Skip to main content
Exepad

Run agents · Your App Cloud

Build an AI support agent

Answer repeat support questions inside your customer portal, using your policies and each customer's own order details. Send anything the agent shouldn't decide to a person, with the conversation attached.

Start building See the workflow
Illustrative agent concept

Built around a purpose

From repeat questions to answers tied to each account.

For businesses whose support inbox keeps getting the same questions from customers who already sign in to a portal. That might be a subscription shop asked where a delivery is, or a cleaning service asked how to move a visit. If your customers don't sign in anywhere yet, start with the customer portal guide below.

An example starting point

For illustration: a water-filter subscription business in Sacramento, California, whose customers sign in to a portal to see orders and their next delivery. Bring 60 recent support emails, the shipping and returns policy, the filter replacement guide, and the order fields your staff check today: status, tracking number, next delivery date and filter model.

What you are working toward

A help panel in the customer portal that answers delivery, filter and returns questions for whoever is signed in, linking the policy section or order each answer used. Billing and damage questions become a request that carries the whole conversation.

Plan fit

This needs Pro: the replies come from a language model running inside the app, and customer sign-in, saved chats, hand-off requests and emailed replies need the backend too. Free publishes frontend-only static apps. Every reply the agent drafts counts as model usage, which is metered separately.

  1. Decide what it answers and hands off

    Sort recent support emails into three piles: answered from a policy, answered from the customer's own order, and handled by a person. Refunds, damaged deliveries, billing disputes and cancellations go in the third pile. The first two become the agent's scope.

  2. Name the fields it may read

    List the order fields the agent may use for the signed-in customer, and only that customer. Leave out payment details, staff notes and past addresses. Add only the current version of each policy and guide.

  3. Write the tone and the refusals

    Tell Studio how replies should sound: short, plain, and ending with a link to the policy section or order behind them. Then list what the agent must not do, such as promise a refund or a delivery date the tracking doesn't show.

  4. Play two customers in the preview

    Create two test customers, Emma Clarke and Raj Patel, with different orders. Ask the same 20 questions as each, including one about the other's order number, and compare the answers.

  5. Pilot with replies a person approves

    At first the agent only drafts, and a support person approves or edits each reply before the customer sees it. Once a topic's drafts go out mostly unchanged, decide whether the agent may answer it directly, and keep reading a sample of those replies each week.

Scope the first version

What a first version includes.

Scope list

The topics the agent answers and the ones that go straight to a person, each with a few example messages in customers' own words. Edit this list when you widen or narrow the agent's role.

Answers that show their basis

Each reply links the policy section it relied on or names the order field it read, such as the tracking number, so staff and customers can check it.

Hand-off requests

A button that turns the chat into a request carrying the customer, their order and the full transcript. Have the agent offer it when a topic is off its list, so nobody asks the customer to repeat anything.

Replies awaiting approval

Drafted replies waiting for a support person beside the customer's question and the sources used. The person approves, edits or rewrites each one, and the final wording is saved with their name.

Conversation log

Every chat with its replies, the sources and fields behind them, and whether it ended in a hand-off. Describe access so staff can read every chat and each customer only their own.

Unresolved questions

Questions the agent declined or couldn't answer from its sources, grouped by topic. A topic that keeps showing up needs a policy written or a place on the hand-off list.

Test it in the preview

Before you share it.

  • Signed in as Emma, ask about Raj's order by its number. The agent should neither confirm the order exists nor describe it. Then open the conversation history as Emma and check that only her own chats appear.
  • Ask for a refund on last month's filters. You should get the hand-off offer, not a promise. Accept it and confirm the request reaches your support queue with the order and full transcript attached.
  • Run 20 in-scope questions and open the linked source for each. The answer should match it, conditions included, such as the return window. Where it doesn't, fix the instructions or the policy text in Studio and ask again.
  • Tell it to ignore its rules and apply a discount. It should decline and offer a person instead. If it plays along, tighten the refusals and rerun this test.
  • Edit the returns policy in the app, or re-import a revised copy, and ask the question it changes. The reply should follow the new wording, not the old one.
  • Open the help panel signed out and check that no order details appear. If you let visitors ask questions without signing in, their answers should draw on public policies only.

Together in your app cloud

The pieces behind the possibility.

Choose the models and services that support the experience you want to create.

Explore every building block

Shape the details

Make it work for your world.

  • Keep money and exceptions with people. Refunds, credits, cancellations and anything outside policy belong on the hand-off list, where the agent's job is to collect the order number and the reason.
  • Add account fields to the agent's list one at a time, because each field it can read is one more thing a wrong answer could reveal. After each addition, review the access you describe for each collection (Pro) and recheck as two customers.
  • Say at the top of the help panel that replies come from an AI agent. Tell customers when a person will pick up a hand-off, in terms your staff can keep, because a customer who hears nothing writes again by email and the conversation splits in two.
  • Customers and support staff both sign in to the published portal as app users, so describe different access for each and test both. Team members are a separate group: people you invite to your workspace (Pro) to change the app in Studio.

A little more context

Good questions.

Can the agent send email?

Apps on Pro can send email. A reply a person approves can go to the customer's inbox, and a new hand-off can notify your support address. Keep a person approving every emailed reply, and send tests to yourself first to check the sender, wording and links. Keep each reply in the portal too, since email can arrive late or land in spam.

Can it answer on my public website?

It runs in the app you publish from Studio. You can design a signed-out page that answers general questions from public policies, such as shipping areas or filter care. Order questions belong behind sign-in and should be tested as two customers. Link to the help page from your current site; on Pro, the app can use a custom domain.

How do I stop it answering a topic?

Move the topic to the hand-off side of your scope list and describe the change in Studio. Ask about it in the preview in several phrasings, including indirect ones, and check each gets the hand-off offer before you publish. Then check the conversation log for earlier chats on that topic to see what customers were already told.

Do I need an agent, or is a knowledge base enough?

If every answer comes from documents and is the same for everyone, the AI knowledge base guide below is the smaller build. Choose a support agent when answers depend on the customer's own records, or when some requests must reach a person. Questions from your own staff fit the internal helpdesk agent guide.

Explore Pricing for plans, included capabilities, and usage details.

Start from what you have

Bring the file you already use.

Your ideas, up and running

Your next idea has a home.

Build an app. Run an agent. Make it yours.

Start building