For most of the past decade, artificial intelligence has worked the way a calculator works: you ask, it answers, and then it waits. In 2025 that pattern began to change. A new generation of systems, often called AI agents, can take a goal, break it into steps, use software tools, and carry the work through to completion with limited supervision. The shift from answering questions to finishing tasks is one of the most significant changes I have observed in my career, and it is already reshaping how organisations around the world operate.
The clearest early examples are in software development. Agents can now read a code base, plan a change, write the code, run the tests, and open a request for human review. Engineering teams that once measured AI assistance in autocompleted lines are starting to measure it in completed tickets. Similar patterns are appearing in customer operations, where agents resolve routine billing and logistics enquiries end to end, and in research, where agents search the literature, extract findings, and assemble structured summaries that once took analysts days to prepare.
Businesses are drawn to agents for three reasons: speed, consistency, and cost. An agent does not queue work, does not tire at the end of a shift, and applies the same procedure every time. In logistics, agents reschedule shipments around weather and port delays. In finance, they monitor transactions and escalate anomalies. In healthcare administration, they chase referrals, prepare discharge summaries, and reduce the paperwork burden that contributes so heavily to clinician burnout.
The risks, however, scale with the autonomy. An agent that can act can also act incorrectly, and a mistake executed automatically can propagate before a human notices. Security researchers have shown that agents can be manipulated through the very documents and websites they are asked to read, a class of attack known as prompt injection. There are also questions of accountability: when an autonomous system books, buys, or communicates on behalf of an organisation, who is responsible for the outcome?
My view is that the answer lies in designing for supervision rather than hoping for perfection. Agents should operate inside clear permission boundaries, keep complete logs of every action, and hand control back to a person at defined checkpoints, particularly for irreversible steps such as payments or external communications. Organisations adopting agents should start with low-risk, high-volume work, measure error rates honestly, and expand autonomy only as trust is earned.
Used this way, agents are not a replacement for people but a multiplier of them, taking on the repetitive middle of knowledge work so that humans can concentrate on judgement, relationships, and the problems that have no procedure. The organisations that learn to manage digital coworkers well, with the same care they bring to onboarding human ones, will define the next chapter of the AI economy.
Written by Prof. Dr. Prabal Datta Barua
Professor Dr. Prabal Datta Barua is an award-winning Australian Artificial Intelligence researcher, author, educator, entrepreneur, and highly successful businessman. He has been the CEO and Director of Cogninet Australia for more than a decade (since 2012), and has served as the Academic Dean of the Australian International Institute of Higher Education since 2022. He was awarded the prestigious UniSQ Alumni Award for Excellence in Research (2023) by the University of Southern Queensland, where he is a Professor and PhD supervisor in A.I. in Healthcare, and he has secured over AUD $3 million in government and industry research grants.
More about Prof. Prabal