Agentic AI is becoming a new architectural layer within the business because it can sit between people, data, applications, and workflows. It can operate inside processes, reason across information, use approved tools, interact with enterprise systems, and take action within defined boundaries. That is a very different proposition from an AI tool used mainly for answers or content.

I see this as an architectural transformation. Businesses have to decide where an agent can take responsibility, what authority it should have, how its actions are governed, and where people remain accountable.

Begin with work the business already understands

The quickest progress is likely to come from processes that organisations already know well. In areas such as internal IT, customer service, invoice processing, or order management, the volume of work is usually visible and there is already a reasonable sense of how long tasks take, where delays occur, and what a good result looks like. That makes it easier to judge whether introducing an agent is actually improving anything.

There is good reason to be disciplined about this. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls among the reasons. Even though this is a forward-looking market forecast rather than a measured current failure rate, it reinforces the need to define value, cost, and controls before scaling.

I would therefore be cautious about promising a standard ROI timeframe. In the right process, a business may see useful results within weeks or a few months. What counts is whether there was a clear starting point and whether the organisation can show that the process is working better than it did before.

South African businesses are still at very different stages. An ITWeb/Incasu survey of 144 respondents, 76% of whom were at executive or middle-management level, found that 43% were exploring AI possibilities, 24% had operational AI in limited areas, 19% were running pilots, and 13% had integrated AI into core processes. Incasu CEO Matthys Marais described South African businesses as slightly behind their global counterparts. The findings offer a useful snapshot of local sentiment, although the sample should not be treated as representative of all South African organisations. For many local organisations, the next step is to turn experimentation into repeatable use inside real workflows.

Microsoft’s 2025 Work Trend Index describes three phases of AI transformation as “human with assistant”, “human-agent teams”, and “human-led, agent-operated”. It is a useful description of how confidence can build without handing whole processes to agents on day one.

What makes a process ready for an agent

Before an agent can take on part of a business process, the basics need to be in place. It needs reliable information to work from, access to the right systems, and a clear understanding of what it is allowed to do. The business also needs a defined outcome, a measurable starting point, an accountable owner, an exception path, and a way to monitor the agent’s decisions and actions.

Take an order cancellation. A customer sends an email asking for the order to be cancelled and the payment refunded. The agent first needs to establish where the order is in the fulfilment process. If it has not left the warehouse, cancellation may be straightforward. If it is already with a courier, the available options change. The agent has to respond to what it finds in those systems rather than simply work through the same sequence every time.

That is where agentic AI starts to look different from conventional workflow automation. Traditional automation follows predefined rules and paths, while an agent can interpret changing context, select among approved actions, and adapt its next step within defined permissions. Microsoft’s current Copilot Studio guidance describes agents that can respond to events and carry out actions within the permissions and controls set for them. How much freedom they are given should depend on the process they are working in and the consequences if something goes wrong.

Some decisions still need a person

There is little benefit in adding an approval step to every routine action simply because an agent performed it. Updating a record or handling a standard service request may be well within the authority the business has given the agent.

The calculation changes when money, contracts, or regulatory obligations are involved. An organisation may be comfortable allowing an agent to prepare or recommend an action, while still requiring a person to approve it before anything consequential happens.

Human approval remains essential for actions that exceed defined risk, financial, legal, contractual, or regulatory thresholds. The person is there to decide whether an action should proceed, handle exceptions, and take responsibility for the outcome. Governance has to be embedded in the process. Access rights, escalation points, audit history, and security controls should reflect what the agent is allowed to do, while the workflow must also account for situations the agent cannot resolve safely.

The architecture of work is changing

Over the next few years, the distinction between a software user and a software agent is likely to become less clear. Employees will increasingly work alongside agents that carry out substantial parts of processes across ERP, CRM, finance, supply chain, and other enterprise systems.

We may also see agents representing different organisations interacting through APIs within predefined commercial boundaries. For example, a sales agent could respond to a procurement agent under agreed authentication, authority, pricing, and escalation rules, with people stepping in when a threshold is reached or an exception requires judgement.

As agents become part of the business architecture, organisations will need to redesign processes deliberately and give them clear authority, while keeping human judgement where the consequences demand it.