Agents Inherit the Data Debt Humans Were Quietly Absorbing

An agent will suppress an entire segment of high-value prospects based on a rule nobody remembers writing. That sentence comes from Margarita Savytska of Sojourn Solutions, writing in AdExchanger this month about marketing-automation stacks that no one has audited in years. She is describing a specific failure, but the mechanism behind it is general, and it reaches well past marketing.

The data layer did not get worse when you added agents. It got read faster, more literally, and by something with no instinct to pause on a value that looks wrong.

The human was the freshness check nobody budgeted for

Most B2B marketing-automation data layers were built three to five years ago, before the current wave of consent regulation. That is Savytska’s estimate, and it is a practitioner’s number rather than a research finding, so treat it as a working assumption. The point holds even if the figure is soft: these layers accumulated suppression rules, consent flags, and segment definitions faster than anyone documented why each one existed.

For years, a human sat between those rules and the decision. A campaign manager looked at a suppression list, felt that a segment was too large, and dug into why. An analyst noticed a consent flag that predated the regulation it claimed to satisfy and flagged it. Nobody wrote “reverify stale records” on a roadmap. The review happened anyway, as a side effect of humans being slow, skeptical, and occasionally annoyed by a number that did not match their intuition.

That review was doing real work. It was absorbing data debt, one hesitation at a time. When an agent takes over the decision, the hesitation disappears. The agent reads the suppression rule, applies it, and moves to the next task in the same second. The rule nobody remembers writing now executes at machine speed, against a segment worth real revenue, with no one in the loop who would have paused.

The debt was always there. The human was paying it down invisibly. Remove the human and the balance comes due all at once.

Marketing found out first, and the reason matters

Marketing is where this surfaced early, and not by accident. Marketing automation was one of the first business functions to hand real decisions to autonomous systems, and its data layer is unusually old, unusually rule-dense, and unusually exposed to regulation that changed underneath it. Consent is the sharpest edge. A suppression rule written in 2022 encodes an understanding of consent that may no longer be legal, and an agent has no way to know the rule is stale. It only knows the rule is there.

Savytska’s recommended response is concrete and worth stealing regardless of function: reverify any consent record older than eighteen months, trace every suppression rule back to the original business reason that created it, and assign explicit human ownership of the data layer the agents depend on. The eighteen-month figure is again a heuristic, not a legal threshold. The discipline underneath it is what transfers. Every function that is about to point agents at an old data layer inherits the same problem marketing hit first.

The asymmetry: reading agents versus serving agents

There are two governance problems here, and they are not the same problem seen from two angles. One is buy-side: your agents read your data and act on it. The other is sell-side: other people’s agents read your API and act on what it returns. We have argued the sell-side case before, that making your product agent-ready is itself a governance decision. The buy-side case is the mirror, and Postman drew it sharply this month.

“An agent rarely fails because it cannot think,” writes Arash Nourian for Postman. “It fails because the API schema is underspecified, the endpoint behaves inconsistently, or the data returned is incomplete, stale, or ambiguous.” Postman frames reliability as a five-layer stack: data, interface, reasoning, execution, and governance. The model everyone obsesses over sits in the middle. The failures cluster at the bottom, in the data and the interface, where the specification meets the agent.

Their heuristic is the cleanest test I have seen for whether your surface is ready: if a new engineer cannot reliably use your API from its specification alone, neither can an agent. A human engineer will Slack a colleague, guess from a field name, or infer intent from a stale example. An agent takes the specification at face value. Every piece of tribal knowledge that a human quietly supplies is a place the agent will fail, silently, and then act on the failure.

The three-layer fix converged this quarter

Three separate responses to the same underlying problem shipped or matured in the same window. Read together, they form a stack.

Ontology, so the meaning is explicit. The semantic layer stopped being a data-team preference and became a platform feature this quarter. Microsoft shipped an ontology workload in Fabric in late 2025. Databricks released Genie Ontology in June 2026, claiming 84.5% first-attempt accuracy against 52.4% for a general agent without the ontology. That figure is Databricks self-reported and measures their own benchmark, so weigh it as a vendor claim rather than an independent result. Google has its Open Knowledge Framework. The convergence is the signal here. No single number carries it: the major platforms now agree that agents need meaning stated explicitly, because they will not infer it the way a human analyst does.

Rule ownership, so debt has an address. Savytska’s prescription is the second layer. Every suppression rule, consent flag, and segment definition needs a traceable owner and a recorded reason. An agent cannot ask why a rule exists. The only defense against a rule nobody remembers is a record of why it was written, attached to a person who can decide whether it still holds.

Agent-ready specifications, so the interface does not lie. Postman’s layer. The API surface an agent consumes has to be complete and honest on its own terms, because the agent has nothing else. Lenny Pruss of Amplify Partners argues that in this shift, the primitive is the product: the well-specified building block that an agent can compose against becomes the thing of value, not the polished human interface wrapped around it. A specification that only works when a human fills the gaps is a specification that fails the moment the consumer is a machine.

Do this now

Pick one data layer that an agent already reads or is about to. A marketing suppression system, a customer data platform, an internal API your agents call. Then run three checks this week.

Trace one rule to its origin. Choose a suppression rule or consent flag and try to name the person who wrote it and the reason it exists. If you cannot, you have found the shape of your debt. Multiply by the number of rules in the system.

Read one API spec as an agent would. Hand your API specification to an engineer who has never used it and ask them to complete a real task from the spec alone, no Slack, no colleague. Where they get stuck is where your agents already fail silently.

Assign one owner. Name a human who owns the freshness and correctness of that data layer. The review that used to happen as a side effect of human slowness now has to be scheduled on purpose, because the agents removed the slowness that was doing it for free.

The agents did not break your data layer. They stopped forgiving it. The forgiveness was a person, and that person is no longer in the loop.

This analysis synthesizes AI agents are making marketing decisions on data no one has checked in years (AdExchanger, Margarita Savytska of Sojourn Solutions, July 2026), Ontology everywhere! (Hands-On Data, June 2026), How we really build production-grade AI agents (Postman, June 2026), and The primitive is the product (Amplify Partners, 2026).

Victorino Group helps teams inventory the rule ownership, ontology, and API specifications their agents depend on before the debt comes due. Let’s talk.