By Gleb Tsipursky, CEO of Disaster Avoidance Experts.
Africa’s telecom operators are moving artificial intelligence out of the demo room and into the systems that customers and network teams rely on every day. Airtel Africa’s 2026 strategy calls for scaling AI-led customer engagement and workflow automation, whilst operators across the continent are expanding digital financial services, cloud infrastructure and data-driven network operations. Airtel Money’s planned London listing this week is another reminder of the scale now attached to African digital services. When automated systems touch millions of customer interactions, the question is no longer whether AI will be used. It is what happens at the moment the machine reaches the edge of its judgement.
That moment needs a standard.
The telecom industry has spent decades engineering redundancy into physical networks. Links fail, power drops, equipment overheats, traffic spikes and systems reroute. AI needs the managerial equivalent of that resilience. GSMA has warned that telecom is an unusually difficult environment for AI because networks are fragmented and have little tolerance for error. African Wireless Communications has already highlighted the infrastructure demands of autonomous AI. The next step is to treat human handoffs as infrastructure too.
A human handoff standard should define, before deployment, when an AI system must stop acting and transfer a decision to a person. It should also define what information arrives with that transfer. A red button is not enough. If the human operator receives an alert without the conversation history, the system changes already made, the uncertainty that triggered the escalation and a clear rollback path, the organisation has created an override mechanism that looks reassuring on a slide but fails under pressure.
Customer service is the easiest place to see the problem. A chatbot can answer routine questions quickly. It can also encounter a billing dispute involving several products, a suspected fraud case, a vulnerable customer, a service outage or a request that does not fit the script. The handoff should not simply say “agent needed.” It should carry a compact decision packet: what the customer asked, what the AI understood, what it already changed or promised, what evidence it used, what remains uncertain and what the employee is now being asked to decide.
That protects the customer from repeating the story and protects the employee from inheriting a mystery.
Network operations need the same discipline. As anomaly-detection systems and AI agents move from recommending actions towards executing them, operators should separate reversible, pre-approved actions from service-impacting changes. An automated system might be allowed to rebalance traffic within a defined envelope, whilst a configuration change that could affect availability requires human authorisation. The important point is not to force a person into every loop. It is to decide which loops require a person before the incident occurs.
Mobile money raises the stakes further. When AI assists with fraud detection, identity checks, customer support or transaction review, speed matters, but so does contestability. A customer whose transaction is blocked should be able to reach a human path that can review the evidence and correct a bad decision. The escalation should preserve the record of what the system saw and why it acted. Otherwise, the human reviewer is left to reconstruct a decision after the fact, which slows resolution and weakens accountability.
A useful handoff standard would also name an owner for each class of automated decision. Too many AI deployments assign responsibility to “the business,” “the model team” or “operations.” Those labels disappear the moment something goes wrong. Someone should own the threshold for escalation, someone should own the quality of the handoff packet, and someone should review repeated overrides to determine whether the system, the process or the policy needs to change.
That last step is where handoffs become a source of learning rather than a tax on automation. Every override is evidence. If staff repeatedly reverse the same type of AI recommendation, the organisation has found a pattern worth fixing. If people almost never intervene, the escalation threshold may be too conservative. If employees approve suggestions without examining them, the problem may be workload or interface design rather than the model itself.
The strongest operators will therefore measure more than adoption and response time. They will track how often AI decisions are escalated, how often humans reverse them, how long recovery takes, which categories generate repeated exceptions and whether customers receive a coherent explanation. Those measures reveal where automation is creating real capacity and where it is simply moving hidden work downstream.
Africa’s connectivity market has strong reasons to automate. Networks must serve fast-growing demand across difficult geographies, uneven infrastructure and a wide range of customer needs. AI can help operators use scarce expertise more effectively. But the faster the systems act, the more important it becomes to know where judgement changes hands.
The goal is not human oversight as a slogan. It is a practiced transfer of responsibility. Telecom engineers already understand that resilience depends on what happens after a component fails. AI resilience will depend on what happens after the model hesitates, misreads the situation or crosses a boundary it should not cross.
Build that handoff before the next layer of autonomy arrives, and African operators can move faster because employees and customers know where the brakes are.
