Let your agent *do* things, not just retrieve information. Look up an order, book a meeting, file a Jira ticket, refund a purchase — wired to your real systems.
A customer asks your agent: "What's the status of order #4521?"
Today, the agent can answer this only if the order details happen to be in your knowledge base — which they aren't, because order data lives in your e-commerce backend, your fulfillment system, or your CRM. The AI says something polite about not having access. The customer files a support ticket. Your team looks up the order manually. Forty seconds of human work for a question that could be answered in milliseconds.
This is the wall every retrieval-only agent hits. Most real customer questions aren't about static documentation — they're about live state in your systems.
Custom tools break through that wall.
Define a tool, the AI uses it. A tool is a small piece of configuration — name, description, parameters, and the API endpoint or function to call. Once defined, the AI knows when to invoke it and what arguments to pass.
Built-in tools out of the box.
- Order lookup for common e-commerce platforms (Shopify, WooCommerce, Stripe)
- Calendar booking for scheduling demos or support calls (Calendly, Cal.com, Google Calendar)
- Ticket creation for engineering escalations (Jira, Linear, GitHub Issues)
- CRM lookups for account data (Salesforce, HubSpot)
Custom tools via webhooks. For anything not built-in, you point a tool at one of your existing API endpoints. The AI calls it as needed, parses the response, and uses the data in the answer.
Permission-aware execution. Tools can require user authentication before firing — "please log in before I can look up your order." High-impact tools (refunds, deletions) can require human approval before execution.
This is the difference between "chatbot" and "agent." A retrieval-only system answers questions. A tool-enabled system completes tasks. The product becomes meaningfully more useful — the kind of useful that justifies a higher price point.
Customer satisfaction goes up dramatically. Self-service that actually solves the problem ("here's your order status, your package arrives Thursday") beats self-service that just delivers information ("orders typically arrive in 3-5 business days"). Resolution rates climb. Tickets that reach humans drop.
Revenue protection. Booking, scheduling, and account-action tools turn passive conversations into active outcomes. A visitor who asks about a feature can book a demo without leaving the chat. A frustrated customer can self-serve a refund without escalating.
Engineering hours reclaimed. Internal teams can use tools too — wire up an "incident lookup" tool to your monitoring system, a "deploy status" tool to your CI, an "on-call schedule" tool to PagerDuty. Your agent becomes the front desk for your operations.
Each tool is defined declaratively:
get_order_status)order_id: string)When a customer's question matches a tool's description, the AI extracts the parameters from the conversation, calls the tool, and incorporates the result into its answer. The user sees a natural-language response — the tool execution happens invisibly in between.
Errors are handled gracefully — if a tool times out or returns an error, the AI falls back to apologising and offering to escalate, never showing raw stack traces or technical details to the customer.
Tools are powerful, which means they need guardrails:
This feature is planned and will ship in waves:
When this ships, your agent graduates from a smart FAQ to a real customer-facing teammate — one that can resolve issues end-to-end without the conversation ever leaving the chat.