Skip to content

How to write store policies an AI can actually follow

The Suvenna team · July 8, 2026 · 6 min read

Open your store's return policy and read it the way a support agent has to: with a specific customer, a specific order, and a specific question in front of you. "Items may generally be returned within a reasonable period, subject to inspection, unless otherwise noted." Returned by whom? Within how long? Inspected against what standard? Noted where?

Most store policies were written to limit liability, not to answer questions. They're defensive documents, drafted once, full of hedges like "generally," "may," and "at our discretion," and they work only because experienced human agents paper over the gaps with judgment. The policy says "a reasonable period" and the agent knows the founder means 30 days.

An AI agent has no folklore to draw on. Neither does the human you hired last week. The fix is the same for both: restructure your policies so that any agent, silicon or otherwise, can apply them deterministically, meaning two different agents reading the same policy against the same order reach the same answer. Here's how, rule by rule, with a before-and-after you can copy.

One rule per sentence

The single highest-leverage change. Legal-style prose packs three rules, two exceptions, and a disclaimer into one sentence, and every reader has to unpack it correctly under time pressure. Unbundle it.

Instead of: "Returns are accepted within 30 days for unworn items in original packaging with proof of purchase, excluding sale items and gift cards, though exchanges may be possible."

Write:

  • Returns are accepted within 30 days of the delivery date.
  • Items must be unworn and in their original packaging.
  • The order number counts as proof of purchase.
  • Sale items marked "final sale" cannot be returned. They can be exchanged for a different size within 30 days.
  • Gift cards cannot be returned or refunded.

Same policy. But now each rule can be checked independently against a real order, quoted verbatim to a customer, and updated without rewriting a paragraph. When a rule is its own sentence, an agent can cite it. When it's a clause buried in a compound sentence, an agent has to paraphrase it, and paraphrase is where errors enter.

Explicit thresholds and time windows

Every vague quantifier in a policy is a decision you've silently delegated to whoever reads it. "A reasonable period" becomes 21 days for one agent and 45 for another. "Significantly damaged" means different things before and after lunch.

Go through your policies and replace every judgment word with a number or a named condition:

  • "Promptly" becomes "within 5 business days."
  • "A reasonable period" becomes "30 days from the delivery date."
  • "Heavily used items" becomes "items with visible wear, stains, or missing tags."
  • "Expensive orders may require review" becomes "orders over $200 require manager approval before refunding."

Anchor every time window to an unambiguous event. "Within 30 days" is incomplete until you say 30 days from what: the order date, the ship date, or the delivery date. For a Shopify store these are real, distinct timestamps on the order, often a week or more apart, and an agent with access to order data can check a delivery-anchored window automatically. Pick one anchor, state it, and use it consistently.

If you genuinely want discretion in some cases, that's fine, but make the discretion itself a rule: "Anything not covered here goes to a human." Discretion that's written down is an escalation path. Discretion that's implied is a coin flip.

Order rules like a decision tree

Agents read policies to answer a question, not to enjoy the prose. Structure each policy in the order an agent actually needs the information: eligibility first, then process, then outcomes, then exceptions.

  1. Eligibility. Who can return, what items qualify, within what window. The rules that produce a fast "no" go first, because most policy questions are eligibility questions.
  2. Process. How the customer initiates, who pays return shipping, what condition and packaging are required.
  3. Outcomes. What the customer receives, refund to original payment method or store credit, and on what timeline.
  4. Exceptions. The named edge cases, each with its own rule.

Within each section, put the general rule before its exceptions, and make every exception point at the rule it modifies. A policy ordered this way reads like a checklist, and a checklist is something an agent can execute rather than interpret.

Name the edge cases

The gaps in a policy don't show up while writing it. They show up at 9pm when a customer asks the question you never wrote down, and the agent either guesses or stalls. The most common gaps for Shopify stores are predictable, so close them in advance:

  • Final sale. Which items, exactly? Is "final sale" marked on the product page, the cart, or both? Can final-sale items be exchanged for size, or is the sale truly final? What if the item arrived damaged, does final sale still apply?
  • Partial refunds. If a customer returns two items from a three-item order, how is the refund calculated? Who eats the original shipping? What happens to a discount code that required a minimum order value the remaining items no longer meet?
  • Exchanges. Is an exchange a return plus a new order, or a distinct flow? What if the replacement item costs more, or less? Does the return window reset on the exchanged item?
  • Damaged or wrong items. Do your normal return rules apply when the mistake is yours? Most stores waive the window and the return shipping here, but almost none write that down.

You don't need to cover everything on day one. You need to cover the questions customers actually ask, and your support history already contains that list. Every time an agent, human or AI, escalates because the policy has no answer, that's the policy telling you what to write next. A well-run store's policy document grows a rule or two a week for the first few months, then stabilizes.

Before and after

Here's a typical returns paragraph, restructured with the four techniques above.

| Before (written for lawyers) | After (written for agents) | | --- | --- | | "Returns may be accepted within a reasonable period of purchase provided items are in acceptable condition, subject to inspection upon receipt. Certain items may not be eligible. Refunds are processed in a timely manner after approval, less any applicable fees, at the discretion of management." | Returns are accepted within 30 days of the delivery date. Items must be unworn, unwashed, and have all original tags attached. Final-sale items and gift cards cannot be returned. Final-sale items can be exchanged for a different size within the same 30-day window. The customer pays return shipping, except when we shipped a wrong or damaged item, in which case we send a prepaid label and waive the 30-day window. Refunds go to the original payment method within 5 business days of the return passing inspection. Inspection fails only for wear, stains, odor, or missing tags, and we email a photo when it does. Anything not covered here goes to a human. |

Every sentence in the right column is a rule an agent can check against a real order and quote to a customer. Nothing in the left column is.

How Suvenna consumes your policies

This structure isn't just good hygiene. It's exactly the shape an AI agent needs, and it's how Suvenna works by design.

Suvenna answers customer questions strictly from two sources: your own policy documents and the customer's real order data from Shopify. When a customer asks "can I return this?", the agent checks your written window against the actual delivery date on their actual order and answers from the rule, not from a general model's idea of what return policies usually say. The answer a customer gets is your policy, applied, and the agent can show which rule it applied.

Just as important is what happens outside the rules. If your policy doesn't cover the question, Suvenna abstains and escalates to your team rather than improvising something plausible. No invented exceptions, no goodwill refunds you never authorized, no confident answers pulled from the internet's average return policy. An unwritten rule produces an escalation, never a guess, which means every gap in your policy surfaces as a handoff you can see instead of a promise you never made. We've written about why that architecture matters in more depth.

And when a policy decision involves money, a refund, a cancellation, a return label, the action is confirm-gated, bounded by the limits you set, and logged. The agent proposes, the rules constrain, the audit trail records.

The practical upshot: an afternoon spent rewriting your policies with these four techniques pays off twice. Your human agents get faster and more consistent immediately, and your AI agent gets a rulebook it can follow deterministically from day one.

If you want an AI agent that treats your policies as the source of truth instead of a suggestion, request early access. We onboard in limited waves so every store gets a clean rollout, and founding members lock in 20 percent off their first year.