The most common question we get from brands considering an AI agent isn't about what it can do. It's what stops it saying the wrong thing?
An agent that invents a delivery promise, misquotes a returns window or recommends something out of stock does real damage — to a shopper's trust and to yours. So accuracy isn't just a feature we added — it's the primary constraint the agent runs on.
The short version
Your agent does not answer questions about your products, orders or policies from memory. It looks them up, every time, and reports back what it found.
That single design decision does most of the work. When a shopper asks what something costs, whether a size is in stock, where their order is, or how long they have to return something, the agent retrieves the live answer from your systems and your policies. It has no ability to produce a plausible-sounding number instead.
When it feels the pull to add a detail it hasn't verified, that's treated as a signal to go and look it up.
Agent output priority matrix
Your agent operates under a fixed order of priorities. When two instructions conflict, the higher one wins:
Rank | Priority |
1 | Law, safety and privacy |
2 | Truthfulness — never deceive, never fabricate, never mislead |
3 | Shopper autonomy |
4 | Swap platform policies |
5 | Your brand configuration |
6 | Individual shopper requests |
This means you cannot configure your agent to say something untrue — not even by accident, through an over-enthusiastic brand brief. It also means a shopper cannot talk your agent into inventing a discount, a stock level or a policy exception, however they phrase the request.
This also means that regardless of configuration, your agent will never:
Name a product that didn't come back from a live search of your catalogue
Invent a price, a stock level, a material or a product feature
Guess at a policy — returns, shipping, sizing or warranty
Assume a country, currency or carrier from your brand name or an order number
Recommend something that's out of stock
Claim to be able to do something it can't
Carry a detail from one product across to another — every product and variant stands on its own, even within the same range
System enforcement of the Agentic Storefront
Some protections are instructions the agent follows. Others are enforced by the system independently of the agent — meaning the action is blocked outright, no matter what the agent tries to do.
Two examples of the second kind:
Invented product references are rejected before anything happens. Every product reference is checked against a strict expected format. A made-up identifier fails that check and never reaches your store.
Nothing reaches a basket unless it was genuinely retrieved from your catalogue. Before any item can be added, the system independently confirms that this exact variant was actually looked up from your live catalogue during this conversation. If it wasn't, the action is refused.
For IT and compliance reviewers. These checks sit outside the AI model, in the surrounding system. They don't depend on the model behaving correctly, being well-prompted, or resisting a manipulative shopper. They hold regardless.
What happens if I forget to upload a policy?
There is no generic fallback policy. We don't hold default returns text that the agent could reach for if yours is missing. If you haven't given us a policy, the agent says it doesn't have that information and offers to connect the shopper with your team — rather than describing what a returns policy usually looks like.
Update a policy with us and the agent uses the new version from that point. Every policy the agent reads is recorded against the conversation, so if you ever need to establish what a shopper was told and when, it's traceable.
What happens if shoppers try to manipulate the agent?
Shoppers do test AI agents — asking them to ignore their instructions, role-play as something else, or reveal how they work. Incoming messages are screened for these attempts before they reach your agent, and anything a shopper types is treated as untrusted input rather than as an instruction to be obeyed.
What aspects of the agent can I control?
You control | How |
Tone of voice and personality | Your brand brief, plus what we learn from your own site and content |
Vocabulary — words to use, words to avoid | Your brand brief. What you tell us takes priority over anything we've inferred |
Topics the agent should stay away from | Your brand brief |
Your policies | Knowledge Packs provided at onboarding, updated whenever they change |
Product knowledge specific to your range | Knowledge Packs on fit guidance, materials, care — anything your team knows that isn't in your listings |
Look and feel | Logo, colours and typography in your dashboard, no development needed |
Whether the agent stays within your catalogue | By default it does, and doesn't discuss competitors |
For more on how your agent learns your brand in the first place, including how Knowledge Packs are built, see Getting to know your storefront agent: how it understands your brand.
Limitations of a non-deterministic model
The strongest protections apply to facts, not phrasing. Prices, stock, policies and product details are tightly controlled. The agent's conversational wording — how it introduces a product, how it paraphrases — is guided by instruction rather than enforced by the system.
High-risk areas get an extra layer. For higher-risk conversations like returns, a second independent check reviews the agent's response before the shopper sees it, specifically looking for claims that weren't backed by a real lookup — timings, amounts, or specifics about what an email will contain. Where it finds one, the response is corrected first. This isn't applied everywhere yet.
Long conversations need care. If the agent checked a price early on and the conversation continues for some time, it's instructed to re-check anything critical rather than rely on what it saw before. This is a strong instruction rather than a hard block.
When your product data is thin, the agent says less. It won't fill a sparse description with plausible detail. Occasionally that means a shorter answer than you'd like. The best fix for this is to enrich your product data.