The AI pilot works. Does the business case?

The timing of the sales call is not an accident. The global technology industry has sunk hundreds of billions of dollars into data centres and model training, and every vendor in that chain is now under pressure to show enterprise revenue on the other side of the capital spend. Having saturated the banks of New York and London, the sales teams have arrived in Nairobi, Lagos and Johannesburg with the same deck. Walk into almost any bank boardroom in Kenya this quarter and you will see it: a triumphant slide showing how a generative AI pilot shaved two minutes off customer-support calls. On the demonstration screen the technology looks effortless, yet inside the operating accounts the unit cost of serving that customer has barely moved. Financial institutions are developing a costly habit of confusing a faster conversation with a cheaper operation, signing enterprise-wide licences before asking whether the minutes saved ever reach the income statement. A 2025 review of enterprise generative-AI deployments by researchers at the Massachusetts Institute of Technology reported that about 95 per cent of corporate pilots produced no measurable return on the profit and loss account, not because the models failed, but because the organisations around them never rewired the workflow that turns saved minutes into saved money.

The most careful evidence available says the gains are real but unevenly distributed. A 2025 study published in the Quarterly Journal of Economics by Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,172 customer-support agents after a software firm introduced an AI assistant. Issue resolution rose by an average of 15 per cent per hour, but the headline hid a sharp split: less experienced agents captured almost all of the gains, while top-performing veterans gained virtually no speed and saw a slight dip in quality. Buying an identical monthly licence for every seat in the call centre therefore means paying a recurring fee that slows down your best people while subsidising a training shortcut for novices. For a Kenyan chief executive preparing the 2027 technology budget, a promising pilot never justifies a blanket rollout. It forces a hard capital-allocation choice between equipping new hires and fixing broken back-office settlement pipes.