Autonomous actions

AI Assistant doesn't just answer, it acts too. Create booking, message customer, modify. But destructive operations need confirmation.

The 5 tool-use rounds

Per message, AI can work in max 5 steps:

Read data
Maybe read more (e.g., "what did they book before?")
Calculate or analyze
Propose action
Execute after confirmation

Infinite-loop prevention. If task takes more than 5, AI confesses and asks.

The 2 operation categories

Safe (no confirmation)

Read-only tools, never change data:

search_appointments
search_customers
get_services
get_staff
get_statistics
get_inventory
get_promotions
check_availability
get_schedules
create_customer (write but low-risk)

Confirmation-required

Destructive or costly:

cancel_appointment
modify_appointment (mainly price/date)
send_message (SMS/email to customer)
redeem_voucher
redeem_package

Confirmation flow

When AI wants to use a confirmation-required tool:

AI analyzes: "Cancel Anna's 14:00 booking?"
Confirmation card appears in chat:
- Action description
- Affected object (booking, customer)
- Context (no-show fee? package session restore?)
- Options: "Yes, do it" / "Cancel" / maybe alternative
You decide
AI acts only after

Context-aware confirmation

Card provides context:

Cancellation window: "Within 4h, 50% no-show fee?"
Package session: "Restore session to package?"
Waitlist: "Notify waitlisted customer immediately?"

Not just "Yes/No," but informed decision.

Team members and autonomous actions

Capability permissions per role:

Owner: all capabilities
Manager: all except subscription changes
Reception: booking, customer, message (but not other staff data changes)
Staff: read-only (queries)

Permissions configurable: Capabilities and role permissions

When what is autonomous

Scenario 1, booking creation You: "Book Anna tomorrow 14:00 for haircut" AI: "OK. 14:00, haircut 60 min, 8,000 HUF. Book?" → confirmation You: "Yes" AI: books, confirmation email to customer

Scenario 2, cancellation You: "Cancel Anna's Friday booking" AI: "Beyond 4-day window, no no-show fee. Cancel and email customer?" → confirmation You: "Yes" AI: cancels, email out

Scenario 3, marketing You: "Launch Black Friday SMS campaign to every Premium customer with 20% off" AI: "150 customers, 150 SMS = 150 credits. Template: '...'. Launch?" → confirmation You: "Yes" AI: campaign starts, status update

Scenario 4, dispute resolution You: "Refund Anna's 5,000 HUF payment" AI: "Refund 5,000 HUF to card? Financial record generated." → confirmation You: "Yes" AI: refund initiated

Scenario 5, AI uncertain You: "Book me 'something' tomorrow" AI: "Multiple interpretations (haircut, manicure, color). Which?" → AI doesn't act without confirmation

Safety safeguards

Restricted topics: turn off off-topic responses
Custom instructions: e.g., "no discount above 30%"
Activity log: every AI action logged (Settings → Activity Logs)
Role-based capabilities: turn off specific functions per role

Tips

Always read confirmation card, don't auto-click Yes.
Autonomous actions cost more credits, complex tool-use chains.
"Cancel" option always available, anytime regret.
Activity log is gold for disputes, replayable if something's off.
Use custom instructions for exception rules (e.g., "no cancel fee for Premium").
Capability permissions per role, don't give Staff autonomous actions.

Autonomous actions and confirmation