AI agents cannot fix fragmented commercial data
Agency turns ambiguity into action. An agent working on unresolved identity and authority does not clean the data; it repeats the confusion faster and further.
Independently created. Contains no employer or client implementation detail, internal names or figures.
An agent can act across systems faster and more often than a person. That does not resolve fragmented commercial data. It turns the data’s ambiguities into actions, at a scale where they are harder to notice and harder to undo.
The hope is understandable
Many commercial organizations carry years of fragmented data: the same customer in several systems, accounts that mean different things to different teams, prices and statuses that disagree. Cleaning it has been on the roadmap for a long time. AI looks like a way around the problem. If a model can read messy text, compare records and fill gaps, perhaps the agent can work on top of the fragmentation instead of waiting for it to be fixed.
Part of that is true. Models are good at reading inconsistent inputs, suggesting matches and explaining differences. Used that way, they can make data work faster. The difficulty starts when reading becomes acting: when the agent does not only suggest that two accounts are the same customer, but creates the opportunity, sends the message or applies the price on the strength of that reading.
Agency turns ambiguity into action
A person who finds two similar accounts usually hesitates. They check the address, ask a colleague, or leave the record alone until someone confirms. That hesitation is slow and informal, but it contains the ambiguity. Nothing irreversible happens while the question is open.
An agent does not hesitate in the same way. It resolves the ambiguity, because resolving is what it was built to do, and then acts on the resolution. The same unresolved identity that was a reporting nuisance becomes duplicated outreach, a quote attached to the wrong legal entity, or a commitment made against a customer record that should have been merged last year.
This is the argument in one line: agency does not reduce the cost of ambiguity. It moves the cost from the reporting layer to the transaction.
Retrieval is not authority
An agent that retrieves a value has found something. It has not established whether that value may be trusted, changed, or used for a consequential decision. In fragmented data, the same question often has several answers: a price in the CRM and another in the ERP, a customer status that is active in one system and blocked in another, a contact who left the company but still appears as the decision-maker.
A person who knows the organization knows which source to believe for which question, usually without being able to explain the rule. The agent needs the rule to be explicit. Which system owns price? Which one owns credit status? Which one decides whether a customer exists as a legal entity? If those answers live only in people’s heads, the agent will pick a source, and it will pick consistently, which is worse than picking randomly when it is wrong.
Confidence scores do not close this gap. A confidence score describes the model’s certainty about its own reading. It says nothing about whether the source it read was the right one. A highly confident match to the wrong legal entity is still a match to the wrong legal entity.
Scale changes the failure mode
Human errors in commercial data are usually local and visible. One seller creates a duplicate; someone notices; the cleanup is annoying but bounded. Agent errors are systematic. The same rule, applied to the same ambiguity, produces the same mistake hundreds of times before anyone looks.
The downstream systems make it worse. They tend to trust what arrives from upstream, so a wrong record becomes a wrong order line, a wrong forecast and a wrong segment. By the time the error surfaces, it has been copied into places that were designed to be read, not corrected.
Build the control plane before widening agency
The practical answer is not to avoid agents. It is to make the parts of the data model that agents depend on explicit before giving them more authority. Canonical identity, with a cross-reference between the commercial account, the legal entity and the transactional customer. Authority per concept, so that every price, status and owner has one place that may change it. A classification of every action an agent might take by its consequence and its reversibility, with people deciding the consequential ones whatever the confidence. And an audit trail that shows what the agent read, decided and did.
There is a useful irony here. One of the best first uses of AI in fragmented commercial data is helping people resolve the fragmentation: suggesting duplicates, explaining contradictions and preparing corrections for a data steward to approve. That work builds the control plane that wider agency needs later.
Agency can then grow in steps, action by action, as identity and authority become explicit enough to carry it.
Where the argument is applied
AI & Agentic Workflows- AI case / 03Designing AI-assisted data stewardship without moving master-data authorityAI that makes stewards faster without becoming the steward: candidates found, ranked by impact and explained with evidence and differences; a governed steward queue; identity, merge and consequential classification decided by accountable people; accepted decisions executed by the system’s own processes; and outcomes that improve evaluation and rules—never the model’s permissions.
- Systems case / 03Designing customer and master-data identity across CRM and ERPSeparating commercial, legal and transactional identity—with one issuer, explicit cardinality and governed merges—and giving authority to concepts and attributes rather than whole records.