Case Study · Telecom
A bilingual telecom chatbot that resolves more than half of all chats
A leading Middle East telecom operator wanted customers to get answers on WhatsApp and Facebook Messenger instead of waiting for the call centre. AMT built an English and Arabic AI chatbot that handles around 100,000 conversations a month and resolves more than 50% of them without a human agent.
Schedule Technical DiscoveryYour first conversation will be with a senior technical partner, not a salesperson.
Published
At a glance
- Client
- Leading telecom operator offering fixed line, mobile, internet and digital TV across the Middle East
- Industry
- Telecom and communications
- What AMT built
- AI chatbot with its own personality, live on WhatsApp, Facebook Messenger, web portal and mobile apps
- Languages
- English and Arabic
- Live since
- 2019
- Results
- More than 50% of chats resolved by the bot. Around 100,000 conversations a month
- Stack
- Angular, Node.js, IBM Watson Assistant, Watson Natural Language Understanding
The short version
The operator's customers already lived on WhatsApp and Facebook Messenger, so that is where the support had to be. AMT built a chatbot that understands more than 500 questions in English and Arabic, each one in more than 40 different phrasings. It handles bill payments, roaming activation, recharges, order tracking and account registration, and it passes customers to a live agent when it should. Since launch in 2019 it has taken a large share of routine queries off the call centre and service centres.
The problem
Routine questions were landing on human agents. Bill payments, recharges, roaming and order status are simple to answer, but at telecom volumes they keep call centres busy and customers waiting.
The operator had three clear requirements:
- Meet customers where they already chat. WhatsApp and Facebook Messenger are widely used in the UAE, so the bot had to live there first, then on the web portal and mobile apps.
- Sound like a brand, not a menu. The company wanted a jovial, witty bot with a name and character traits, not a computer responder.
- Learn what customers feel. Every conversation needed sentiment analysis so the business could see how customers were reacting.
What AMT built
A bot with a personality
The chatbot has its own name and personality traits, agreed with the operator. Its tone is consistent across every channel, so a customer gets the same character on WhatsApp as on the website.
Language understanding that holds up
The bot handles more than 500 general questions in English and Arabic. Each question was trained in more than 40 ways of asking it, so the bot still understands a customer who phrases things loosely, mixes words or skips context.
FAQ and transactional queries in one place
Customers can ask general questions and complete real tasks in the same chat:
- Bill payments
- Roaming service activation
- Online order tracking
- Mobile recharge
- Account registration
Two-factor authentication for anything transactional
Transactional queries go through two-factor authentication, which identifies and verifies the customer before the bot touches their account.
Knowledge base, live agents and CSAT
The bot draws answers from an integrated knowledge base. When a conversation needs a person, it hands over to a live agent. After the chat, it triggers a CSAT survey so the operator can measure satisfaction.
Sentiment analysis that feeds marketing
Watson Natural Language Understanding scores the sentiment of each conversation. The operator uses those insights to plan targeted offer campaigns, so the support channel also tells marketing what customers want.
Results
| Measure | Result |
|---|---|
| Chats resolved by the bot | More than 50% |
| Monthly conversations | Around 100,000 |
| In production since | 2019 |
| Effect on support teams | Fewer queries reaching call centres and service centres |
| Customer satisfaction | Good CSAT scores, measured by the in-chat survey |
Technology stack
| Layer | Technology |
|---|---|
| Front end | Angular |
| Back end | Node.js |
| Natural language understanding | IBM Watson Assistant |
| Sentiment analysis | Watson Natural Language Understanding |
| Channels | WhatsApp, Facebook Messenger, web portal, mobile apps |
Before you build something similar
If you run support for a telecom operator or any high-volume consumer business, settle these first:
- Which queries are transactional? Anything that reads or changes an account needs authentication inside the chat. Plan it before you design the flows.
- Which channel do your customers actually open? Launch on that one first. A bot on the wrong channel gets no traffic, however good it is.
- How many ways do people ask the same thing? Collect real phrasings from your chat logs and call notes. Training on a single phrasing per question is why most bots feel dumb.
- Where does the bot stop? Define the handover to a live agent, and make it fast.