Agentic AI: what autonomous agents mean for African business

Many AI tools answer questions or draft text. Tool-using agents can also interact with other systems, which introduces both useful possibilities and additional risks.

For example, an agent could retrieve an order and prepare a refund request. Actual refunds should pass deterministic policy checks and approval controls appropriate to their risk; a model's output is not authorisation.

What agents can do today

Start with bounded tasks such as drafting reports or triaging documents. Feasibility depends on data access, integration quality and error tolerance. Measure accuracy, exceptions and review effort in a pilot before expanding access.

• Customer service: resolving account queries end-to-end on web and WhatsApp
• Finance: three-way matching, reconciliation and exception queues
• Operations: compiling daily reports from multiple systems before the morning meeting
• Sales: qualifying leads, drafting proposals and chasing follow-ups

The safety question is the design question

A useful design question is: what stops an agent doing something expensive? The answer is that safety is designed, not hoped for. Well-built agents run with least-privilege access, hard spending and action limits, mandatory human approval for defined categories, and audit logs of every step.

Treat an AI agent like a capable new employee: clear job description, limited authority, supervised probation, and expanding trust earned through review.

That framing changes procurement conversations. An agent remains software: permissions, monitoring, rollback and accountability must be designed around its limitations.

Where to start

Pick one process that is high-volume, rule-based and measurable. Map it honestly, including the exceptions. Pilot the agent in shadow mode against historical cases before it touches production. Then widen its mandate as the review record justifies. Compare the pilot with a simpler rules-based workflow and continue only if the measured benefit justifies the risk and cost.