According to Gartner, more than 40% of agentic AI projects will be cancelled before the end of next year. One of the main reasons, says the research house, is unclear business value. Organisations are deploying autonomous systems because of FOMO and without defining measurable outcomes. If anything, agents get pointed at whatever is loudest in the organisation, which is often an executive with a preference. “The agent will follow its instructions to the letter,” says Jan van Niekerk, senior vice president: Data and AI strategy at Netsurit.
“If the instructions were never tied to strategy, that obedience is the problem, not the safeguard.” What gets automated in those conditions is what Van Niekerk calls “vanity work”, the visible tasks that show the organisation is “doing” AI without changing what it costs to run.
But is any of this a technology problem? South African enterprises license the same frontier models as everyone else and pre-built agents now ship for processes that have been documented for decades. Van Niekerk says enterprise AI maturity runs through four stages, from personal AI, to workflow AI, and process AI, to being AI-native. “Most South African enterprises are stuck at personal AI, where a tool gets handed to everyone and the world is expected to change, with no P&L impact to show for it,” he says. He adds that markets like the US, the UK and parts of the Gulf have more enterprises operating at process level, mapping and redesigning end-to-end work before deploying agents and assigning ownership of the value.
Agentic AI promised to free staff up for other tasks, which is easy to report, but essentially meaningless on its own. An agent that removes six hours of work a week from a finance team has changed nothing until somebody decides what those hours can be used for. “Hours saved is a vanity metric unless the capacity is redeployed,” says Van Niekerk.
His advice is to baseline the process first, measuring the capacity that’s freed up, and then tracking what it was redirected to, such as more customers served, a cost line reduced or increased revenue. “If all you can report after 90 days is ‘time saved’, you haven’t proven value, you’ve proven activity,” he says.
Nazia Pillay, SAP
Nazia Pillay, managing director, SAP Southern Africa, says measurable improvement should be evident within months when an enterprise picks a well-defined process. The quickest payback comes from work that is repetitive, time-consuming and governed by clear rules. “Starting with the outcome, rather than the technology, makes it easier to demonstrate value and decide where to expand next,” she says.
Van Niekerk says there are a few signs to look out for if a process is not ready for an agent. One would be if staff members can't describe it end-to-end, without contradicting one another. Another is if the exceptions outnumber the rule, or the handoffs between people and systems are undocumented and the real knowledge is in somebody's head or inbox. Lastly, if there is no single owner or baseline metric. “Agent readiness isn't a property of the process, it's the output of having mapped it, where needed, and redesigned it first,” he says. A great example of this is credit decisioning and fraud monitoring, where the rules are already written down, which is why banks have deployed agents here first. “Where there is a decision that needs to be made, and it's an explainable decision, that's where agents really work well,” says David Pretorius, customer advisor manager at SAS.
Source: McKinsey, The State of AI 2025 Global Survey, November 2025
Pretorius believes one problem enterprises are least prepared for is agent containment, or defining the limits on what an agent can reach, change or break once it starts acting on systems rather than answering questions. An agent that reasons, retries and chains calls can act on systems. “The model is the engine of the vehicle and the harness around it is the steering, the brakes, the dashboard, the seatbelts, everything around it, keeping it contained,” says Pretorius. “It should have a bounded blast radius.” If a deployment causes problems, it should be cheap to fix and quickly reversible, which implies project management rather than model selection. This is also true of policy, because a document written for chatbots will have little defence against agents acting autonomously. “The objective is not to require human approval for every routine action, but to establish clear thresholds so that agents can handle predictable work efficiently while people retain control over sensitive decisions, exceptions and consequences,” says Pillay.
Source: End User Analysis, 2025-2035
Agent density is already a capacity metric in infrastructure planning, but Pretorius says that the enterprise version will rather count agents against people. “By 2030, the question will be how dense is your agent landscape,” he says. “And how does that ratio sit with regards to the amount of people you have, and what is the payoff rate?” Getting to this kind of number means knowing what each agent is for, and who owns it, which most enterprises haven’t yet considered. This is why Van Niekerk’s recommendation is to treat agents as organisational assets, with ownership tied to a role.
He says agents need to be registered, and their purpose, permissions and prompts documented. “In a business of 10 000 people running 10 agents each, agent lifecycle management becomes as fundamental as identity management is today.”
BETTER TOOLS, SAME RULES
FNB parent FirstRand went from dabbling with agentic AI to scaling it inside six months, says Kevin Mitchell, chief digital officer. Project scoping took 30 minutes, compared to two weeks.
No exceptions
Every control built up over decades of regulated delivery still applies once agents enter the pipeline, he says. Segregation of duties, risk review and release management all need to be included in the production workflow. All code, whether written by a staff member or generated, must still be reviewed before it’s committed.
Keep the estate closed
Pay attention to data flows; a tool that sends customer information to an unintended location will be flagged by risk and compliance. Running agentic tooling in a closed environment isolates the prompts, code and data. Mitchell says the data must be kept under the bank’s jurisdiction. Usage telemetry will also show who is actually working with the tools, and for what purpose.
Build it into the SDLC
An AI program that runs alongside a development process should be included in the development lifecycle at the outset. Mitchell says AI agents should span ideation, product development, risk, legal and operations, rather than only being used by a single team. Bring the risk and legal people in early, he advises, because teams that meet an agent for the first time at go-live may become a bottleneck.
Tie tokens to outcomes
Consumption dashboards will tell you an agent is busy with a task, but not whether it’s providing value for money. Mitchell suggests linking token spend to specific backlog items and features. He says the bank is already tracking these metrics, including time-to-production.
Plan for the sceptics
The appetite for agentic usage inside a large engineering team will not be uniform, and a rollout should assume that enthusiasm will wane. One group at FNB queued for access immediately, while another team said they didn’t trust a coding agent to fulfil tasks as well as they could. The latter group needs evidence rather than encouragement, says Mitchell. The argument he makes is that the tooling has changed, not the responsibility. “Humans are ultimately accountable. You just have a better tool,” he says.