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. 鈥淭he agent will follow its instructions to the letter,鈥� says Jan van Niekerk, senior vice president: Data and AI strategy at Netsurit.

鈥淚f the instructions were never tied to strategy, that obedience is the problem, not the safeguard.鈥� What gets automated in those conditions is what Van Niekerk calls 鈥渧anity work鈥�, the visible tasks that show the organisation is 鈥渄oing鈥� 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. 鈥淢ost 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,鈥� 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. 鈥淗ours saved is a vanity metric unless the capacity is redeployed,鈥� says Van Niekerk.

His advice is to baseline the process first, measuring the capacity that鈥檚 freed up, and then tracking what it was redirected to, such as more customers served, a cost line reduced or increased revenue. 鈥淚f all you can report after 90 days is 鈥榯ime saved鈥�, you haven鈥檛 proven value, you鈥檝e proven activity,鈥� he says.

Nazia Pillay, managing director, 麻豆原创 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. 鈥淪tarting with the outcome, rather than the technology, makes it easier to demonstrate value and decide where to expand next,鈥� 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. 鈥淎gent readiness isn’t a property of the process, it’s the output of having mapped it, where needed, and redesigned it first,鈥� 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. 鈥淲here there is a decision that needs to be made, and it’s an explainable decision, that’s where agents really work well,鈥� says David Pretorius, customer advisor manager at SAS.

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. 鈥淭he 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,鈥� says Pretorius. 鈥淚t should have a bounded blast radius.鈥� 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. 鈥淭he 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,鈥� says Pillay.

Agent density is already a capacity metric in infrastructure planning, but Pretorius says that the enterprise version will rather count agents against people. 鈥淏y 2030, the question will be how dense is your agent landscape,鈥� he says. 鈥淎nd how does that ratio sit with regards to the amount of people you have, and what is the payoff rate?鈥� Getting to this kind of number means knowing what each agent is for, and who owns it, which most enterprises haven鈥檛 yet considered. This is why Van Niekerk鈥檚 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. 鈥淚n a business of 10 000 people running 10 agents each, agent lifecycle management becomes as fundamental as identity management is today.鈥�

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鈥檚 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鈥檚 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鈥檚 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鈥檛 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. 鈥淗umans are ultimately accountable. You just have a better tool,鈥� he says.

This article first appeared in .

The post Bloody agents appeared first on 麻豆原创 Africa News Center.

Emmanuel Raptopoulos, Global President of Customer Success for Europe, Asia-Pacific, the Middle East and Africa at 麻豆原创, spoke to ITWeb following the recent 麻豆原创 Connect 2026 event in Las Vegas.

鈥淵ou need enterprise context to make meaningful decisions from information held in your systems. That is the foundation for moving from reactive towards predictive operations,鈥� said Raptopoulos.

He added that organisations need an AI layer anchored in their operational context to respond faster and improve predictions. This involves more than data; it also encompasses how a company operates, makes decisions and uses its transactional systems.

He believes Africa could use emerging technologies to bypass some stages of technological development.

“We have seen technologies emerge in the past that allow markets to leapfrog previous stages of development. You do not necessarily have to go through every preceding step. I think Africa has another opportunity to leapfrog from a technology perspective. With an agentic approach, there is potential to jump a stage and apply the technology to areas such as transparency and efficiency in the public sector, citizen trust, access to public services and education,” said Raptopoulos.

He said there is a lot of opportunity, but guardrails are extremely important.

“We should continue to apply judgment. Human judgment is here to stay. Governance, security and clear accountability are essential to managing autonomous systems safely.”

Risks of AI agents

Raptopoulos acknowledged that AI agents could behave unpredictably even when rules and controls are in place, including measures to ensure they are task-specific and appropriately prompted.

鈥淭here is always a possibility. Our role as a vendor, together with our customers and their governance frameworks, is to control that probability. If you respect the design principles and put appropriate governance around the technology, you can reduce the probability significantly. Cyber security is also evolving alongside AI capability. The same underlying technological capabilities can be deployed defensively,鈥� he said.

鈥淵ou therefore need the right governance internally while heightening external defences to create a shield around your systems. The objective is to prevent this becoming an issue in future.鈥�

Dr Pierre le Roux, MD of digital consultancy MOYO, said AI is moving from technology that tells people what they can do to technology that can increasingly do it for them.

That shift could pose a major governance challenge as businesses move from generative AI and copilots to autonomous AI agents that can access systems, initiate transactions, communicate with customers, change data and execute business processes, said Le Roux.

鈥淯ntil now, much of the AI conversation has been about the accuracy of the answer,鈥� he added. 鈥淎gentic AI changes the question. If the technology is able to take an action, the issue is no longer simply whether the answer was right or wrong. You have to ask who gave it permission to act, what it was allowed to access, where its authority ended and, ultimately, who is accountable for the outcome.鈥�

Le Roux said the issue is becoming more urgent as businesses move beyond AI experimentation. He cited a Gartner prediction that by 2027, 40% of enterprises could demote or decommission autonomous AI agents because governance gaps are discovered only after deployment.

鈥淎I may execute the decision, but accountability remains with the organisation that designed the process, gave the agent access and determined its authority. The question is no longer whether you trust AI, but whether your organisation has engineered the boundaries within which AI is allowed to act.鈥�

Raptopoulos said companies were increasing their investment in AI and integrating it into their operations, but questions remained about their ability to control agents operating within their systems.

鈥淎wareness has risen. Recent high-profile events have helped make the issue visible even beyond people who are highly technically literate,鈥� he said.

鈥淭he concern is that you do not want agents moving around enterprise systems in ways that could make data available to the wrong person. That is a potential boardroom nightmare. One important design principle is to make agents very specific about what they are supposed to do and give them clearly defined objective functions. An agent is ultimately an application trying to fulfil an objective. If that objective is defined too broadly, it has greater scope to find unintended ways of fulfilling it. Agents should therefore be specific to the task at hand and grounded in the enterprise objective.鈥�

This article first appeared on .

The post 麻豆原创 says Enterprise Context is Key to Predictive AI Operations appeared first on 麻豆原创 Africa News Center.