Agentic Commerce is Replaying an Old Enterprise Data Mistake

For a long time, B2B commerce worked under a simple assumption: Humans browse.

They read product pages, skim spec sheets, and tolerate vague language because they know how to ask follow-up questions. When something is unclear, they email a sales rep. When a rule is buried in a footnote, experience fills in the gap.

B2B product data evolved entirely around that behavior. It never had to stand on its own; it only had to be interpretable by a human. With AI, that assumption is no longer the case.

We’ve Been Here Before with Enterprise Data

If this feels familiar, it should. A decade ago, enterprises were having a very similar conversation about data. Warehouses were full, data lakes were overflowing, and ultimately every system was exporting something. On paper, companies were data rich. In practice, nothing moved fast because business users couldn’t answer basic questions without analysts as their translators. SQL became a choke point.

Enterprise data was organized around how systems stored information, not how people reasoned about the business. Rows and columns existed, but concepts did not. Revenue lived in three tables. “Customer” meant five different things depending on who you asked and when. Metrics were debated endlessly because no one had defined them clearly.

The breakthrough in enterprise data came from accepting complexity and containing it. Semantic layers are one example, but they were part of a broader shift. Enterprises stopped pretending that raw data was usable by default and started building translation layers that matched how the business actually thought and operated.

Metrics models did this by defining calculations once instead of re-deriving them in every report. Revenue meant the same thing everywhere because someone had taken the time to encode it. Data models and dimensional schemas did the same thing structurally. They turned operational tables into concepts like customer, product, order, and time. Business users no longer had to understand how many joins were required to answer a basic question. The relationships were already there.

Data catalogs and governed definitions handled another part of the problem. They captured meaning that used to live in people’s heads. What does this field represent? When should it be used? What are its limitations? Context stopped being tribal knowledge and became part of the system.

Together, these layers absorbed complexity and made it operable. They created stable abstractions that allowed more people — and more systems — to reason correctly without reinterpreting the world from scratch every time. That is exactly what B2B commerce is missing today.

Agent-Led Discovery is Triggering the Same Reckoning

Agentic commerce is forcing B2B product data through the same test. Manufacturers and distributors are not short on product information. They already store enormous amounts of it: from specifications to configurations to pricing logic to contractual constraints.

The problem is that almost all of this data was structured for humans. Specifications live in PDFs. Rules are explained in a physical product catalog that never made it online. Exceptions are implied in a back-office sales process, rather than encoded. Far too much depends on institutional memory when context lives in sales teams’ heads.

An AI agent doesn’t skim a PDF and “get the idea.” It doesn’t know which sentence is a hard constraint and which one is sales language. It can’t safely infer rules from formatting or tone. If the meaning isn’t explicit, the agent treats it as unknown.

This Isn’t About Unstructured Data Being Bad

It’s worth being clear about something. Unstructured data is not the enemy. It never was.

In enterprise analytics, unstructured data didn’t disappear when semantic layers showed up. It got layered on top of structure. Structure handled rules and relationships. Unstructured content handled nuance, explanation, and context.

The same pattern applies here.

Agents need structure to reason. They need explicit rules, relationships, constraints, and states. They need to know what is compatible, what is configurable, what is allowed, and under what conditions something applies. Unstructured content alone can’t reliably provide that.

But structure alone isn’t enough either. Agents don’t just retrieve attributes. They compare options. They evaluate tradeoffs. They decide both what something is and when it should be recommended.

Narrative is the layer that explains intent, positioning, and use cases. It’s the difference between “this product exists” and “this is when you should choose it.” In the enterprise data world, this showed up as definitions, documentation, and business context. Here, it shows up as a product-level explanation that agents can learn from. While structured product data tells the agent what is true, narrative helps it decide what matters.

Commerce Has Been Optimized for Presentation, Not Reasoning

This is the uncomfortable part. Commerce infrastructure never really made the leap enterprise data did. We built better PIMs. We built richer catalogs. We built prettier product pages. But we never built a true semantic layer for products; we optimized for presentation.

As long as humans mediated B2B buying, that was fine. Sales reps explained edge cases. Buyers tolerated ambiguity, and everyone knew how to work around the system.

Agents remove that buffer. In B2B, the cracks show immediately. Prices vary by account. Availability changes by region. Compatibility depends on configuration. Contracts override defaults. Entitlements matter. None of this is safely guessable.

When an agent evaluates a product, it isn’t impressed by a well-written description. It wants to know what fits, what’s allowed, what’s compatible, and what happens next. If that information isn’t explicit, the agent doesn’t ask for clarification; it just moves on.

What Commerce Companies Need to Do Now

This is the inflection point. Commerce companies can keep treating product data as content that humans interpret. Or they can start treating it as infrastructure that machines reason over.

That means specifications need to become attributes with defined meaning. Compatibility needs to be encoded as relationships, not explained in paragraphs. Pricing needs to be expressed as logic. Entitlements need to be explicit. Availability needs to be stateful and precise.

This is exactly the same move enterprises had to make with analytics. When raw data and tables weren’t enough, meaning had to be defined. And once that structured core exists, narrative stops being the only source of truth for AI and becomes the layer that teaches agents how to apply that truth in real situations.

Manufacturers and distributors who do this will become legible to agents. Their products will be easier to evaluate, easier to recommend, and easier to trust. Those who don’t will still “have data,” but it will function like old enterprise warehouses did: technically present, but practically unusable.

The Pattern is Old, but the Consequences are Not

None of this is speculative. We’ve already watched enterprise data go through this exact cycle. The only difference now is the user. Instead of business analysts, it’s autonomous agents. Instead of dashboards, it’s recommendations. Instead of slow decisions, it’s instant exclusion.

Agentic commerce is exposing a decades-old enterprise data problem. The companies that recognize that — and treat product data the way enterprises learned to treat operational data — will adapt quickly. The ones that don’t will keep adding PDFs, rewriting descriptions, and wondering why agents never seem to choose them.

History is repeating itself. This time, the machines are paying attention.