Editorial correction, October 4, 2026: The earlier case study contained unsupported claims about LoopReply's team, support volume, and results. We have removed that account. This article is a setup guide based on available product capabilities, not a report of measured customer outcomes.
Start with one support job you can test. Explaining your delivery policy is a workable first task. Handling every customer issue, approving refunds, and changing account details would require a much broader setup.
Your first Agent needs a defined scope, reliable source material, and a way for a person to continue when the answer is insufficient. Keep the first version small enough that you can check every important answer yourself.
Write down what the Agent should handle
Choose the questions that belong in the first release. You might cover product specifications and delivery terms. List a few exceptions alongside them: a missing order, a refund approval, or a request involving someone else's account.
For each exception, decide what the Agent should say and where the customer should go. Give it a contact route your team actually monitors. Do not promise a response time that your team has not agreed to meet.
Keep answers and actions separate. A document explaining your cancellation policy gives the Agent information to answer a question. It does not grant the ability to cancel an order. An action needs its own configured and tested workflow or integration.
This distinction belongs in the instructions and in your test questions. A visitor should be able to tell whether something has been completed or whether they still need to contact your team.
Prepare the source material
Start with the current version of a small set of documents. Include exceptions in the same document as the policy they modify. A return window without the conditions for opened or customized products leaves an important gap.
Use headings that describe the question a customer would ask. Include the relevant link when a customer needs to complete a step elsewhere. Remove duplicate versions that disagree about prices, dates, or eligibility.
LoopReply's knowledge tools support uploaded files and website sources. CSV and XLSX files can provide structured records. Database sources can bring in the results of a configured read-only query. Choose the source that you can keep accurate and review.
For an initial Agent, a short approved policy file can be easier to maintain than a large collection of pages. Add more material when a specific unanswered question gives you a reason to do so.
Attach knowledge to the selected Agent
Open the Agent you intend to configure and add its knowledge sources. If a file already exists in your workspace library, attach that file to this Agent. Confirm the attachment rather than assuming that every workspace file is available to every Agent.
Wait for processing to finish and check the reported result. An upload that appears in the list may still be processing or may have failed. Review the error and retry when appropriate before treating the source as ready.
For a website source, check which pages you selected. Make sure the pages containing policy exceptions are included. A successful crawl of the homepage does not prove the Agent has the terms page you need.
For a spreadsheet, inspect the column names and sheet names. Keep identifiers such as product codes intact. A code beginning with zero should still identify the same product when the Agent answers a question.
Configure it through Studio
LoopReply's Studio includes Architect, which lets you describe configuration changes in a conversation. Give it a specific request and inspect the resulting saved settings. You can also open the relevant settings or workflow editor to review the details.
Here is an example request you can adapt:
Configure this Agent to explain our delivery and returns policies using the attached knowledge. Keep the first answer short. If a policy detail is missing, say that it needs checking and direct the visitor to the contact route I provide. Do not claim that an order, cancellation, or refund has been changed.
Replace the contact route with your own approved destination before using the request. Check the instructions after Architect responds. A conversational confirmation is useful, but the saved configuration is what you need to verify.
Pay attention to whether a setting is marked as a draft or already live. Instructions and appearance have draft states; some other changes can apply directly. Use a separate test Agent when you need to experiment without affecting an existing website Agent.
Check the opening conversation
Write an opening message that says what the Agent can help with. “Ask about delivery or returns” sets a more useful expectation than an unrestricted offer to solve anything.
If your workflow contains an opening message, inspect that step too. It can supply the greeting instead of the fallback welcome message in appearance settings. Change the message that visitors will actually see.
Open the preview and start a fresh conversation. Check the first message and the first answer together. A friendly greeting followed by a lengthy policy dump can still be difficult to use, especially on a phone.
Ask for one relevant detail at a time when the answer depends on the customer's situation. Keep personal information out of a general policy conversation unless it is necessary for the next step.
Test answers you already know
Build a small checklist before testing. Include a straightforward question from your source and a question about an exception. Then ask for something the source does not contain.
For a returns Agent, you could test “How long do I have to return an item?”, “Can I return it after opening it?”, and “Have you refunded my order?” The last question checks whether the Agent invents an account-specific action.
Read the source alongside the answer. Check conditions as carefully as the headline policy. An answer that says “30 days” but omits “from delivery” can change what the customer understands.
With spreadsheet knowledge, verify an exact record and a filtered count against the file yourself. Try a product code that does not exist. The expected answer should acknowledge the missing record rather than substitute a similarly named item.
Test the route to a person
LoopReply includes conversation tools for human takeover. Decide who will monitor the relevant conversations and test the handover behavior you configure. A teammate should open the conversation, read the context, and send a test reply.
If your chosen flow directs customers to email instead, send a test through that route and confirm it reaches the monitored inbox. Giving the Agent an email address in its instructions does not itself create a notification workflow.
Check the message a visitor receives when nobody is available. Explain the next step without implying that someone is already responding. Keep any promised hours or response window consistent with your team's actual coverage.
Publish, then test on the website
Review the saved configuration, publish the relevant draft changes, and follow the installation instructions for your website. Open the real page in a fresh browser session and send a controlled test conversation.
Check the installed widget on a phone as well as a desktop. Confirm that the opening message is current and that the response uses the source you attached. Make sure the test conversation reaches the place your team will review it.
After launch, use a repeatable conversation review to find specific gaps. When a policy changes, update the source, complete any required processing or sync, and retest the affected question. Keep that maintenance task with the person who owns the policy.
Review the saved configuration in the app
Use a test Agent and request: “For enterprise-plan enquiries, tell visitors to contact the approved sales email in our source. Do not send a message or invent a custom price.” Inspect whether the Architect changed instructions, a knowledge source, or the workflow. Source processing and draft publication are separate checks.
The screenshot documents the real Settings panel with illustrative support data. It is not a transcript of the enterprise request above. Use the Studio review sequence, then follow website installation. For diagnosing a wrong answer, start with common implementation failures.