Every answer runs the same governed loop
From a question to a verified answer with receipts — and back again, because the recorded outcome sharpens the next one. Here is a real refund question moving around the loop.
Ask
“Where is my refund for order #A-2093?”
Ground
Looks up order #A-2093 and answers only from approved records.
Decide
Applies the refund policy — issuing a new refund would need approval.
Act / hand off
Reads the record; no money moves and no false promise is made.
Verify
Confirms the evidence and cites the source — no invented limit.
Record
Logs every step to the audit trail, ready to replay.
The outcome is recorded and becomes context for the next question — the loop closes.
26 real actions, grouped by what they do
Talking is the easy part. The agent can also look things up and take guarded actions. This is everything it can actually do today, organised into five families.
Knowledge & answers
3 tools
Answers grounded only in approved business knowledge — every claim carries a source.
Orders & commerce
6 tools
Reads the real order, product and account state — and can start a checkout or a test-mode refund.
Design Studio aftercare
13 tools
Full self-service support for every design-studio project a customer owns.
Everyday utilities
5 tools
Deterministic helpers that never touch secrets, the network, or customer data.
Escalation & tickets
3 tools
When a request needs a person, it opens a ticket with full context — never a dead end.
It sees, it hears, and it speaks
Support is not only typing. Customers can show a screenshot or just talk — the agent understands all three.
It sees
Reads customer screenshots and photos as native image input.
- PNG · JPEG · WebP · GIF
- Up to 3 images per turn
- Magic-byte verified · 5 MB each
It hears & speaks
Real-time voice in and out, with natural barge-in when the customer interrupts.
- Speech-to-text: ElevenLabs Scribe
- Streaming text-to-speech
- Interrupt any time
It reads knowledge
Retrieves approved knowledge and cites exactly where the answer came from.
- OpenAI File Search
- Inline citations
- Fails closed if unindexed
Every answer shows its work
The agent is wrapped in guardrails and writes a durable receipt for every step, so an operator can replay exactly what happened — and why.
Four guardrails, always on
Relevance classifier
Off-topic messages are stopped before the model even runs.
Jailbreak shield
Prompt-injection and jailbreak attempts are blocked up front.
Privacy refusal
Cross-customer data requests are refused deterministically, before any model call.
Tenant isolation
Every tool is scoped to one merchant and the signed-in customer only.
Three levels of autonomy
Not every action carries the same risk — so the agent does not treat them the same. Reads are instant; money and accounts always pause for a person.
Auto
Read-only · answered instantly
- Order & product lookups
- Knowledge answers
- Status & utilities
Confirm
Writes that need an explicit step
- Create a checkout session
- Open a support ticket
- Send a project link
Approval
Money & accounts go to a human
- Refunds & cancellations
- Account changes
- Anything with no evidence
Answers you can act on, not walls of text
When the agent commits a real action, it renders a structured card from the verified result — not a paragraph you have to re-read. Cards are built only from committed tool output, and scanned for secrets before they render.
Order #A-2093
- Amount
- RM 120.00
- Refunded on
- 2 Jul 2026
- Items
- 1 × Brand Identity Pack
Cited from approved order records
Refund review opened
- Ticket
- RV-7742
- Requested
- RM 89.00
- Routed to
- Human queue
No money has moved. A person approves before anything is refunded.
The operator cockpit — the same verified state, shown for review and approval.
The fastest way to understand it is to use it
Open the live product and ask it something real, or bring one workflow from your business and we will map it into the agent.