When businesses talk to me about agentic AI, the conversation usually starts with efficiency. They want to know whether an agent can take enough repetitive work off an overstretched team to justify the investment, and how soon that value will show.
However, in many organisations, adoption is being driven less by a defined AI strategy than by cost pressure and labour constraints. In South Africa, skills scarcity adds to that pressure because some technical and administrative roles are difficult to fill. Boards are also asking what the business is doing about AI before the problem it should solve has been properly defined.
Pressure is accelerating adoption
Automating the repeatable parts of hard-to-fill roles can be sensible, provided the work is suitable for an agent. High-volume processes with clear inputs and outputs are usually better candidates than work filled with open-ended judgement calls.
If nobody in the business can explain how a task is completed, including what happens when something unusual occurs, the process is not ready for an agent. It needs to be mapped first. I look for processes where the path is well understood and the exceptions are known. That gives the business something concrete to test.
The economics sit in the exceptions
The quickest returns tend to come from repeatable work such as first-line support triage or document and data processing. For a narrow, well-scoped deployment, I would consider three to six months a realistic timeframe to start seeing a return.
Once an agent has to work across several systems or handle more nuanced decisions, 12 to 18 months is a more honest expectation. Organisations often underestimate how much effort goes into handling exceptions.
The harder part usually starts when the process does not behave exactly as expected. Information may be missing, a request may not fit the normal pattern, or the agent may reach a point where the business has not clearly defined what should happen next.
That is why testing only the straightforward cases is not enough. The business needs to see how the agent handles the messy situations that happen in real operations.
Keep people where the consequences are real
Human oversight should depend on what happens if the agent gets something wrong. A low-risk task that can be easily corrected does not require the same level of supervision as a decision that affects a payment, a compliance obligation, safety, or a client relationship.
That is the distinction I would make. Agents can handle routine work and prepare information, but once the consequences become significant or difficult to reverse, a person must remain accountable for the decision.
Low-consequence, repetitive work can run with less supervision. Over-supervising trivial activity destroys much of the value of automation, while giving an agent too much freedom around consequential decisions creates obvious risk.
Governance needs to exist before rollout
A sensible deployment should begin with a shadow period. The agent runs alongside a human and produces an output, but nothing is acted on automatically. Once the business is comfortable with the results, lower-risk work can be handed over gradually while unusual cases continue to escalate.
The boundaries should already be clear. The business needs to define what the agent may do and when it must escalate. Its actions also need to be logged so that someone can review what happened.
If an automated action causes a problem, “the agent did it” is not an acceptable answer. Someone must remain responsible for the process and for deciding how much autonomy the system receives.
Automation will change the human role
Over the next five years, I expect agents to take on more of the repetitive workload, but businesses will still need people to oversee what those systems are doing. Employees may spend less time completing routine tasks and more time reviewing exceptions, checking outcomes, and stepping in when a decision needs judgement.
As more work is automated, that oversight becomes increasingly important. Someone still needs to understand the process, recognise when something has gone wrong, and remain accountable for the outcome.
That changes the human role, but it does not remove accountability. The more work agents handle, the more clearly organisations will need to understand who owns the outcome when something goes wrong.
Businesses that approach agentic AI first as an accountability and governance problem will be better positioned to extract value from it. Those that automate work without redesigning responsibility may simply find themselves making the same mistakes faster.
