As AI agents become more autonomous, human accountability becomes more important. In fact, says Pieter du Toit, Head of Marketplace and CSP at Mint Group, without clear responsibility for what agents do – and what they cost – experimentation can become an expensive governance problem.
According to Du Toit, to get real value from AI, businesses must look beyond AI’s productivity promise to the control needed to manage agents as digital workers.
“Access to a system does not automatically mean authority to act with it,” he says. Organisations must define what each agent can do, which actions require human approval and who is accountable for the outcome.
The cost of agent sprawl
Many organisations began their AI journeys by encouraging experimentation. But without an inventory, clear ownership or oversight, teams can build agents for similar purposes, duplicating functionality and adding costs.
Du Toit says this resembles cloud application sprawl, where businesses ended up with several applications performing similar functions. Because agents can be built and moved into production faster, duplication has become harder to spot.
Tools such as Microsoft Agent 365 can provide visibility into agents, ownership, consumption and security – up to a point. But businesses must still decide which agents deliver value and which should be consolidated or switched off.
From licensing to consumption
Du Toit notes that the cost of AI extends beyond licence fees. Consumption-based pricing, model selection, data access and the technology required to run an agent all affect total cost of ownership.
He warns that an inexpensive pilot can become much costlier in production. A proof of concept may include components not charged for during testing, while scaling introduces additional workloads and consumption.
He cites a customer that inadvertently selected the wrong deployment option, turning a cost of around 10 cents per gigabyte into $100. The problem was not the agent itself, but the technology chosen to run it.
Frequent changes to models and pricing add complexity. Organisations may adopt new models without considering the financial implications. Du Toit says FinOps practices, consumption monitoring and collaboration with finance teams are essential to manage costs. Businesses should forecast low, medium and high usage, set spending thresholds and monitor consumption continuously.
Governance before scale
However, controlling spend is only part of the challenge. Organisations must also assess whether agents deliver measurable business value.
“It boils down to not just measuring the agent, but measuring the work it completes,” says Du Toit. Performing a task faster than skilled humans could does not necessarily improve output quality or give a business a competitive advantage.
He argues that human creativity, experience and business-specific knowledge remain critical differentiators for any business. AI makes it easier to generate good looking reports, build applications and automate tasks, but it cannot automatically provide the insight that distinguishes one business from another.
“The differentiator is the information that a person has that is both exposable through experience and through correctly grounded data,” he explains.
Organisations risk confusing productivity with progress. Completing a task faster may save time, but it does not necessarily improve decision-making. Without a clear problem to solve, agents can automate existing processes – and risk multiplying their inefficiencies and costs.
Du Toit recommends appointing one accountable business owner for each agent, supported by IT, security and finance. Organisations should maintain an inventory, classify agents by risk and impact, and establish access permissions, audit trails and approval processes. Compliance with South Africa’s Protection of Personal Information Act (POPIA) must be built in from the outset.
Controls should reflect the consequences of an agent’s actions: the greater its autonomy and potential impact, the stronger the oversight required. Postponing governance can expose sensitive information, create security breaches and leave businesses with agents they cannot prioritise or switch off.
Du Toit advises starting with one or two small pilots, establishing ownership and validating business value before expanding.
The principle is straightforward, he says: build a secure foundation, prove the value of individual use cases, then scale. Without that discipline, organisations risk multiplying not their returns, but their costs and complexity.
Share
