Safaricom is using artificial intelligence to reduce the cost of serving more than 40 million customers, but the most important design choice may be the handoff to a human when the automation fails.
The Kenyan telecom group’s FY26 investor presentation says it has more than 70 AI and machine-learning models in production, with AI used across fraud prevention, recommendations, network operations and customer experience. Zuri, its conversational assistant, operates across Telegram, the web, Facebook and the M-PESA app.
That scale makes Safaricom a useful test of what AI customer service looks like when the chatbot is not a demo but part of a national communications and payments platform.
Automation works best on predictable requests
Telecom support contains a large volume of repetitive questions.
Customers check balances, buy bundles, retrieve account information, manage subscriptions and resolve routine service requests. A conversational system can handle many of those tasks without a human agent, at any hour and at much lower marginal cost.
Safaricom explicitly frames AI as a way to lower its cost to serve. That is commercially important because customer-care operations become expensive as a subscriber base grows.
The assistant also creates a consistent front door. Rather than making users understand which internal department owns a problem, a conversational interface can attempt to route the request based on intent.
The problem begins when the system’s confidence is wrong or the issue does not fit a standard workflow.
Escalation is part of the product
In July, Safaricom had to clarify publicly that it had not removed all human agents after a customer complained that Zuri could not resolve an issue and the path to live support was unclear. The company explained that callers can still reach a human after navigating self-service options.
That exchange captures the risk of over-automation.
A chatbot does not need to solve every possible problem. It does need to know when it has failed and make escalation easy. Otherwise, the efficiency gain for the company becomes frustration for the customer.
This is especially sensitive for Safaricom because support can involve money. M-PESA reversals, account access and fraud concerns carry higher stakes than a simple data-bundle query.
The more financial services a telecom operates, the more careful its automation needs to be around irreversible actions and disputed transactions.
AI is becoming a layer across the whole telco
Customer support is only one part of Safaricom’s AI strategy.
The FY26 presentation also describes machine-learning use in fraud and anti-money-laundering controls, personalised recommendations, credit cross-selling, predictive maintenance and network efficiency.
That matters because a telecom has unusually rich operational data: network behaviour, payment history, customer service interactions and product usage. AI can find patterns across those datasets that would be difficult to manage manually.
The risk is governance. A recommendation model can affect what customers are offered. A fraud model can affect whether a transaction is blocked. A credit model can influence financial access.
At that point, model performance becomes a customer-rights issue, not only an engineering metric.
Why it matters
African telecom operators serve some of the continent’s largest digital user bases. If they make AI work at scale, the technology will touch millions of people long before most standalone AI startups reach similar adoption.
Safaricom’s experience shows the opportunity and the restraint required. Routine requests can be automated aggressively. High-friction and high-stakes cases need clear human escalation.
For builders, the lesson is that customer-service AI should be measured on resolution quality, not deflection rate alone. A system that prevents customers from reaching an agent can look efficient internally while damaging trust externally.
What to watch
Watch whether Safaricom publishes clearer measures for Zuri’s resolution rate, escalation time and customer satisfaction, not only the number of AI models deployed.
Also watch how the company applies responsible-AI controls to fraud, credit and personalisation systems. These models have greater consequences than a support chatbot.
The call centre is becoming more automated. The winning design will not be the one that removes humans completely, but the one that uses them where automation stops being reliable.
