RUFARO MAFINYANI | An enterprise-wide adoption strategy that shows its work

This is the 16th and final instalment of AI Fluency Corner, a weekly series in Business Day building one connected mental model of artificial intelligence (AI) in plain language. This edition sets out the components of a smart enterprise-wide AI adoption strategy, and the discipline that links it to key performance indicators (KPIs) and proves real return on investment.

The Massachusetts Institute of Technology’s Networked Agents and Decentralised AI (Nanda) initiative reviewed more than 300 enterprise generative-AI deployments this year and found roughly 95% produced no measurable effect on the bottom line.

Gartner forecasts that six in 10 corporate AI projects will be abandoned by the end of 2026, for the unglamorous reason that the underlying data was never ready.

Neither statistic describes a technology failing. Both describe a strategy never written down: money spent, a pilot demo-ed, no line connecting the demo to a number the business already tracked.

The South African Revenue Service (Sars) built automated risk selection and refund processing into its core assessment workflow, not alongside it. Capitec’s fraud models are credited with protecting clients from hundreds of millions of rand in attempted losses, measured against what manual checks would have cost to run at the same speed.

Standard Bank’s SmartNudge recommendation engine is not judged on how clever it sounds in a boardroom demo, but on a 66% acceptance rate; its conversational assistant now resolves 65% of digital queries without a human queue. None of these are technology stories. They are KPI stories that happen to use AI.

Treat every AI initiative as a design problem with a financial hurdle, not a software purchase. Empathise and define: name the process, its owner, and the KPI it already sits on a scorecard for — cycle time, cost-to-serve, claims leakage, error rate, conversion.

If that KPI cannot be named before the vendor meeting, the initiative is not ready to fund. Ideate and prioritise: rank candidate use cases by value, data readiness and risk, then choose fewer, deeper bets over a portfolio of disconnected demos.

Prototype for production, not for applause: build the pilot so its data pipeline, governance rules and measurement framework are the same ones that will run it at scale, because a pilot built as a throwaway rarely graduates.

Test against a baseline for a fixed window, typically 90 days, then scale or kill without ceremony — only what clears the hurdle rate earns the next budget cycle.

Most reported AI wins collapse on inspection because they mix activity with outcome. Use one calculation instead: value created equals hours reclaimed per week, multiplied by headcount, multiplied by 52, multiplied by the fully loaded hourly rate — plus errors avoided multiplied by the cost of each error, plus any revenue directly attributable to the AI-enabled step.

Subtract total cost: licensing, integration and data clean-up, and ongoing monitoring.