When the Catalog Becomes Chaos, AI Starts Deciding Based on a Truth That Doesn’t Exist

TL;DR

• General-purpose agents magnify a problem that predates AI: the disconnect between what systems report and what actually applies to a B2B transaction.
• If pricing, inventory, credit, and terms are fragmented, an agent may execute with confidence without validating the authoritative source.
• Before delegating commercial decisions to automation, the CTO must define data authority, context, and traceability.
• AI’s opportunity lies in operating within explicit governance—not in replacing business rules with probabilistic answers.

Who Ensures the Agent Consulted the Truth Before Making a Decision?

For CTOs, the arrival of agentic AI creates an uncomfortable tension: the company wants to accelerate decisions and reduce manual work, but the catalog meant to support that automation often does not represent a single source of truth.

The problem emerges when a system displays one price while the valid terms depend on the customer, available credit, inventory, business unit, negotiated agreement, or a rule maintained in another environment. Each piece of information may be correct in isolation and still produce the wrong commercial decision when combined without the proper context.

The premise “When the General-Purpose Agent Arrives Before the Catalog Truth” captures this risk: the proliferation of general-purpose models amplifies the disconnect between what systems say about pricing, inventory, and terms and what actually applies to a B2B transaction.

An agent trained on or directed by outdated or fragmented data can confidently submit orders based on information its architecture never validated. In that scenario, model quality is not the only issue. The more fundamental question is architectural: who ensures that the agent consulted the authoritative source before making a decision?

The Catalog Is No Longer Just a List

In a B2B operation, a catalog is not simply a list of available products. It represents the combination of what can be sold, to whom, under which terms, and within what commercial context.

That is why two concepts must be distinguished:

• Data availability: The agent can access a piece of information.
• Data authority: The architecture can prove that the information is valid for that specific decision.

The first makes it possible to generate an answer. The second makes it possible to trust the execution.

Without this distinction, connecting an agent to more systems does not necessarily solve the problem. Access may increase, but so does ambiguity. The agent encounters multiple versions of prices, inventory levels, catalogs, or terms and must act without clear rules governing precedence, validity, and authorization.

AI does not create this fragmentation. It removes some of the human friction that previously slowed its conversion into an order, quote, or commercial commitment. Without governance, greater speed also shortens the path between an inconsistency and its operational impact.

Model Confidence Is Not Commercial Validity

There is an important difference between producing a coherent answer and making a valid commercial decision.

An agent may correctly interpret a request, locate an item, and generate a plausible response. None of that proves that the price is current, the inventory can be committed, the customer qualifies for those terms, or the combination has been authorized.

For CTOs, the question should not be limited to “Which model should we use?” Before that, the organization must answer:

• Which source is authoritative for each data element?
• How are conflicts between sources resolved?
• Which context determines the applicable terms?
• What may the agent recommend?
• What may it actually execute?
• Which decisions require additional validation?
• How can the data and rules used in a decision be reconstructed later?

Without these answers, the autonomy granted to the agent may exceed the maturity of the architecture supporting it.

The Problem Is Not Solved by Choosing One More System of Record

In organizations with local or business-unit-specific systems, declaring a single platform the universal source of truth may ignore how operations actually work. One source may control inventory, another may maintain commercial data, and another may contain the terms applicable to a particular context.

The central issue is not necessarily replacing all those systems. It is establishing governance that determines which information applies, at what time, and to which decision.

That requires treating business rules as an explicit part of the architecture. Pricing, credit, catalogs, and commercial terms cannot remain implicit knowledge scattered across systems, spreadsheets, and employee judgment while an agent is expected to reconstruct the correct logic on its own.

The company’s negotiation DNA lives in those rules: limits, exceptions, approval thresholds, and combinations that turn an available product into a valid commercial offer. If that knowledge is not represented in a governed way, AI may reproduce fragments of it without necessarily preserving the operation’s underlying logic.

The Cost of Inaction

Postponing catalog governance does not leave the company in the same position. As agents begin accessing data and executing tasks, long-standing inconsistencies gain speed and scale.

The cost of inaction appears in three areas.

First, reliability. If sales teams or customers discover that prices, inventory, and terms vary depending on the channel they use, automation loses credibility. People begin manually confirming every decision, reducing the expected value of AI.

Second, accountability. Without clear data authority, an error can be passed among the model, integration layer, ERP, and sales organization without any way to identify where the decision ceased to be valid.

Third, scalability. An agent may work within a limited scope but become difficult to expand when every new business unit, system, or policy adds another possible interpretation of the catalog.

The risk, therefore, is not limited to executing an incorrect transaction. It also includes building an automation system that depends on parallel verification and a growing number of exceptions just to remain operational.

Principles for Governing Agents Across B2B Catalogs

• Define authority before expanding access: Every relevant data element needs clear rules for origin, validity, and precedence.
• Separate recommendation from execution: An agent can support a decision without receiving authority to complete it from day one.
• Preserve commercial context: Product, customer, credit, inventory, and terms must be evaluated as parts of the same decision.
• Make rules explicit: Exceptions and approval thresholds should not depend solely on institutional memory.
• Log the decision: It must be possible to reconstruct which data and rules supported an action.
• Treat conflicts as an expected scenario: Diverging sources are not anomalies to ignore; they are conditions the architecture must resolve.
• Govern before automating: Speed creates value only when the accelerated decision remains valid.

FAQ

Are General-Purpose Models the Problem?

Not by themselves. The risk emerges when an agent is given autonomy over fragmented or outdated data without an architecture that validates the authoritative source.

Is Integrating the Agent With the Company’s ERPs Enough?

Integration provides access. By itself, it does not determine which business rule takes precedence or which information is authorized in each context.

Do Existing Systems Need to Be Replaced?

The material analyzed points to another approach: an orchestration layer can centralize business rules and return only what each context authorizes to each system or touchpoint—without replacing the existing ERPs.

Where Does AI Create Value in This Scenario?

In interpreting requests, assisting users, and executing tasks within clear boundaries. The opportunity grows when AI operates using authorized data, explicit rules, and traceable decisions.

Who Is Already Experiencing This

On Software Advice, Leonardo C., a verified reviewer in the automotive industry at a company with 1,001–5,000 employees, wrote:

“We work with B2B solutions on CWS”

The statement is brief and should be interpreted within that limitation: as a public reference to the use of CWS in B2B solutions.

An Illustrative Case

Case LI-042 from the archive demonstrates the architectural dimension of the problem: in B2B operations with multiple ERP systems, the challenge is not merely technical. What is missing is an orchestration layer above local systems that can centralize business rules—such as pricing, credit, and catalog rules—and return to each touchpoint only what its specific context authorizes.

This approach does not require treating a general-purpose agent as the arbiter of truth or replacing every ERP. It establishes a governed boundary.