Skip to main content
RoutMe

Playbook

AI chatbot for SaaS support: setup and limits

By the RoutMe team. Last updated .

An AI chatbot for SaaS support combines two sources: your help documentation for how-to questions and a read-only lookup against the account record for facts like plan, seats and billing status. Together they answer a large share of routine tickets. Bugs, billing disputes and pricing negotiation should go to a person with the conversation and lookups attached.

The two kinds of questions

SaaS tickets split cleanly. How-to questions, such as how to invite a teammate or export data, are answered by documentation. Account questions, such as which plan someone is on or why an invoice looks different, need data about the specific account. A documentation-only bot cannot answer the second kind and often bluffs. The difference is the core of RAG vs tools for customer support. The SaaS support industry page shows how RoutMe handles each.

Make the help center usable by an assistant

Retrieval quality depends on content quality. Before connecting anything:

  1. Give each article one clear topic and a descriptive title.
  2. Put the answer near the top, then the steps.
  3. Remove or mark outdated pages, especially after a redesign.
  4. Write plan-specific differences explicitly, for example which features exist on which tier.
  5. Keep terminology consistent with the product's interface labels.

Add the docs as knowledge sources. A page added by URL is a snapshot, so re-add it after significant changes.

Account lookups

Expose a small set of read-only tools: current plan and seats, billing status and last invoice summary, usage against limits, and whether a feature is enabled. The guide on connecting your REST API to an AI agent describes how to choose endpoints and shape responses, and what MCP is explains the protocol that connects them. Return only the fields a support answer needs.

Identity and permissions

An anonymous visitor on the marketing site is not a logged-in admin. Decide how the assistant learns who is asking. If the widget runs inside your authenticated product, you can pass a signed identity from your own backend so the assistant knows the account and role. If it runs on the public site, limit it to docs and general plan information. Respect roles too: a member should not see billing details meant for an owner.

Status and incidents

Add a tool that reads your status page, so when users ask whether something is down the answer is current. During an incident the assistant can acknowledge the issue and link the status page instead of troubleshooting from scratch, which also keeps people from filing duplicate tickets.

Handoff to the right queue

Different problems need different owners. Set rules that route:

  • Suspected bugs to support engineering, with steps, account and any error text the customer gave.
  • Billing disputes to finance or billing support.
  • Pricing, contract and expansion questions to sales or account management.
  • Security reports to your security contact, not through general chat.

The assistant should gather the details a human needs, such as browser, time and what was expected, before handing off, so the first human reply is a fix and not a question. See human handoff best practices and the support deflection and handoff job.

Where it should not act

Refunds, plan changes, deletions and permission changes are consequential. A read-only assistant explains how a customer can do these in the product, or passes the request to staff. If you later automate any of it, keep it a separate, explicit flow with confirmation. What read-only AI means explains why this default matters.

Check the results

Define what success means before launch: conversations the assistant finished without a person, conversations handed off and why, and customers who came back with the same question. How to measure AI support quality gives definitions. Read the handed-off conversations first, since they show documentation gaps and missing tools.

How RoutMe approaches this

RoutMe lets you add help content as knowledge sources, connect account data through MCP or APIs in read-only mode, embed a chat widget with a script snippet or use the Chat SDK for deeper integration, and hand off to a shared inbox. See the developer documentation for connection details and pricing for plans.

Sources

Explore the product: RoutMe for Shopify, solutions and the documentation.

Frequently asked questions

Can it use our help center?

Yes. A URL or pasted text becomes a knowledge source that the assistant searches by meaning. URL sources are a snapshot taken when added.

Will it know which plan a customer is on?

Only if you connect a lookup that returns it, and the assistant knows who is asking. Without that, it should say it cannot see account details.

Can it process refunds?

Not by default. It is read-only, and refunds stay with your team unless you build a deliberate, confirmed flow.

Should it be in the app or on the marketing site?

Both can work with different scopes. Inside the authenticated app it can use account data, while on the public site it should stick to docs and general information.

Put these ideas to work