The questions you’d ask before signing.

On the storefront

Where and when does it open?

The shopper opens it from the launcher, or from an “Ask” button you place in your own page. A small prompt bubble also appears after a few seconds, or when a shopper reads most of a page, lingers, or (on desktop) moves to leave. One tap dismisses it, and it can be switched off, reworded, delayed or capped for your store.

One thing opens on its own today. On screens wider than a phone, the full chat opens about eight seconds after a visitor lands on your home page, at most once per browsing session, and not at all once they have opened the chat themselves in that session. Dismissing the prompt stops it only for that page view: if the visitor reloads or comes back to the home page later in the same session, it can still open. It also opens when you have switched the prompt off, and it can’t be switched off per store yet. Making it off unless you turn it on, and letting you choose which pages prompt at all, are on the list.

Can shoppers dismiss it?

Always. It closes to the launcher and the conversation is preserved, so reopening continues rather than restarts. There is no interstitial the shopper can’t get past, and nothing blocks the underlying page.

Does it respect inventory, variants and promotions?

Inventory and pricing come from your catalog as of its last sync, so items that were out of stock at that sync aren’t recommended as available. The catalog is synced at setup and re-synced when we refresh it, not continuously, so tell us when prices or stock change materially. Variants are handled where your feed exposes them at variant level. Promotional pricing depends on where the promotion lives — catalog-level price rules are picked up; checkout-time cart promotions are not visible to the agent and it won’t claim them.

Geography and delivery restrictions?

Delivery and shipping questions are answered from your own delivery and policy pages. Where your content doesn’t cover it, the agent says delivery needs checking rather than assuming.

How do add-to-cart and checkout work?

On Shopify and WooCommerce the agent adds the item to your store’s own cart and hands the shopper to your normal checkout. On WooCommerce, choosing a product’s options (size, pack, colour) inside the chat is still being finished. On other platforms it links to the product page, and checkout runs exactly as it does today. We don’t replace, wrap or intercept checkout, and no payment data touches HOLLOU.

When it isn’t sure

What happens on low confidence?

This is the part worth reading twice. When nothing in your catalog or your pages matches the question well enough, the turn is flagged as thin and routed to the stronger model on Pro and above, and the model is instructed to answer only from what it was given — to say it doesn’t know rather than produce a plausible answer to stay useful. On those turns, if your store has a contact page set, the agent can also offer the shopper a link to your team. There is no separate fact-checker reading the finished answer back; the grounding is in what the model is given, not in a review of what it wrote.

Every one of these is logged as a catalog gap in your console, which over time is one of the more valuable outputs — it’s a list of the questions your product data can’t currently answer.

What if a model provider goes down?

Each model sits behind a circuit breaker. Your instance has a primary provider (Anthropic Claude or Microsoft’s Azure OpenAI Service) and the other as its automatic backup: if the primary can’t answer, the reply comes from the backup. If no provider can answer, the agent says the assistant is temporarily unavailable and shows the catalog products that best match the question, rather than hanging or erroring at the shopper.

Integrations

What’s production-ready today?

Any website, via one script tag — that’s the primary path and it’s platform-agnostic. Shopify has a direct catalog sync in production. WooCommerce, Magento and custom stacks are supported through feed or API ingestion, which is production-ready but involves a mapping step during onboarding rather than being one-click.

We’ll tell you which category you’re in on the first call, and we won’t describe a mapping exercise as a one-click install.

Data

What’s retained, and where?

Conversations and events are stored on your merchant instance to power your console. Your catalog data is stored to serve retrieval. Personally identifying information in transcripts is masked by default, with reveal-on-demand recorded in an audit log.

Is customer data used to train models?

No. Your customer data is never used to train external models.

What does the platform side see?

Aggregate counts — sessions, tokens, cost, health. The platform console shows no message content, and central reporting rejects any payload that carries it. Your transcripts are stored on your own instance, scoped to you, and on Pro and above you read them in your own console. We also keep copies off the instance: nightly database backups, where off-server backup is switched on for your store, and a full export archived if your store is deleted. Those copies, and the ways our engineers can reach your instance, are set out on our trust page.

Compliance

Data is encrypted in transit on every connection, under a certificate issued for your own instance. Disk encryption at rest is not switched on yet — it is a small change and it is on the list, and we would rather tell you than let you assume it. On data protection: any shopper’s data can be erased on request, personal details are masked in transcripts by default with every reveal audit-logged, and your data is never used to train external models. Your instance runs on AWS, in the region chosen for your store: Sydney (ap-southeast-2) or N. Virginia (us-east-1). Our sub-processors are AWS, Anthropic, NVIDIA and Microsoft (Azure OpenAI Service); a reply is generated by Anthropic or Azure OpenAI, whichever is primary for your instance, with the other as the automatic backup. We do not publish a DPA yet; Enterprise includes one. Because we never take a payment or ask for card details, and checkout stays on your platform, we are outside PCI-DSS cardholder-data scope. We are working toward SOC 2 Type II and do not hold a report today — if your procurement process requires a current report, we are not the right fit yet, and we’d rather say so now.