Everett Zufelt
Connected data, not smarter models, is what lets agents make decisions you can trust
Most of the customers we speak to about agentic and autonomous commerce eventually ask the same question: “Does getting ready mean replatforming everything, again?”
It doesn’t, but it does mean facing what many modernization programs have skipped: accurate and accessible data.
Agents don’t act on your storefront, they act on your data, and in most companies that data is scattered across a dozen systems that each describe the business a little differently. You already have reporting and other integrations holding those systems together, but they access and act on data without understanding what it means, through point-to-point mappings that break the moment a field changes. Tolerable when a person reads the results. Not when an agent makes the decision.
Agentic commerce is already here, deciding which of your products a shopper’s AI assistant ever surfaces. Autonomous commerce, where AI automates tasks like pricing and inventory, is where the industry is heading next. Give either one a decision to make, and it works from what your data says. In most companies different systems say different things, and the agent has no way to know which is right.
People blame the AI. The real gap is that nothing in the stack holds a single, trustworthy answer to what a product or a customer even is. That gap is what a commerce ontology closes: a shared model of your business, defining what a Product, Customer, or Order is and how they connect, layered over the systems you already run.
This is where AI modernization should start, ahead of or in parallel with investments in autonomous agents. A commerce ontology makes agentic commerce easier now and is the precondition for success in autonomous commerce.
What breaks when agents run on bad data
When a report looks off, a person stops and checks. An agent doesn’t. It acts on the answer it’s given, at the speed and volume it was built for.
Say your product system lists an item as active and your inventory system has it discontinued. A merchandiser reconciles that in seconds. An agent may pick one version and move forward: repricing the category, adding the item to a campaign, triggering a reorder. The error may be quiet, fast, and already downstream before anyone reviews it.
Automate more on top of inconsistent data, and the mistakes multiply, each one made with full confidence, across your whole catalog.
What a commerce ontology actually is
An ontology is a model of the things your business works with and how they relate. In commerce that means entities like products, variants, categories, customers, orders, inventory, locations, suppliers, and prices. It also means the relationships between them: which variant belongs to which product, where a product is stocked, which customers qualify for which price. It isn’t another database or another system to run. It’s a layer of shared meaning above the systems you already have, so “product” means one thing whether the question comes from your storefront, your planning tool, or an agent. Integration code moves fields between systems. An ontology says what those fields mean.
Take Pureinsights Discovery as one example of how that layer gets built. It ingests product and catalog data from wherever it lives, reconciles records that point to the same real-world thing, and adds the attributes and relationships that were only ever implicit, then publishes one connected model that other systems and agents can query. The sources underneath keep running as they are. What changes is that the meaning on top of them is finally written down.
Build it as a layer, and keep it alive
It’s tempting to treat this as a cleanup project: fix the catalog, dedupe the records, done. That fails, because commerce data changes daily. New products, new suppliers, renamed fields, another channel. An ontology has to be assembled from your existing sources and then kept current as those sources shift. In practice that means automated ingestion, entity resolution (deciding when two records describe the same thing), and enrichment (filling gaps and adding the attributes agents need to reason). None of it requires replacing your systems of record. They stay where they are.
Where to start
Don’t model the whole business first. Start from a decision you actually want to hand off. Repricing a category. Reordering stock. Surfacing the right products to a shopping agent. Take one, trace back to the objects and relationships that decision depends on, and model that slice well. Get a single agent-driven decision working on data it can trust, then widen the model from there. The work stays tied to real value, and you can have something running in weeks.
Before you fund the next AI pilot, ask question: can the agent reach your commerce data wherever it lives and get one consistent answer about what it means? That, far more than the model you choose, decides how far autonomy can go.
Disclosure: Pureinsights is committed to sharing a wide range of perspectives on search and AI technologies. The views expressed in this guest post are those of the author and do not necessarily reflect the views of Pureinsights or its other partners.