Agentic AI in PIM isn’t a feature. It’s an architecture.

Since late 2025, many PIM vendors have announced “agentic AI,” and many of the announcements describe an internal copilot that writes content, classifies products and drafts channel descriptions.

Some are demos; some are in private preview.

That confuses a feature with an architecture.

An AI assistant inside your PIM is useful, the way a spellchecker in Word is useful. It improves one workflow inside one interface.

It doesn’t make your product data agentic. It doesn’t make your catalog callable by outside AI systems. It doesn’t run product work at scale without a person driving each step, and it doesn’t prepare you for the commerce landscape that is already here.

What an agent actually does.

An agent takes a goal and breaks it into steps. It calls tools, checks the results, and tries again until the goal is met. A person doesn’t approve each decision along the way.

In product work, it looks like this. The first four examples are built into CataZenta; the last depends on buyers’ agents that are still early.

• A merchandiser adds CataZenta as a connector in Claude.ai, signs in, and asks which products are least complete. Claude calls the catalog tools and returns a ranked list with the attribute gaps named.
• The same merchandiser asks for German descriptions for the ten least complete. Claude reads each product’s attributes for context, writes the copy and saves it in German with the merchandiser’s own permissions. Those saves apply directly, so the version history is where the team checks them or rolls them back.
• Inside CataZenta, an editor on a product page asks Zen AI to run the SEO optimization agent on this product. Zen AI proposes the run, and nothing starts until the editor confirms. By default, the agent’s suggestions then wait in the approval queue.
• The Compliance agent runs on a repeating schedule over one category and flags products missing required compliance fields such as country of origin. It suggests a value only where the product’s own data supports one. By default, those suggestions wait in the approval queue, and the rest are raised as findings for a person.
• Still ahead. A buyer’s AI purchasing agent compares your product against three competitors on specific attributes. Whether your product makes that comparison depends on data that is machine-readable and complete, which is work you can start in your catalog now.

None of the first four is a chatbot answering from memory. Each calls tools on the catalog itself, under your rules.

The first four don’t need custom development. Tools, user permissions, specialist agents and an approval queue are enough. The fifth waits on buyers’ agents and commerce protocols maturing, and on your data being ready when they do. Not all of this exists off the shelf today, CataZenta included.

A PIM whose only AI is a built-in copilot isn’t an AI-native PIM.

Much vendor messaging blurs that line. A PIM whose only AI is a built-in copilot has a specific shape.

• The data model was built for human editors, not for machines reading it.
• The AI can read the catalog, but it has limited ability to run a multi-step job on its own.
• People start every action in the UI, and the AI answers prompts.
• Outside systems often can’t call it without a custom integration.
• Rules come last. The AI does things, and the log says what it did.

That helps with one bottleneck. If your team writes product descriptions by hand, a copilot accelerates that task. Fine.

The larger problems stay where they were. Supplier onboarding, running the checks in the right order, compliance per channel, enrichment that runs without a person, and access for outside AI. Those need a system built for agents, not one retrofitted with a copilot. Some PIM vendors now ship MCP servers of their own; ask what they expose and under whose permissions.

Before and after, one operation at a time.

This is the target shape. CataZenta does parts of it today; see the examples above.

The change isn’t an assistant that saves your team keystrokes. It is agents running whole workflows under your rules, with people stepping in where judgment is needed.

What needs to call your catalog.

A closed copilot assumes one reader of your catalog. Your own team, inside the vendor UI. In 2026 the list of systems that need structured access to your product data is longer than that.

The AI systems that decide whether your products are visible, buyable and compliant sit outside your PIM. A copilot locked in the vendor UI is invisible to all of them.

What an AI-native PIM is made of.

A PIM built for agents from the start, rather than retrofitted later, has five layers.

1. Expose the catalog as typed tools (MCP).

Search, product edits, tasks, imports and exports are typed tools on the Model Context Protocol (MCP), the open standard AI assistants use to call outside tools. An AI client discovers the tool list itself, so you don’t write an integration per assistant. CataZenta works this way today, and Claude and ChatGPT can connect to a CataZenta catalog as a connector.

2. Every assistant call runs as the signed-in user.

An AI connection signs in as a person through OAuth. An assistant’s tool calls carry that person’s roles and workspace, not a service account with everything switched on. A merchandiser’s Claude session can’t do what an admin’s can.

CataZenta’s built-in agents work under a dedicated agent account; a suggestion someone approves is applied with that person’s permissions.

3. Agents can be chained into flows.

Specialist agents can be chained into flows, in sequence or side by side. One finds the gaps, the next drafts the missing attributes and the channel-specific fields. By default, their suggestions land in the same approval queue.

Steps coordinate through the catalog and those approvals rather than handing data to each other. A failed run is recorded, not swallowed.

4. By default, agent suggestions wait for review; assistant edits leave a version.

Many teams hesitate to give AI systems write access to product data, and the reason is a good one. An AI with nothing holding it back can damage a catalog at scale, more quickly than any person could. One invented specification published to Amazon can bring returns, a listing suspension, and a compliance violation.

By default, agent suggestions wait in an approval queue. Values based on estimates stay with a person even when auto-apply is switched on. Where web lookups are set up for your workspace, a sourced value links to the page it came from. Confidence scores on generated values show which outputs likely need a person to check them.

Limits are set per agent: fields it can’t change, the channels and languages it may write to, and a confidence level above which its suggestions may apply on their own. A run is scoped to one product, a category, a product family or the catalog in batches, and platform rules can hold or block an agent change, even one that would otherwise auto-apply. New variants, category moves and new marketplace feed plans still wait for a person.

Zen AI and connected assistants write directly, within that person’s permissions, and deletes and schema changes need an explicit confirmation. Every product keeps a version history, so an edit that turns out wrong can be restored.

5. Every decision leaves a record.

Approvals and rejections of agent suggestions are recorded with who made them. Every product edit, AI edits included, creates a numbered version with who made it and when, and people with restore permission can roll it back.

One thing many PIMs still lack is a complete, customer-facing log that tells “a person clicked a button” apart from “Claude called a tool.” CataZenta doesn’t have that yet either, so ask every vendor exactly what gets recorded.

Ask three questions when a vendor says “agentic.”

When a vendor puts agentic capabilities in “private preview” or “coming soon,” three questions tell you whether it is architecture or marketing.

1. Can outside AI systems call these capabilities? Claude, ChatGPT, my own agents, my CI pipeline. Or only your built-in copilot?
2. Is it an open protocol that works across AI clients, such as MCP or OpenAPI with tool semantics? Or is it proprietary to your platform?
3. Can I use it today, with approval controls and each workspace kept separate? Or is it a demo reserved for a handful of pilot accounts?

If the answers are “only our copilot,” “proprietary,” and “sometime next year,” you are reading a feature announcement. The vendor is still working out the architecture. Meanwhile your catalog can stay out of reach of the AI systems already sitting between buyers and products.

Run this checklist.

Check these when a vendor calls their PIM “agentic.”