This phase is about moving beyond basic documentation answers. With your agent live on its first channel, you can now connect integrations, add user data, and set up simple actions and workflows. This enables the AI to handle personalized and more complex requests.
Think of this phase as upgrading your AI from answering general questions to resolving account-specific and task-oriented requests.
Steps
1) Add user data
Give your agent access to account and customer details. This can be done in two ways:
Via Integrations
Link your AI agent with your helpdesk (Intercom, Zendesk, Freshdesk, etc.).
Why it matters: Integrations provide customer context quickly and prepare the agent for account-specific requests.
Sync user fields or organization fields from your helpdesk
Pass user IDs so the AI can call APIs for personalized info
Via Actions
Add actions that fetch user-specific data directly from your systems.
Why it matters: Actions are essential for delivering personalized answers, not just general ones. They let the agent pull live data such as plan details or trial status.
Example questions the agent can now answer:
“What is my current plan?”
“When does my trial end?”
“How many seats do I have?”
“Is this feature included in my plan?”
Note: Always test these queries in Playground to confirm correct results.
2) Set up Actions
Give your AI agent the ability to fetch or perform tasks in real time.
Why it matters: Actions can also perform operations (e.g., update account status or trigger workflows), letting the AI resolve issues end-to-end.
Here are 3 popular SaaS use cases to start with (examples):
Retrieve subscription details
Purpose: Let the AI answer account-related questions without human help.
Use case: Get subscription status
Action: GET /subscription/{userID}
Answers: “What plan am I on?”, “When does my trial end?”
Returns: plan_type, trial_end_date, seat_count
Fetch invoices or order history
Purpose: Reduce ticket volume by automating simple lookups.
Use case: Retrieve invoices
Action: GET /billing/invoices/{userID}
Answers: “Can I get my last invoice?”, “Show me billing for March.”
Notes: Support invoice_id or date range filters if available.
Cancel or pause a subscription
Purpose: Automate repetitive requests that don’t require judgment.
Use case: Cancel subscription
Action: POST /subscription/cancel/{userID}
Behavior: Changes account status immediately.
Notes: Can run standalone or inside a cancellation workflow.
Safety: Confirm user identity and capture reason before running.
Tip: Actions can be used on their own or combined inside Workflows for multi-step processes. Always test each Action in Playground before enabling it in production.
Example: Add Action screen where simple API-based actions can be configured.
3) Build your first workflows
Combine steps into guided scenarios.
Why it matters: Workflows give the AI a structured, step-by-step way to handle multi-step requests, making support consistent and predictable.
Here are 3 popular SaaS support workflows to start with (examples):
Password reset
Purpose: Automate one of the most frequent requests.
Use case: Password reset flow
Trigger: User says “I can’t log in” or “reset my password”
Steps:
1. Confirm user email
2. Send reset link
3. Ask if reset was successful
4. Escalate if unresolved
Basic troubleshooting
Purpose: Help users solve common technical issues quickly.
Use case: Troubleshooting flow
Trigger: User reports error or problem with setup
Steps:
1. Ask clarifying questions (device, browser, version)
2. Suggest 1–2 common fixes from docs
3. Confirm if issue resolved
4. Escalate if unresolved
Data collection before escalation
Purpose: Save human agents time by gathering details upfront.
Use case: Data collection flow
Trigger: User request needs escalation
Steps:
1. Collect email and workspace ID
2. Ask for a brief issue description
3. Pass all info to ticket with context
4. Escalate to human agent
Tip: Workflows can call Actions inside them (e.g., cancel a subscription or pull billing details). Always test in Playground before enabling.
4) Expand to a second request category
Once your general product questions are covered, move into the next biggest opportunity. Choose one or two of your most common, repetitive categories that have straightforward solutions.
Examples:
Common troubleshooting steps
Repetitive account or workspace questions
Service status or outage queries
5) Test after every layer
Validate the agent each time you add new data, actions, or workflows.
Use Playground for structured testing
Confirm accuracy with >80% correct answers before moving forward
Layered validation path showing how each layer (knowledge base, integration, actions, workflows) is tested before rollout.
6) Expand integrations
Once stable, embed the agent across your infrastructure.
Examples:
Website, help center, or developer docs
Intercom or Zendesk
Community channels (Slack, Discord)
Best Practices / Tips
Start with integrations — they’re faster to set up and provide context quickly.
Add Actions for personalization — they’re required if you want the AI to fetch or change user-specific data.
Pick one new request category at a time.
Test each new integration, action, or workflow in Playground.
Keep workflows simple at first — 3–5 steps max.
Use Actions for live data and Workflows for guided scenarios.
Common Mistakes to Avoid
Adding multiple actions at once without testing
Automating sensitive categories (billing disputes, data privacy issues, etc.)
Overcomplicating workflows too early
Expanding to too many channels at once
Cross-references
✅ Expected outcome: Advanced scenarios are supported with Integrations, Actions, and Workflows.