Telecommunications · chat triage

The national telecommunications carrier of a small island nation.

A thousand customer chats a day, moved through the day instead of building. Agents open each chat to a pre-built brief, not an empty draft.

Mia is our AI employee. Email her — she’ll book your 15-minute call. That’s the demo.

Telecommunications · chat triage

The national telecommunications carrier of a small island nation.

A thousand customer chats a day, moved through the day instead of building. Agents open each chat to a pre-built brief, not an empty draft.

Volume

~1,000 chats/day

Across the carrier’s primary channel

First-response time

Minutes → under 30 sec

Through the day, including peak

Autonomous resolution

60–70% containment

With re-contact rate held flat

Agent-side AHT

15–25% faster

On the cases that reach a human

The open loop

A thousand customer chats a day across a small, dense market. Most of them were the same handful of questions — bill enquiries, prepaid top-ups, plan changes, roaming, SIM activations, outage check-ins. Each one looked routine until you’d answered it for the eighth time that morning.

Agents handled what they could during the day and the evening peak built into a queue that the overnight shift inherited. By the weekend the queue had texture — actual incidents tangled up with sixty repeat questions about the same evening’s mobile data top-up.

Day 1–2 · Named employee

A chat triage officer.

Scoped to first-touch chats only. No billing changes, no service credits, no plan adjustments, no enterprise accounts.

Connections

  • Customer account database (read-only)
  • The chat platform
  • Knowledge base and standard reply library
  • Outage and incident status feed
  • Escalation routing rules
First-pass scope
  • Read every incoming chat and classify against the top six categories
  • Pull account context — plan, recent usage, last billing event, last service activity — before drafting
  • For routine categories with a clear answer, draft the reply in house tone
  • For service credits, complaints, enterprise accounts, or outage-impact compensation, route with one-line context
  • Surface to the agent with the draft pre-loaded
Week 2 · what got tuned

The outage pattern got missed. During the first incident in week one, the employee was confidently answering “your service should be working from your side” to a cluster of customers whose service in fact wasn’t. The reply was technically correct against the account record and exactly wrong against reality.

The rule changed: any cluster of three or more similar-pattern messages within a ten-minute window triggers a hold on outbound drafts and a flag to the network operations queue. The outage status feed got wired into the employee’s context the same day.

What humans own
  • Billing adjustments, credits, refunds
  • Plan and service changes
  • Enterprise and B2B accounts
  • Complaint resolution and any escalation
  • Anything during an active outage — the wrong autonomous reassurance at the wrong moment is the worst-case interaction
What the employee owns
  • Reading the chat and classifying it
  • Pulling account context before drafting
  • Drafting the routine reply in the right voice
  • Detecting outage patterns and pausing drafts
  • Routing edge cases with reasoning attached

The deployment sits in the band published comparable telecom AI cases describe: first-response times under thirty seconds during peak, 60–70% of triaged chats resolved without an agent typing, and 15–25% faster handle time on the cases that do reach a human — with re-contact rate held flat as the honesty check on resolution.

What the employee deliberately doesn’t do
  • Touch billing. Adjustments, credits, refunds — all human.
  • Change a plan or service. The employee can describe options; only an agent makes the change.
  • Handle enterprise accounts. B2B routes immediately, no exceptions.
  • Speak during an active outage. Outage-pattern detection holds drafts and routes to a human.
  • Send without review. Every customer-facing draft passes through an agent.
What the team does now

Agents now open each chat to a pre-built brief — customer context, recent activity, draft reply already prepared for the routine categories, edge cases pulled into the right queue with a summary attached. The evening peak still happens — it always will — but it clears inside the evening instead of waiting for the next morning. The work that needed a person — the complaints, the outage conversations, the enterprise escalations — is the work agents are actually doing.

A moment

Chat #2814 — prepaid data top-up question, account in good standing, drafted from top-up template. Held for Aishath’s approval.

Cluster detected — 4 messages re: data signal in last 6 mins. Drafts on hold. Flagged to NOC. No outbound.

48-HOUR START

Bring us one workflow. We will name the employee in 48 hours.

Email Mia

Mia is our AI employee. Email her — she’ll book your 15-minute call. That’s the demo.