Deploying a Chat Agent

This guide walks you through deploying an AI-powered chat agent on your website using Outrun. By the end, you'll have a live chat widget that answers visitor questions using your actual business data.

Before You Start

You'll need:

• An Outrun workspace with at least one connected data source (a CRM like HubSpot, Pipedrive, or Zoho works best)
• Access to add a script tag to your website (or your developer can handle this step)
• A few test questions to verify the agent's responses

Step 1: Choose a Template

Outrun includes three chat agent templates. Pick the one closest to your primary use case:

Pre-Sales Agent — Best for product-led growth sites. Answers questions about features, pricing, and capabilities. Qualifies visitor interest and routes hot leads to your team.

Lead Qualification Agent — Best for B2B sites with a sales team. Asks qualifying questions conversationally, captures contact details, and creates CRM records.

Customer Success Agent — Best for logged-in customer portals. Answers account questions, troubleshoots issues, and pulls context from the customer's actual data.

You can deploy multiple agents — the triage router will automatically route visitors to the right one based on their questions.

Step 2: Connect Your Data

Each agent template specifies what data sources it can use. In the agent settings:

1. Select your connected sources — choose which CRM, support, or knowledge base integrations the agent should have access to
2. Review workspace context — the agent pulls live data from your workspace to answer questions accurately

The data is already synced from your connected tools. There's no separate training step — the agent queries your data in real time.

Step 3: Configure the Agent

Customise the agent for your business:

• Agent name — this appears in the chat widget (e.g. "Sarah from Acme")
• System prompt — the template includes a default prompt; customise it with your company's voice, specific product details, or response guidelines
• HITL settings — choose whether to review all responses, only low-confidence responses, or let the agent respond directly
• Email capture delay — set how long to wait before asking for the visitor's email address (configurable per agent)

Step 4: Embed the Widget

Add the Outrun SDK script to your website. It's a single script tag that loads asynchronously:

<script src="https://cdn.getoutrun.com/sdk/outrun.min.js" async></script>

If your visitors are logged in, you can identify them so the agent knows who it's talking to:

Outrun.identify({
  email: "[email protected]",
  name: "Jane Smith",
  company: "Acme Corp"
});

The widget appears as a chat launcher in the corner of your site. Visitors click to open and start a conversation.

Step 5: Test the Agent

Before going live, test with a few scenarios:

1. Product question — "What integrations do you support?" — verify the agent gives an accurate, specific answer
2. Pricing question — "How much does it cost?" — verify it responds appropriately (or escalates if pricing is sensitive)
3. Off-topic question — "What's the weather?" — verify it stays on-topic and redirects politely
4. Support question (if using multiple agents) — "I can't log in" — verify it routes to the right agent

Check the workflow run history to see how each message was triaged and which agent responded.

Step 6: Go Live and Monitor

Once you're satisfied with test results:

1. Enable the widget for all visitors (or a percentage to start)
2. Monitor the HITL queue — review and approve agent responses for the first few days
3. Check the metrics — response time, resolution rate, escalation rate
4. Make corrections — when you edit an agent response, the correction is stored and used as a reference for similar future questions

Ongoing Improvement

After the first week:

• Review correction patterns — are the same types of errors recurring? Update the system prompt to address them
• Adjust HITL thresholds — if accuracy is consistently high for certain question types, move those to auto-respond
• Add more agents — if visitors frequently ask questions outside your first agent's scope, deploy a second specialist

After the first month:

• Check correction memory — the agent should be handling previously-corrected question types correctly now
• Review routing accuracy — if using multiple agents, check that the triage agent is sending conversations to the right specialist
• Measure business impact — compare lead conversion rates and support ticket volume before and after the chat agent