Personalized AI infrastructure
Superintelligence for Ecommerce: Catalog Memory and Returns Rules
Organize catalog facts, return policy versions and supplier data as company memory so an AI agent answers from current rules and sends exceptions to staff.

Superintelligence for an ecommerce brand means a stronger AI agent that answers customers, staff and suppliers from a governed record of the business, not from whatever it absorbed during training. The decisive layer is company memory: current catalog facts, the return policy version that applied on each order date, and the supplier terms behind each product. Be Human designs that memory, its permissions and its approval boundaries before an agent is given any customer-facing role.
Key takeaways
- Treat catalog facts, return policy versions and supplier terms as dated records with named owners, not as text pasted into a prompt.
- Answer return questions from the policy version in force on the purchase date, and let the agent state which version it used.
- Give the agent read access to the catalog and policy, and approval rights only for narrow, reversible actions.
- Test the agent against real customer questions before widening its scope, and keep that test set current.
- Start with one role, such as returns triage or product fit questions, and expand only after the memory has earned trust.
Why catalog and returns questions are a memory problem
A customer asks whether a jacket bought in the spring can still go back, and the agent answers with confidence. The answer is wrong, but the model is rarely the root cause. The policy changed after that purchase, the product page still shows the older window, and a warehouse note says the supplier moved this style to a different rule. The agent did what it was built to do: it produced a fluent answer from the material it had.
Ecommerce generates this problem constantly. Prices, sizes, materials, stock status, shipping cut-offs and return windows change on different schedules, and each lives in a different tool: the storefront, the product information system, the supplier portal, the helpdesk macros and a spreadsheet someone keeps for exceptions. An agent connected to all of them without a clear hierarchy will sometimes quote the newest source and sometimes the most visible one. The customer only sees that the answer contradicts the last one they received.
Superintelligence, in the sense Be Human uses the term, does not fix this by becoming more knowledgeable on its own. It requires the business to organize what it knows so that every answer can be traced to a current, owned source. That is the memory layer. Brands that already keep customer history in one place will recognize the pattern, which is explored in superintelligence for consumer brands and customer memory across channels.
What belongs in company memory for an ecommerce brand
Not every document belongs in memory. A useful memory for a commerce agent holds the records that change how it should answer, and each record needs an owner, an effective date and a status that says whether it is current. Retrieval designs for this pattern, where current records are fetched at the moment an answer is composed, are described in Google Cloud's reference architecture. A practical starting set includes:
- Catalog facts: product name, variant, dimensions, materials, care instructions, compatibility and the stock status rule the storefront uses.
- Return and exchange policy: each version with the date it took effect, the products or categories it covers and the exceptions that apply.
- Shipping and cut-off rules: carriers by region, handling times and the rule for orders placed near a holiday or a promotion.
- Supplier data: lead times, minimum order quantities, warranty terms, country of origin and the contact path for defects.
- Exception decisions: approved one-off outcomes, such as a goodwill replacement, with the reason and the person who approved it.
- Service language: tone guidance and the wording the brand has agreed to use for refunds, delays and damaged goods.
Keep the list short at the start. A catalog with thousands of products does not need to enter memory on the first day. The first version should cover the products and policies that generate the most customer questions, and the rest can follow once the team sees where answers are weakest.
How to version return policies so answers match the order
Return policies are where memory most often fails. A brand changes its window for a season, adds a rule for final sale items or starts charging for returns from one region, and the old version remains visible on a help page, in a confirmation email and on a marketplace listing. An agent that reads only the newest version will misstate the terms for every order placed before the change. An agent that reads every version without a date rule will mix them.
A workable approach has four steps. Give each policy version an identifier, an effective start date and an end date when it is replaced. Attach each order to the version in force when it was placed, using a field the order system can supply. Have the agent cite the version it applied in every answer, so a reviewer or the customer can see the basis. Retire old help page text only after the orders covered by that version have closed their return windows.
The steps translate into a checklist the operations lead can review before any agent goes live:
- Every return policy version has an identifier, an effective start date and a named owner.
- Each live order can be matched to exactly one policy version by a field the agent can read.
- Exceptions are logged with the approver, the reason and the order they apply to.
- The customer-facing answer states the policy version or its date in plain language.
- Superseded versions are marked as superseded, not deleted, so past orders remain explainable.
Illustrative scenario: a home goods brand shortens its return window and removes free return shipping for one category. Orders placed before the change keep the earlier terms. A customer writes about a lamp bought two weeks before the change and asks for a refund by mail. The agent checks the order date, identifies the version in force, confirms the earlier window still applies, and drafts a reply with the return steps and the last day to send the parcel. Because shipping costs are involved, the permission rules route the draft to a person before anything is sent.
Supplier data and catalog facts need named owners
Catalog and supplier information disagree more often than most teams expect. The product sheet says a fabric is machine washable because the buyer copied the claim from an earlier season, while the supplier's current care label says otherwise. The agent has no way to know which is right unless the business has decided. Memory is useful only when every field has a rule for which source wins and who corrects the other.
A simple rule works in most cases. The supplier is the source for technical and warranty facts the supplier controls. The brand is the source for merchandising, pricing, promotions and policy. When the two conflict, the record is marked disputed, the agent gives a cautious answer that does not repeat the disputed claim, and the item goes to the person who owns the catalog. The agent should not resolve the conflict itself, because a guess about washing instructions becomes a warranty dispute later.
Supplier data also changes outside the storefront, for example when a manufacturer discontinues a colorway or moves production to another site. Ask each important supplier how it sends updates, and record that channel as a source in its own right. The aim is not to automate every supplier relationship. It is to know which facts the agent may state as current and which it must flag.
Permissions and approval boundaries for a commerce agent
A customer-facing agent that can read the order system, the policy memory and the helpdesk needs a clear line between what it may do and what it may only propose. Read access to catalog facts and the current policy is low risk and high value. Writing to the catalog, issuing refunds above a set amount, changing a policy or promising compensation are different in kind. Each needs an explicit approval path, and each permission should be granted to the agent's own identity rather than shared with a human login.
Authorization guidance from OWASP stresses that access should be checked on each request against the specific action, not assumed because a tool happens to be connected. For a commerce agent, the refund tool should verify the amount and order state every time, and the agent should have no way to widen its own rights through a prompt. Permissions also need scheduled review, because teams change, suppliers change and a temporary exception often outlives its purpose. The review routine is described in how to run a permission review for your AI agents.
Approval thresholds deserve the same care. Start with a narrow set of actions the agent can complete alone, such as sending an order status update or issuing a standard return label for an eligible order, and send everything else to a person with the draft already prepared. Thresholds can rise later, once answers have proven reliable on real cases and the approvers trust the records behind them.
How to measure the agent and keep the memory current
An agent that answers ecommerce questions needs a test set before launch and again after each change. Collect real customer messages from the helpdesk, include the difficult ones about partial returns, bundles, marketplace orders and damaged items, and write the expected answer for each, including the policy version it should cite. Anthropic's guidance on evals describes this kind of repeated testing as the way to catch regressions when prompts, tools or memory change. Run the full set after every policy update or catalog migration.
Record enough in production to explain any answer. A useful log names the memory sources read, the policy version applied, the tool calls made and whether a person approved an action. OpenTelemetry's observability concepts describe the same principle in general engineering terms: traces and logs should let a team reconstruct what happened, not merely confirm that a request completed. When a customer disputes an answer, the team should be able to see its basis quickly.
Memory also needs a change process. A new return policy, a supplier change or a seasonal exception should go through a written request that names the record affected, its owner, the effective date and the test cases to rerun. The routine for this is set out in how to request and approve changes to a managed AI agent.
Be Human operates this layer as a managed service. The first engagement defines the one role, the memory sources it may read, the tools it may use, the approval boundaries and a success measure, such as the share of return questions in the test set answered from the correct policy version. Ongoing work covers monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, delivered remotely in English or Portuguese. The records and the decisions about policy remain under the business's control.
To find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
What does superintelligence for an ecommerce brand actually require? It requires the business to keep catalog facts, return policy versions and supplier terms as current, owned records that an agent can read and cite. The model matters less than whether each answer can be traced to the right source. Permissions and approval boundaries are set before the agent speaks with customers.
Should a return policy be stored as one current document? Usually not. Keep each version with an effective date, attach each order to the version in force when it was placed, and mark older versions as superseded rather than deleted. That allows the agent to answer correctly about past orders as well as new ones.
Who should own catalog and supplier data? Assign one owner per record type, typically the merchandising lead for catalog facts and policy, and the purchasing or supplier manager for technical and warranty details. When the two sources disagree, the record should be marked disputed and sent to its owner rather than resolved by the agent. This keeps guesses out of customer answers.
Can an agent issue refunds on its own? Only for narrow, reversible actions defined in advance, such as a standard return label for an eligible order. Larger refunds, exceptions and any policy change should go to a person with the draft already prepared. Approval thresholds can be raised later, once tested answers have shown they are reliable.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


