Cilliers Naudé, AI Delivery Manager at Retro Rabbit / SmarTek21.
While AI is capable of autonomous action within business workflows, it is still too soon to take humans completely out of the loop. This is according to Retro Rabbit / SmarTek21, which positions itself as a leading software design and engineering consultancy with extensive experience in the banking and insurance sectors.
Cilliers Naudé, AI Delivery Manager at Retro Rabbit /SmarTek21, oversees the company’s enterprise AI implementations and comes from a software engineering background. Naudé confirms Retro Rabbit / SmarTek21 has been delivering enterprise AI solutions in South Africa for over four years and has developed a practical understanding of what guardrails should be in place and where humans should remain in the loop.
He says: “We keep a human in the loop for consequential decisions and authorisations. Fully autonomous AI remains a risk in many processes, so any autonomous task should be tightly scoped. AI can be given permission to perform clearly bounded tasks, including classification, data extraction and routing, depending on the use case. However, approval should still sit with the people responsible. We are comfortable allowing AI to automate routine, reversible work, but when an outcome can affect a customer's: money; cover; or access to a service or legal position, an appropriately authorised person needs to remain accountable for the final decision.”
This viewpoint is endorsed by many leading organisations. One example is the IDC research conducted for Leah. This is an enterprise agentic AI company, which recently reported that 66% of organisations surveyed were already using agents in production, with six times as many deployments planned by January 2027. However, none had a fully autonomous agent in production. Of those surveyed, 46% allowed agents to execute low-risk actions autonomously, while 37% required approval before an agent acted and 9% permitted execution across workflows with audit controls. In Retro Rabbit / SmarTek21’s AI implementations, human oversight, guardrails and visibility into AI actions are part of every solution design.
Naudé adds: “If we give AI enough authority to do useful work and expand its responsibility, we should only do so when the evidence supports it. There cannot be a one-size-fits-all design, and the right level of human involvement depends on the use case. At Retro Rabbit / SmarTek21, we make these decisions by looking at the full business process, not only the model. We work with process owners to understand where time is being lost, where judgment is required, which exceptions occur and what a wrong action would mean for the customer and the business. This gives us a practical basis for deciding where AI can act, where it can recommend and where approval remains necessary.”
He elaborates: “We start with the business problem and the outcome the client needs. We map the current workflow, including judgment points, exceptions, handovers and approvals. We assess the consequence of each action and whether it can be reversed. We then set clear limits on the data, systems and actions available to the AI and identify which actions can run automatically and which require approval, and which remain with a person. We start conservatively and expand authority only when production results support it.”
He notes that Retro Rabbit / SmarTek21 also believes in being honest when AI is not the right answer. “Sometimes a workflow needs better integration, conventional automation, cleaner data or a simpler process. Because Retro Rabbit / SmarTek21 also builds custom software, we use the right combination instead of forcing AI into every step,” he says.
Naudé says visibility is also crucial. “A lot of monitoring goes into our implementations. We design our systems with clear reporting and decision logs so that reviewers can see what the AI did, inspect the source evidence and understand any exceptions. If a business cannot see what the AI did or why it did it, that system should not be given more authority.”
Naudé highlights implementations in which AI is transforming operations while keeping strict guardrails in place and retaining humans in the loop. Smartboxx, Retro Rabbit’s re-usable enterprise AI framework, brings together a broad and growing range of capabilities. These include document and contact centre intelligence, classification, extraction, workflow automation, human review, dashboards, monitoring and audit history, with further capabilities added as new business use cases are implemented.
He cites an implementation for a medical scheme that uses AI to assess e-mail bodies and attachments and route queries to the right department. A contact centre quality assurance solution compares what was said during calls with the contracts and documents subsequently sent to customers and gives the business oversight of 100% of calls. A call and its related quality assurance review, which could previously take around 90 minutes, can now be analysed and reported on in under three minutes. In claims processing, AI can gather and structure the available evidence and prepare a recommendation, while an authorised employee remains responsible for the final decision.
“Our goal is to give AI enough authority to remove repetitive work without losing control of the outcome. The right level of autonomy is one the business can measure, explain and stand behind,” Naudé concludes.
