What Salesforce reference data is, and why an agent depends upon it
The Salesforce reference data an AI agent reads, the relationships among its objects, and the four properties each record must possess.
Reference data is the set of records a system consults in order to act: the account, the contact, the product, the contract, the price. It is distinguished from transactional data, which records that something happened, by the fact that it is consulted repeatedly and assumed to be true. Data management professionals usually call records such as accounts, contacts and products master data, and keep the term reference data for code lists such as country codes and units of measurement; this guide uses reference data in the broader sense. A sales representative who doubts a record can draw on what they know beyond it; an Agentforce agent reasons from the data it is given, and a record that is wrong but looks right can mislead it.
How reference data in Salesforce is structured
Within Salesforce, reference data is held in a small number of standard and managed-package objects whose relationships form a hierarchy. Accounts sit at the top and are related to one another by a parent field, which is how a subsidiary rolls up to a parent; each Product is priced through a Price Book Entry, one for each price book and currency in which it is sold. An agent handling a support case, amending a contract or scoring a lead traverses several of these relationships in a single task.
The properties every record must possess
| Property | Meaning | Failure it prevents |
|---|---|---|
| Complete | Every field the consuming process requires is populated. | An agent unable to email a contact whose email address is blank. |
| Consistent | Related records agree, for example a Subscription in the same currency as its Contract. | A renewal quote priced in the wrong currency. |
| Current | The record reflects the present state of the world, not the state at entry. | An agent addressing a contact who left the company two years ago. |
| Unique | One real-world entity is represented by exactly one record. | A lead routed to the wrong representative because the account exists twice. |
Each property is measurable. Completeness is a null count on the fields that matter. Consistency is a join across related objects with a filter on inequality. Whether a record is current is measured against an authoritative external source or a last-modified threshold. Uniqueness is a match rate. A remediation that does not begin with a measurement of these four properties, by object, has no baseline against which to be scoped or judged.
Why an agent depends upon Salesforce reference data
The dependency is not symmetrical. A human user can often recognize a defective record and work around it; an agent can recognize only what the data reveals, and it can act faster than human review can typically keep up with. The condition of reference data therefore limits the reliability of any agent that reads it, whatever the model, the prompt or the vendor. What an agent can tell apart can also depend on how fields and picklist values were mapped, and what it can see depends on the permissions of the user it runs as and on what is retained.
Assess the data before the agents depend upon it
The first call is with the practitioners who would do the work, not a sales intermediary. The measures come first and the estimate follows.