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

Story audio is generated using AI

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. Design thinking for measurable AI optimisation 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. The metric that actually proves it 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. Divide the difference by total cost and multiply by 100. Value created = (hours reclaimed/week × headcount × 52 × fully loaded hourly rate) + (errors avoided × cost per error) + attributable revenue ROI % = (Value created − Total cost) ÷ Total cost × 100 Take a mid-sized insurer’s claims team: an intake and triage tool reclaims five hours per agent per week across 60 agents. At a fully loaded R320 an hour, that is 5 x 60 x 52 x R320 — just less than R5m a year in reclaimed capacity, before counting fewer reprocessed claims from data entry errors. That number only becomes ROI once someone decides what the five hours buy. Redeployed into the claims backlog and complex dispute resolution, it lowers cycle time and lifts a KPI the executive committee already watches — the compounding path that funds process redesign and, eventually, genuine innovation capacity.Left unassigned on desks that are simply quieter, the same R5m evaporates into payroll and shows up nowhere. The tool worked but the strategy did not exist. Tag each use case by what it actually does — extraction, drafting, retrieval or action — because that decides which KPI moves and how the value should be counted. Where adoption succeeds, fails and goes nowhere Success looks like Sars, Capitec and Standard Bank above: a named KPI, a baseline, a number that moved. Outright failure looks like the shelved proof-of-concept and Gartner’s data-readiness casualties — money spent, nothing shipped. The third category is the one nobody audits: adoption without impact. PwC’s Africa research found 82% of organisations already running AI pilots and 64% of staff already using AI daily, yet the top fifth of companies worldwide still capture 74% of all AI-driven financial returns. Enthusiasm was never the bottleneck. A separate enterprise survey found 79% of organisations still struggling to convert that adoption into enterprise-wide value, with only 29% reporting a significant return despite individual staff reporting five-times productivity gains.Five-times productivity and 29% ROI are not a contradiction — they are proof that hours saved and value proven are two different measurements, and only one of them survives a finance committee. The emerging playbook Anchor every initiative to a KPI that already exists before a tool is chosen. Run a short, honest readiness check on data, process clarity and workforce fluency — that check prevents most of the 95% failure rate, not better prompting. Prioritise fewer use cases and design each one to travel into production unchanged. Decide, before launch, what the reclaimed capacity is for — headcount reduction, faster service or the design-thinking time to redesign the process itself — because that decision is what separates compounding value from a rounding error in the payroll. That is also this column’s argument, stated one final time across 16 editions: the technology was rarely the obstacle. The discipline of naming the number before the tool is bought, and deciding what the freed capacity is for, is what turns AI from a cost centre that eventually gets switched off into a capability that earns its place on the profit and loss statement. Our task this week Name one AI-touched process and its baseline KPI — cycle time, error rate, cost-to-serve or conversion. Calculate the value it should be creating: hours reclaimed × headcount × 52 × fully loaded hourly rate, plus errors avoided. Decide, in writing, what the reclaimed capacity is for — headcount, service speed, or process redesign. Set a 90-day date to measure the KPI again against its baseline. If you cannot complete line three, the tool is working and the strategy still is not. This brings the AI Fluency Corner run of 16 instalments to an end. Thank you for reading them — the KPI outlives the tool that was bought to move it. • Mafinyani is senior partner in financial engineering and AI at specialised finance, risk and applied technology firm Intellica Analytics.Business Day