Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
Knowledge Base

Keeping Policy Answers Accurate: Returns, Fees and Terms

AI SNS CHAT EditorialPublished 7 min read

How to keep AI answers about returns, pricing and terms correct: one source of truth, clear ownership, human handoff and regular testing in AI SNS CHAT.

A customer asks your website chat whether they can return a sale item after 30 days. The AI answers confidently, but the policy changed last month and the answer is wrong. Now you have an unhappy customer and a decision to make: honor the mistake or contradict your own assistant.

Policy questions about returns, refunds, fees, delivery, cancellation and warranties carry more risk than most. This article shows how to keep those answers accurate: where the wrong answers come from, how to structure your knowledge, and how to test and maintain it over time.

Why policy answers go wrong

Wrong policy answers rarely come from the AI "making things up" out of nowhere. They usually come from the material it was given. The common causes are:

  • Conflicting sources. The returns page says 30 days, an old PDF says 14, and a pasted FAQ says "contact us".
  • Stale content. A policy changed, but one of the places that mention it was never updated.
  • Vague wording. "Most items can be returned" leaves the AI to guess which items.
  • Missing exceptions. The standard rule is documented, but the sale-item, custom-order or digital-product exceptions are not.
  • No owner. Nobody is responsible for noticing that a policy changed.

Notice that none of these is a model problem. Switching to a different model will not fix two documents that disagree. Accuracy on policy topics is mostly a content management discipline.

Start with one source of truth

Pick one canonical place for each policy and make everything else point to it or be removed. If your returns policy lives on your website, in a help-centre article, in a PDF and in a staff handbook, you have four copies that will drift apart.

A practical approach:

  1. List every policy topic customers ask about: returns, refunds, shipping, fees, cancellation, warranty, privacy.
  2. For each, name the single document that is authoritative.
  3. Find every other place the policy is repeated and either delete the copy or replace it with a link to the authoritative one.
  4. Record the owner and the date of the last review.

Write policies so an AI can apply them correctly

A policy written for lawyers or for an internal audience is often a poor source for chat answers. Rewrite the authoritative version so that it is explicit.

State the rule, the conditions and the exceptions together

Compare these two versions of the same policy.

Vague wording
Answer-ready wording
Most items can be returned within a reasonable time.
Unused items can be returned within 30 days of delivery.
Sale items may be subject to different terms.
Sale items are final sale and cannot be returned, except if faulty.
Refunds are processed promptly.
Refunds go back to the original payment method after the item is received and inspected.
Fees may apply.
A return shipping fee applies unless the item arrived damaged or incorrect.
Vague policy text vs. answer-ready policy textIllustrative example

Use question-and-answer pairs for the high-risk topics

For the questions customers ask most, write the exact answer you want given. A Q&A entry removes interpretation. Include the edge cases that cause disputes: opened items, gifts, items bought with a discount code, international orders, and what happens after the deadline.

Be explicit about what you do not say

If you do not offer price matching, say "We do not offer price matching." An absence in the documents invites the AI, or the customer, to assume the opposite.

Put dates and currencies on anything that changes

Fees, thresholds and promotional terms should state the currency and the date they took effect. For international audiences, also state which regions a rule applies to, since a US returns window and a UK or Australian one may differ.

Set guardrails for what the AI should not decide

Some questions should not be answered from general knowledge at all. Decide in advance which topics the agent must answer only from your documents, and which it should hand to a person.

Good candidates for human handoff include:

  • Disputes about a specific order or charge.
  • Requests for exceptions to the standard policy.
  • Legal questions, such as consumer rights in a particular country.
  • Anything where the customer is already upset.

In your agent's persona instructions, say plainly: answer policy questions only from the knowledge provided; if the knowledge does not cover the question, say so and offer to connect the customer with a person. An honest "I don't have that detail, let me get a colleague" is far better than a plausible guess.

  1. 01Customer asks about a policy
  2. 02Agent checks the approved knowledge
  3. 03Clear answer found: reply with it
  4. 04No clear answer: hand off to a person
How a policy question should be handledIllustrative example

Test before you publish and after every change

You would not publish a new returns policy without proofreading it. Treat the AI's answers the same way.

Build a small policy test set

Write 15 to 30 questions covering your highest-risk topics, including awkward phrasings and edge cases. For example:

  • "I bought this on sale, can I send it back?"
  • "It's been 35 days, is it too late?"
  • "Who pays for return postage?"
  • "Do you refund the delivery fee?"
  • "Can I cancel my subscription mid-month and get money back?"

Next to each question, write the answer you expect. Run the set whenever a policy or its source document changes.

Compare, then correct

For each test, check three things: is the answer factually correct, is it complete (does it mention the exception), and does it avoid inventing details such as a promised timeframe you never stated? When an answer is wrong, fix the source rather than only patching the symptom, otherwise a differently worded question will bring the error back.

Review real conversations

Real customers phrase things in ways you did not anticipate. Skim recent conversations on policy topics each week at first, then monthly. Look for answers customers rated poorly, chats that were escalated, and questions the AI could not answer. These show you where the knowledge has gaps.

Keep policies current with a simple routine

Accuracy decays unless someone maintains it. A lightweight routine is enough:

  • Assign an owner for each policy, ideally the person who decides the policy itself.
  • Update the source first, then the AI. Make "update the knowledge" part of the checklist for any policy change, alongside the website and the email templates.
  • Review on a schedule. Quarterly for stable policies, and before peak seasons such as holiday sales, when return and shipping questions surge.
  • Keep a change record. Know who changed what and when, so a surprising answer can be traced to its cause.
  • Plan for temporary rules. Seasonal extended return windows should have an end date, and someone responsible for removing them.

Doing this in AI SNS CHAT

AI SNS CHAT gives you the pieces for this workflow, and you can start without any special setup.

Add your policy as controlled knowledge. Under Data sources in the sidebar you can add Files, Text snippets, a Website, Q&A entries and more. For policies, the Q&A source is a good fit for your highest-risk answers, since you write the exact wording. If your policy already lives in a help centre or wiki, connectors under Channels cover tools such as Confluence, GitBook, Zendesk Guide, Google Drive and Notion, so the agent reads from the same place your team edits. Wiki connectors sync up to 1.5M characters and 400 pages at a time, and the app reports when content was truncated, so you know if a long policy library was cut off.

Set guardrails in the persona. Each agent has a persona and guardrails. Write the rule there: answer policy questions only from the knowledge base, and offer a person when the answer is not covered.

Hand off to humans. Use Helpdesk → Scheduling (Standard plan or higher) to set shifts so handoffs go to whoever is on duty, with alerts to Slack, Teams, email, WhatsApp, LINE WORKS or a webhook. When a handoff happens, the server reads the rota and adds the "what happens next" line itself; it never invents a wait time. In the inbox or chat logs you can take over a chat, and the bot pauses while you reply.

Teach the agent from corrections. When a reply is wrong, edit it into the ideal answer to train the agent, in both the email Inbox and Activity → Chat logs. Chat logs also support filters, flags, escalation and thumbs up/down ratings, which helps you find the answers worth reviewing.

Test safely. Use the Playground to run your test questions before customers see changes. Playground A/B model compare lets you check two models side by side on the same question.

Track changes. Workspace settings → Activity records every create, edit, delete or upload with name, email, feature and time, and each data source shows who added it and who last edited it. That makes the "who changed the returns text?" question easy to answer. Restore points (up to 40 per agent) let you roll back if an update makes things worse.

Plans and limits vary, so check pricing and features for what your plan includes.

Frequently asked questions

Can the AI guess a policy answer if my documents don't cover it?

You control this through the persona and guardrails. Instruct the agent to answer policy questions only from the knowledge provided and to offer a human when it has no clear answer. Then use the Playground to confirm it behaves that way on questions your documents do not cover.

How do I fix a wrong answer without it coming back?

Correct the underlying source, not just the one reply. Update or remove the conflicting document or Q&A entry, then re-run your test questions with different phrasings. Editing a reply into the ideal answer in chat logs or the Inbox also trains the agent on that case.

Who should be allowed to edit policy content?

Limit editing to the people who own the policies. AI SNS CHAT has roles such as Owner, Co-owner, Member, Support and Viewer, with per-member permission overrides, and the workspace activity log shows who changed what and when.

How often should I review policy answers?

Review right after any policy change, before busy seasons, and on a regular schedule such as quarterly. In between, skim poorly rated or escalated conversations in Chat logs to catch gaps early.

#knowledge base#policy answers#accuracy#returns#customer support

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.