Back to Insights
Enterprise AI6 days agoJustin Pennington

Your AI Doesn't Need More Data. It Needs One Definition of "Customer."

Your AI Doesn't Need More Data. It Needs One Definition of "Customer."

Ask three systems how many active customers you have and you'll often get three different numbers. The CRM counts anyone with an open opportunity. The ERP counts anyone with an invoice in the last twelve months. The marketing platform counts anyone who hasn't unsubscribed. Everyone in the room knows which number they trust, and nobody has ever written that down.

Your team handles this gracefully. They know the CRM list is inflated, that the ops team maintains a side spreadsheet of "real" accounts, and that two customers are actually the same parent company under different names. That reconciliation lives in people's heads, and it happens so fast nobody calls it work.

Then you point an AI tool at the same data and it falls apart — not because the model is weak, but because it has no access to the quiet corrections your people make all day.

Why This Breaks AI Harder Than It Breaks People

A new hire who gets a confusing answer asks someone. An AI system doesn't. It picks a row, writes a confident paragraph, and moves on. Ambiguity that a human treats as a prompt to check becomes, for a model, just another input.

This is why so many pilots stall at the demo stage. The demo works because someone hand-picked clean records. Production doesn't, because production is full of duplicate accounts, abandoned fields, status codes that meant something in 2019, and three competing notions of "active." The fix isn't a better prompt or a bigger model. It's agreement.

Pick Three Nouns, Not Thirty

A full master data program is a multi-year project most mid-market companies don't need and won't finish. You need far less than that to get value.

Choose the three nouns your business actually runs on. For most companies it's customer, order (or job, or project), and product (or service). Everything else — leads, tickets, invoices, timesheets — hangs off those three. If those three are solid, an AI assistant can answer most of the questions people actually ask it.

For each noun, you're settling four things:

  • The system of record. One system owns each field. Not "both, synced" — one. Billing address lives in the ERP. Pipeline stage lives in the CRM. Pick, then enforce.
  • The definition. What makes a customer "active"? Write the rule in a sentence an ops manager could repeat from memory.
  • The identity key. What makes two records the same entity? Email is fragile. Company name is worse. A durable ID — account number, tax ID, an internally issued key — holds up.
  • The change rule. Who can create or edit the record, and where. Most duplicate data is created by well-meaning people in the wrong screen.

That's a workshop, not a quarter. The hard part isn't technical; it's getting finance, sales, and operations to agree out loud about something they've each quietly assumed for years.

Write the Definitions Where Machines Can Read Them

A definitions document that lives in someone's drive helps nobody. Put it somewhere both people and systems can reach: a data dictionary in your Business Hub, a shared reference your integrations and AI tools pull context from, a set of views in your warehouse that encode the agreed logic.

This is the step teams skip. They settle the definition in a meeting, then let every report, dashboard, and script re-implement it independently. Six months later you're back to three answers. If the definition of "active customer" exists as a named, reusable view or field that every tool reads from, drift gets much harder.

The Duplicate Problem Is a Process Problem

You can clean duplicates with a one-time merge. They'll come back within a quarter unless you fix the entry point.

Usually the entry point is the problem: a web form that creates a CRM record without checking for an existing account, a salesperson who can't find a company spelled slightly differently, an import that runs without matching logic. Add a lookup step before creation, require the durable key on anything that touches billing, and route the ambiguous cases to a human queue rather than letting the system guess. Entity resolution sounds exotic. In practice it's a matching rule plus a place to send the maybes.

How to Know It's Working

Test it the way your team will use it. Ask the same business question three different ways — "how many customers do we have," "how many accounts bought from us this year," "who are our active clients" — and see whether you get one answer. Then ask it of your AI tool, your dashboard, and your controller. If all three agree, you have a foundation worth building on.

If they don't, you've just found the most valuable work on your AI roadmap, and it has nothing to do with AI.

Where This Usually Leads

In our experience, companies discover this problem on the way to something else — an ERP rollout, a CRM consolidation, a reporting project, a first serious AI pilot. The definitions work pays for itself regardless of which project surfaced it, because every system downstream inherits the clarity.

If your AI pilots keep producing answers nobody trusts, the model probably isn't the issue. Reach out to Infraxio and we'll help you find where your definitions disagree — and what it takes to settle them.