Master Data Expert Agent
An assessment and recommendation engine that scores your master data the way your business consumes it — Sales, Supply Chain, Manufacturing, Marketing, Finance — at the single-object scope and at the population scope. The other four agents build and run the platform; the Expert Agent tells you whether the data inside it is ready for the work that depends on it.
Most "data quality" reports talk past the business
The standard quality output reads like a database report. True, useful to a steward, and entirely opaque to the business owner whose work depends on that data.
The conversation a business owner actually wants is shaped differently:
What the business is actually asking
- "Is this SKU ready to launch on Amazon next week?"
- "How much of our finished-goods catalog is ready for the Q3 sales push?"
- "Which suppliers can I source from for the regulated chemicals program, today?"
- "What single fix would unblock the most SKUs for digital commerce?"
Those questions don't get answered by attribute-level completeness. They get answered when you score the data against what each consumer needs — what Sales needs from a SKU, what Manufacturing needs, what Marketing needs — weight those concerns by how much they matter to that consumer, and aggregate the answer. That is what the Expert Agent does.
What the agent produces, on real data
Two example assessments. The single-object one evaluates one SKU. The aggregate one rolls the same evaluation across ten thousand SKUs. Both linked in full — this is what the agent's output looks like, not a mockup.
One product, evaluated against nine functional dimensions: Sales Readiness (15%), Category Management (40%), Finance Readiness (20%), Supply Chain (18%), Manufacturing & Production (15%), Logistics (12%), Digital Commerce (12%), Search Marketing (8%). Each dimension has weighted sub-metrics with percentage scores, specific data values, and gap callouts.
Excerpt — the agent doesn't just emit a number, it puts the values on the table:
The same dimensional framework, rolled up across the whole finished-goods population. Adds two things you can't get from a single object: systemic gaps (which sub-components are dragging which dimensions down) and priority actions ranked by how many SKUs each one unblocks.
Each action is ranked by how many records it unblocks and how many dimension points it adds. The business owner sees what to fund next, not a list of nulls.
Three jobs, one agent
Assess a single object
One SKU, one supplier, one customer, one material. The agent evaluates it against the consumer-domain framework, produces a weighted score, drills into each dimension's components, calls out specific gaps, and recommends fixes ranked by which consumer the fix unblocks. The single-SKU example above is the artifact.
Assess a population
All SKUs in a category, all suppliers in a region, all customers above a credit threshold. The same framework rolls up. Systemic gaps surface. Priority actions are ranked by impact — how many records they unblock and how many dimension points they add. The aggregate example above is the artifact.
SME on call
Both kinds of assessment ground a chat surface where business and MDM teams ask questions against the live data, the lineage, and the assessments. The questions look like this:
- Procurement lead time is marked "Estimated" and last updated 6 months ago — the rest of the category has confirmed values updated within 90 days.
- No secondary supplier is configured — 78% of the category has at least one.
- SEO keyword set is at 4 entries — category median is 12.
The dimensional framework
The framework is metadata, not code — same as everything else in ZMDM. Each domain ships with a default framework that you tune to your business. The finished-goods framework that produced the two example assessments looks like this.
Category Management
Primary category, sub-category, product family, hierarchy depth, classification consistency, replacement / substitute links.
Finance Readiness
Standard cost, BOM cost, transfer price, regional pricing, tax classification, GL account assignment.
Supply Chain Readiness
Lead times (procurement / manufacturing / total), supplier coverage (primary / secondary / emergency), inventory policy, MOQ, ROP, safety stock.
Sales Readiness
List price, MSRP, channel availability, regional restrictions, launch date, lifecycle status, description quality.
Manufacturing & Production
BOM completeness, routing, work centers, capacity, batch / lot policy, regulatory certifications.
Logistics Readiness
Dimensions, weight, pallet config, hazmat class, storage requirements, handling instructions.
Digital Commerce
Images, descriptions, attribute coverage for marketplace feeds, content localization, channel-specific overrides.
Search Marketing
SEO keywords, search synonyms, A+ content, structured data, category-relevance scores.
The weights are defaults the business tunes. A regulated pharmaceutical company would raise Manufacturing & Production and Logistics; a digital-first retailer would raise Digital Commerce and Search Marketing.
Each domain has its own framework. Supplier ships with Procurement Readiness, Compliance Readiness, Sourcing Readiness, Risk & Financial Health, Relationship Quality. Customer ships with Credit Readiness, Sales Readiness, Compliance, Service Readiness. Material, Location, Employee — each gets a framework tuned to its consumers. The framework is the contract between the master data and the business that consumes it.
How the agent gets to a score
The assessment is not a black-box LLM hallucination. It is a deterministic rollup with an LLM commentary layer on top.
-
Record-level signals
Field-level completeness, validation pass/fail, freshness, distribution-fit, duplicate confidence, lineage events — produced by the Quality Agent and the workflow engine.
-
Component scores
Each component (e.g. "Lead Time Management") is a deterministic rule combining a handful of record-level signals: required fields present, values within plausible ranges, last-updated recency, source-of-record confidence.
-
Dimension scores
Weighted average of component scores, with dimension-specific bonuses for cross-component consistency.
-
Overall score
Weighted average of dimension scores.
-
Narrative
An LLM step grounded on the scored output writes the bulleted "key insights", the recommended actions, and the conversational answers. It does not compute scores — the numbers are deterministic and citable.
The split is deliberate. The number you act on is never the one the LLM made up.
30 minutes, end to end
From the top-10 SKUs through a category aggregate, to a routed integration fix and a steward task — in the time of a standard meeting.
What the Expert Agent doesn't do (yet)
Honest scope. The first two are deliberate scope decisions; the second two are short roadmap items.
- Historical trending The two example assessments are point-in-time. Time-series views of "Supply Chain readiness over the last six quarters" are on the roadmap, gated on the lineage store landing field-level history.
- Forecasting The agent scores what is, not what will be. It does not predict that lead times are about to drift; it tells you they already have.
- Dollar-impact ranking Priority actions are ranked by SKU count and dimension-point impact. Dollar-weighted ranking (revenue at risk, COGS impact) requires consuming Sales and Finance data the agent doesn't yet read.
- Cross-domain assessments A SKU's readiness depends on its supplier's readiness, which depends on that supplier's compliance data. Cross-domain rollups are designed but not yet shipping.
If any of these are critical for your first deployment, tell us before we start.
How it fits with the rest of ZMDM
The Expert Agent is the fifth of five — the one that tells the business whether the whole thing is ready for the work that depends on it.
Design
Canonical model + lifecycle workflows
Architect
Workflow templates + integration wiring
Quality
Continuous perception + scored findings
Steward
Activity execution alongside human stewards
Expert
Consumer-weighted assessment + recommendation
- It consumes record-level signals from the Quality Agent — completeness, validation, duplicates, freshness, lineage events.
- It reads the canonical model from the Design Agent — knows which attributes are mandatory, which valuesets exist, which fields are display-required vs. operational-required.
- It reads workflow history from the Architect Agent's output — lifecycle events, approval timestamps, who-did-what.
- It feeds the Steward Agent — priority actions become workflow assignments. High-confidence fixes get agent-completed; nuanced ones escalate.
- It speaks MCP — any compliant agent (yours, a vendor's, a partner's) can ask the Expert Agent for the readiness of a SKU or a population and get back a cited, scored answer. Master data becomes the AI context layer the rest of your AI surface grounds itself in.
Start Your Success Story
Join the growing list of manufacturers who have transformed their master data management with ZMDM.