Skip to content

MDM

Your master data is why the AI pilot failed

The model was fine. The problem is that four systems disagree about who your customer is, and nobody owns the answer.

18 November 2026 · 7 min

A pattern we see often enough to predict: a company runs an AI pilot, the demo is good, the rollout is quietly abandoned, and the post-mortem blames the model. The model was almost never the problem.

What actually went wrong

The pilot ran on a curated extract — a clean spreadsheet somebody prepared, or a single system's view. Production ran on the real estate: a CRM, an ERP, a billing system and a spreadsheet that one team swears by, each with its own idea of what a customer is.

An agent reading across that does not fail loudly. It produces an answer, with confidence, and the answer is wrong in ways that take a quarter to surface.

Master data management, without the acronym

Master data is the small set of records everything else refers to: customers, vendors, products, materials, employees, the chart of accounts. Managing it means three unglamorous things.

  • Definition: agreeing what a customer actually is. One legal entity? One billing relationship? One site? Most organisations have never written this down, and different teams answer differently
  • Ownership: naming the person who approves a change to a vendor record, and the rule they apply
  • Distribution: making sure the agreed version reaches every system that needs it, rather than each system drifting on its own

None of that is software. The software comes after, and it is the easy part.

Cleanup is not governance

If you only cleanse, you will be back where you started within about a year.

The most common waste in this area is a one-off deduplication project with no governance behind it. Records are merged, a report is produced, everyone is pleased, and the same duplicates reappear as soon as people resume creating records the way they always did.

Governance is the boring part that makes the cleanup stick: the approval workflow, the validation rules, the survivorship logic written down so the result is reproducible rather than one analyst's judgement.

What good looks like

A simple test: two teams pull the same number from two different systems and get the same answer, without anyone reconciling by hand. If that is not true today, it is the thing to fix before any agent is allowed near the data.

Do it during the ERP project, not after

An ERP migration already forces these decisions. You cannot load a customer master without deciding what a customer is. Doing the governance work at that moment costs a fraction of retrofitting it later, when every downstream system has already encoded its own assumption.

It is also the least glamorous line on the project plan and the first one cut. That decision is usually visible two years later, in the form of a reporting dispute nobody can settle.

Facing this in your own business?

Book a 30-minute call