Skip to content

Playbook

Master data lifecycle

Request to retirement, governed. Master data is the least visible process in the business and the one that decides whether anything built on top of it — reporting, automation, agents — can be trusted.

  • SAP MDG
  • SAP Master Data Management
  • Odoo

The process

What the standard flow looks like.

  1. 01

    Request

    A change is requested through a form that knows what it is asking for, not by someone with create access going straight into the system.

  2. 02

    Validate

    Rules check completeness, format and plausibility at the point of entry, where the person still has the context to fix it.

  3. 03

    Enrich

    Attributes are added by the functions that own them — finance its account assignments, logistics its dimensions — rather than guessed by the requester.

  4. 04

    Approve

    A workflow routes to the owners of each affected area. Approval is recorded against the change, which is what makes the audit possible.

  5. 05

    Distribute

    The approved record is replicated to every consuming system, with key mapping so each one can still recognise it.

  6. 06

    Govern

    Quality is measured continuously against the same rules, and drift is remediated. Governance is a process, not a project.

Failure modes

Where we expect to find it broken.

Master data is created directly in the ERP
Anyone with create access is a governance process of one. The data is not wrong on day one; it diverges steadily and nobody can say when.
Every system has its own version of the customer
CRM, ERP and the warehouse each hold a record, none is the master, and reconciliation is a quarterly spreadsheet exercise nobody enjoys.
Duplicates are found by reporting, not prevented at entry
The same supplier exists three times with different spellings. Spend analysis understates concentration and the negotiation is weaker for it.
AI is deployed on top of ungoverned data
The agent reads the master data it is given and answers confidently from it. Bad master data does not produce obvious errors — it produces plausible ones.

Scope

What is standard, and what is the actual work.

The distinction matters commercially. Paying to build what the platform already does is the most common way an ERP budget is wasted.

Out of the box

  • Data models for material, business partner and finance objects
  • Change request types with workflow and approval by domain
  • Validation and derivation rules at entry
  • Replication to downstream systems with key mapping
  • Data quality rules with scoring and remediation lists
  • Full audit of who changed what, when and under which request

Configured or built

  • The data model itself: which attributes exist, which are mandatory, and for whom
  • Change request types and the workflow behind each one
  • Business rules — validation, derivation and the thresholds that make something an exception
  • Replication models per consuming system, and the key mapping between them
  • Duplicate-check logic, which is the hardest part to get right and the most valuable
  • The quality rules you will actually act on, as opposed to all the ones you could measure

The agent layer

Where an agent earns its place.

At the edges, where documents arrive unstructured and matching needs judgement. The transaction flow itself stays deterministic.

Duplicate detection
Compares a new request against existing records on name, address, tax identifier and bank detail, catching the near-matches that exact-match logic is built to miss.
Attribute enrichment
Proposes classification, dimensions and compliance attributes from supplier catalogues and documents, for a data steward to accept or reject.
Rule-violation triage
Groups quality failures by likely root cause rather than by record, so remediation fixes the source instead of the same symptom repeatedly.

Outcomes

What changes.

  • Before

    Anyone with create access can make a master record.

    After

    Every record arrives through a request that was validated and approved by the people who own the consequences.

  • Before

    Duplicates are found in a report, after the spend has happened.

    After

    Near-matches are surfaced at the point of request, before the record exists.

  • Before

    Nobody can say whether the master data is good enough for automation.

    After

    Quality is scored against stated rules, so the answer is a number with a trend rather than an opinion.

Questions

What people ask.

Do we need MDG, or will discipline do?
Discipline works while one team creates records in one system. It stops working at the point where several systems consume the same record and each can change it. That is the threshold, not company size.
Should we fix master data before or after the ERP project?
During. Migration is the only moment when the business is already looking at every record. Deferring it means migrating the problem and paying to fix it twice.
Why does this matter for AI?
An agent cannot tell that a record is wrong. It will answer from whatever it is given, fluently. Governance is what makes the difference between an assistant that is useful and one that is confidently misleading.

Recognise any of this in your own process?

Book a call