AI Policy
Last updated: September 7, 2026
This AI Policy explains, specifically, where and how Returner uses AI (large language models) as part of the Service, what data reaches the AI provider, and how it is controlled. It supplements our full Privacy Policy, which lists Anthropic, PBC (maker of Claude) as a Sub-Processor. This AI Policy does not replace that Privacy Policy.
AI features are off by default
None of the AI surfaces described below run for a store until its owner explicitly accepts this AI Policy in Settings > Return Policy > AI concierge. Until accepted, these features are structurally disabled: the app falls back to non-AI, deterministic behaviour instead of silently degrading or queuing the request. Acceptance can be reviewed (who accepted, and when) from the same settings page.
Where AI is used
| Surface | Who sees it | What's sent to Claude | Merchant control |
|---|---|---|---|
| Sidekick (in-app assistant) | Merchant / store staff, in the dashboard | The staff member's typed question, plus store-scoped analytics figures (return rate, reasons, products) needed to answer it. If the staff member asks about a specific order, return, or customer, the matching records (which can include that customer's name, email, and order number) are shared with the model to answer that question. Simple navigation and common analytics questions are answered without any AI call at all. | Requires AI Policy acceptance (Settings) |
| Insight briefings & enriched insight cards | Merchant / store staff, in the dashboard | Aggregated store metrics (return rate, top reasons/products, refund SLA, financial summary). No individual customer data. | Requires AI Policy acceptance (Settings) |
| AI concierge (return tracking chat) | Shoppers, on the return tracking page | The shopper's typed messages, plus the details of their own single return/order (including the name and contact details on that order) and the store's return policy. Nothing about other customers or other returns is ever accessible. | Off by default; requires both AI Policy acceptance AND a separate "Enable AI concierge" toggle (Settings) |
| Support-ticket triage (internal) | Returner's own engineers, in our internal support queue | When store staff submit an in-app bug report, the report text and its captured page context (page URL, browser info, recent console lines) may be AI-summarised to speed up our response. No store analytics or customer records are included beyond what the reporter typed or what appeared on their screen's console. | Requires AI Policy acceptance; without it, reports are triaged by a human only |
The shopper-facing concierge additionally asks each shopper to acknowledge a short in-chat notice before their first message is sent. That acknowledgement is separate from, and in addition to, the merchant's Settings acceptance.
The AI provider
AI-assisted features are powered by Anthropic's Claude API. Under Anthropic's standard commercial API terms, data submitted through the API is not used to train Anthropic's models.
We want to be precise about retention, because it is the part people most often assume. We have not entered a zero-retention agreement with Anthropic. Inputs and outputs from our AI features are therefore retained by Anthropic for their own standard operational and trust-and-safety period, under their published data-retention terms, and not deleted on transmission. Zero retention is an account-level agreement, not something an application can switch on per request. If we enter one, we will say so here.
What AI does not do
- It cannot access data outside the current store (Sidekick, insights) or outside the current return (concierge): every tool call is scoped server-side to the authenticated tenant or verified shopper, not something the AI model can override.
- It never executes refunds, approvals, or other money/state-changing actions on its own. It can only surface a suggested action for a human to click.
- Figures shown by the AI (return rate, financial totals, etc.) are always computed deterministically by the app and handed to the model as fact; the model narrates, it does not calculate.
Questions
For questions about this AI Policy, contact us via the details in our Privacy Policy.