Master Data Won't Tell Your Agent the Customer Went Quiet

An agent that reads a health score of 42 will act on it. Draft the outreach, open the ticket, flag the account before the pipeline review. Every step is permitted, logged, and auditable.

What is not auditable is the 42. The reasoning that produced it was never in the payload.

We wrote about that gap in August: permission is not provenance. Stibo Systems and Accenture have since published a version of the same argument.

The industry agreed on the first half

On 3 September 2026, Stibo Systems published “5 master data failures that turn reliable AI models into unreliable agents”. Stibo sells master data management and was named a Leader in this year’s Gartner Magic Quadrant for that category. Their opening claim: “Most agent errors trace back to ungoverned data, not the model itself.”

In AI at Scale, a white paper by Accenture in collaboration with Stibo Systems, the same point about autonomy (p. 10):

An AI agent acting on incorrect or inconsistent master data does not make a single error. It propagates that error systematically, at scale, before any human has the opportunity to intervene.

We would put it the same way.

What twenty-one pages of governed data covers

Across that white paper, the data an agent needs is a record about a thing. A product. A supplier. A price. A regulatory classification. Govern those, make them consistent across systems, give every value a traceable lineage, and an agent has a foundation to reason from.

It is a good foundation. It tells an agent what a customer is: which plan, which contract, which region, which renewal date, which support tier. What it does not capture is how the relationship is changing.

Master data management is not failing when it misses that. It was built to answer a different question, which is which version of a record is true across every system that holds one. That question is real and hard, and an agent that gets it wrong will act on a stale price or a duplicated customer.

An agent working on a customer relationship needs that layer and one more.

What the second layer holds

It holds two things, and the distance between them.

The first is what the customer said. “We’re evaluating other solutions.” “What’s the process for cancellation.” Language in a ticket or on a call that means something changed.

The second is what they did, or stopped doing. The champion who has not appeared in a month. The thread that had three people on it and now has one. The conversation volume that fell to under half of what it was.

Both halves can be stored. Neither can be looked up. A conversation sits perfectly well in a governed table, and an event carries a clean schema. But the thing that matters is not the row, it is the change: this champion used to reply, and now does not. Absence is not a value you retrieve. It is inferred by comparing the present against a baseline, which means it lives in the shape of the history and never in the current record.

The disagreement is the finding

Reconciliation exists to resolve conflicting values into one authoritative record. Two systems should not disagree about a price. When they do, somebody fixes it, and the fix is the point.

The Accenture and Stibo operating model extends that instinct to AI. Where a model produces conflicting outputs, they write, that “should trigger data governance review, not just model retraining” (p. 17).

For records about things, that is right, but for people it inverts. A customer who tells you the renewal looks fine, whose champion has gone quiet, is not a data quality problem. Those two facts are supposed to disagree. The disagreement is the earliest honest thing you get, and it shows up before anyone opens a ticket.

Resolve it into one authoritative value and you have deleted the signal.

What an agent can do with it

A person holding two halves that disagree gets uneasy, then goes looking. That unease is the last error check in most customer operations, and autonomy removes it.

An agent can only weigh what comes back in the response. Hand it a number and it has nothing to weigh. Hand it the conversation, the date, and what changed, and it can do the thing a bare verdict makes impossible: decline to act when the evidence underneath is thin.

Where we sit

Resonant IQ is the customer evidence layer. It reads the conversations your teams are already having across support, sales and success, and links every signal back to the conversation it came from and the date it changed. This is the fragmentation problem we built it to close: the evidence already exists, scattered across tools nobody reads together.

It reads absence too, by comparing an account against the way it used to behave.

Agents can query that layer directly, and what comes back carries the source conversation and the date alongside the answer.

None of this replaces a governed data foundation. Agents need one, it sits under the product and supplier and pricing records, and it is real infrastructure.

It just cannot see the customer going quiet.

Sources

Alison Bruford, “5 master data failures that turn reliable AI models into unreliable agents”, Stibo Systems, 3 September 2026.

Damien Fellowes, Michael Fieg and Sal Seno, AI at Scale: Why Enterprise Initiatives Stall After the Pilot, a white paper by Accenture in collaboration with Stibo Systems, 2026. Quotations from pp. 10 and 17.