without human validation. Every ML finding carries a confidence score and a validation step.

Ownership routing template

The pack ships with a routing template. During deployment it is mapped to your own owners and governance forums; it stays editable in Settings → Ownership routing.

DomainFinding typeAccountable ownerReview forum
CustomerCompleteness gap (RULE)Kavita IyerWeekly dealer data huddle — Thursday
CustomerBehavioural anomaly (ML)Sandeep KrishnanMonthly master data governance forum — second Tuesday
CustomerDuplicate cluster (ENTITY)Kavita IyerMonthly master data governance forum — second Tuesday
VendorTax / statutory field gap (RULE)Anjali DesaiPayables control review — fortnightly
VendorDuplicate cluster (ENTITY)Ramesh PillaiMonthly master data governance forum — second Tuesday
MaterialClassification / UoM gap (RULE)Vinod RathiMaterials governance council — monthly

Prerequisites

  • A lake copy of ERP master data, or a read-only ERP OData / RFC service account.
  • Extract of the last 24 months of transactional behaviour per entity (for anomaly baselines).
  • A named accountable owner per master data domain.
  • One governance forum where findings are reviewed on a cadence.

Deployment configuration

Source systemPinnacle SAP S/4 HANA (India)
Secondary sourceTCS BaNCS (Apex Capital)
Read pathLakehouse copy (Fabric), read-only
Domains in scopeCustomer, Vendor, Material
Scan cadenceDaily 06:00 IST + delta on change events
Monthly budget cap₹45,000 (agent fleet policy)
Write accessNone — read-only by design

Agents installed

Extraction Agent
Read-only lake / OData pull
Deterministic
Rules Engine
84 completeness, format, referential rules
Deterministic
Anomaly Agent
Per-entity behavioural profiling
Model
Entity Resolution Agent
Duplicate clustering + LLM adjudication
Model
Summarisation Agent
Plain-language finding narratives
Model
Routing Agent
Owner assignment per routing matrix
Model

All six register in AI Context → Fleet Governance with budget caps and kill-switches.

Outcome at 12 deployments

Median first-scan findings286
Findings closed in 90 days71%
Duplicate vendor spend recovered₹1.2Cr median
Setup effort saved68 person-days