Agents handle the routine requests.
Your team keeps the hard cases.
Agents answer customer emails, WhatsApp messages and calls from your manuals and your own systems. They close routine cases and hand risky or unclear ones to a person.
From the first message to a closed case.
Four requests at a home appliance maker's service desk, and what the customer and the team see.
A customer emails that the dishwasher shows E24. The agent replies with the fix from that model's manual and names the page.
A customer reports on WhatsApp that the oven stopped heating. The agent checks the purchase date, confirms the warranty and books a technician in the same chat.
A customer wants money back for a dented fridge. The agent prepares the payment, and anything above your limit waits for a team lead.
A caller reports a fault that came back after a repair. The voice agent writes a summary into the case and transfers the call to second-level support.
The routine moves to the agent. Decisions stay with your team.
Today every request goes through a person, even when the answer is in the manual. With agents, your team gets the cases that need a decision, with the groundwork done.
Sorting incoming requests
Today: Someone reads every message and forwards it to the right colleague
With connic agents: The agent answers the routine requests and passes the rest on with a summary
Measure quality, cost and handovers from day one.
Every message the agent handles is recorded with its answer, steps, duration and cost, so these figures need no extra tracking.
Answer quality
A judge scores answers against your criteria, such as the right source and a friendly tone, and alerts you when the average drops.
Cost per case
Your tools tag each run with its case number, and a dashboard chart shows the cost per case.
Share handed to a person
Each handover is tagged on its run, and a dashboard card counts them next to the total.
Time per answer
Each agent's page shows how long its runs take, on average and for the slowest ones.
A person decides what costs money or can't be undone.
You set what the agent does on its own. The rest waits for your team, and every conversation leaves a record.
- Refunds wait for approval
- Name the actions that need a person, always or above a limit such as €100. The run pauses until someone approves, and your team gets an email and an in-app notification.
- Card numbers don't reach the model
- In email, chat and messaging, card numbers and IBANs become placeholders before the agent reads a message. Attempts to talk it out of its instructions are blocked.
- Every conversation has a record
- Each run stores the message, the answer and every step. The audit log shows who approved what.
- Changes are tested before they go live
- Test cases name answers the agent must give and actions it must not take. A release goes live only when every case passes.
One file per channel, one file for the rules.
For your engineers: a connector file links the service mailbox to the agent, and the agent file sets its tools, the approval limit and the guardrails. Both go live with the same deployment.
- Each channel is a file
- A file under
connectors/names the channel and the agents it starts. Each deployment creates, updates or removes it in its environment. - A limit is one line
param.amount_eur > 100pauses only refunds above the limit. Withon_rejection: continue, the agent tells the customer instead of failing.- Payment data is masked first
piiinredactmode replaces card numbers and IBANs with placeholders before the model sees the email.
type: email
mode: inbound
connection: service-mailbox
config:
mailbox: INBOX
agents:
- service-deskversion: "1.0"
name: service-desk
description: "Answers Halden customer emails"
model: connic/gpt-5.6-terra
system_prompt: |
Answer from the manual for the customer's model and name the page.
Hand the case to a person when the manual has no answer.
tools:
- retrieval_query
- orders.find_order
- warranty.check_purchase
- refunds.issue
- cases.hand_over
approval:
tools:
- refunds.issue: param.amount_eur > 100
message: "Goodwill payments over €100 need a team lead."
on_rejection: continue
guardrails:
input:
- type: pii
mode: redact
config:
entities: [credit_card, iban]
- type: prompt_injection
mode: block