The agent in the room nobody hired

Autonomous software now acts, transacts and speaks on our behalf, and responsibility cannot be handed to a system that cannot be held to account.

Sign up for my Substack daily AI newsletter here.

Article link: https://johanosteyn.substack.com/p/the-agent-in-the-room-nobody-hired?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

At a recent industry gathering, one of South Africa’s most experienced telecommunications figures, Andile Ngcaba, told delegates that artificial intelligence agents would supersede humans on the internet before 2030, and that the number of living people would soon be a fraction of the number of agents operating online. It is a bold claim from someone with a stake in the industry, and the exact date matters less than the direction of travel. Software that can plan, decide and act on its own is moving from the margins to the mainstream of how business gets done. The question that should occupy boards is not when the machines outnumber us, but a simpler and harder one. When an agent acts in your name, who is accountable for what it does, and can you prove a human was ever in the loop?

CONTEXT AND BACKGROUND

Much of the forecast has already arrived. Automated traffic has overtaken human traffic on the open web, with one network company, Cloudflare, putting bots at 57.5% of requests against 42.5% from people, a crossover its chief executive admitted came faster than he had predicted. A separate industry briefing went further, suggesting that for every human online there are now 144 agents, and that the systems built to manage online activity were never designed for machines at that scale and speed.

These numbers deserve a caveat. Counting requests is not the same as counting attention, and people still dominate the hours spent streaming and scrolling. Much automated traffic is also malicious rather than useful. The trend is real even so, and it marks a shift that South African leaders have barely begun to price in.

INSIGHT AND ANALYSIS

Our online world was built on an assumption that is now breaking. Systems for voting, reviewing, commenting and transacting assumed a person behind the screen. As agents act autonomously, that assumption fails in two ways. One is the flooding of shared spaces by fake accounts under a single operator, and the other is the blurring of whether an agent is acting as a person, for a person, or instead of one, leaving no clear human to hold responsible. The accountable person, long taken for granted, must be designed back in.

I have previously written about this, arguing that South Africa did not need a new law for agentic AI so much as a careful reading of an old one, because our electronic transactions legislation already treats a message generated by an automated system as the responsibility of the party it acts for, even with no human review. The law, in other words, is ahead of most boardrooms. What is missing is control inside the business.

That gap is not hypothetical. Security researchers report that 90% of AI agents operate with permissions broader than they need, and that 80% of organisations have already seen agents take unintended actions, which is why the emerging advice is to treat every agent as potentially compromised. An agent with standing access to systems and money is an employee without a contract, a manager or a disciplinary record.

This is not new. Businesses have delegated to automated systems for years, from algorithmic trading to software that runs without a manager approving each transaction. What changes is the open-ended autonomy, the breadth of access and the speed. The answer is not a human approving every action, which would stall the benefit, but bounded authority and clear accountability.

IMPLICATIONS

For boards and executives, the task is to govern the agent before it governs the risk. That starts with knowing which agents are deployed, who authorised them, what they may spend and touch, and whether every action leaves an audit trail naming a responsible human. The technology that lets agents plug into other services is itself a new attack surface, and even national security agencies now warn that standard controls do not fully address agentic risks and that adopters should proceed with caution.

None of this argues against the technology. Used well, agents promise real gains, and economists estimate that artificial intelligence could lift aggregate labour productivity by as much as 3.8% over the long term, with the strongest returns where it complements skilled people rather than replacing their tasks. The opportunity for South Africa is to adopt quickly while keeping humans as net contributors rather than spectators, investing in the unglamorous infrastructure of trust, identity and oversight at the pace of the tools themselves.

CLOSING TAKEAWAY

The forecast that agents will outnumber us is less a prediction than a warning about design. We are building an internet in which much will be done by software that acts on our behalf, and the one thing we cannot automate is responsibility for what it does. Cognitive governance means keeping a named, accountable person in every loop that matters, with the authority to intervene and the record to answer for it. The agent in the room was never hired, holds no mandate and bears no blame. Our task is to make sure that someone always does.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net