AI ‘agents’ aren’t matching up to the buzzwords

THE biggest issue with the term artificial intelligence (AI) “agent” is that everyone seems to have a different definition for what it means.

Most often, it is used to describe an AI system that can act autonomously and work with outside applications to complete increasingly complicated tasks. The buzziest example from Asia has been Manus, which went mega-viral earlier this year.

But it has also morphed into a marketing buzzword, slapped onto everything from products that surf the Web on their own to bots that will eventually achieve human decision-making skills – and could be coming for your job.

In China, where reports of a new agentic tool seem to emerge every week, some firms have been accused of labelling their products AI agents just to capitalise on the hype.

Against this confusing backdrop, SoftBank Group founder Masayoshi Son – known for adding zeros to his bold aspirations – says that he plans to deploy one billion AI agents within his company by the end of the year. Speaking at a business conference in Tokyo last week, Son waxed poetic about how these systems will be able to think for themselves, self-replicate and work 24 hours a day.

They will participate in meetings, make phone calls and send e-mails, he said, and will evolve on their own. He envisioned a boon in productivity. A comparison he used was to give workers the power of Senju Kannon, a Buddhist figure with a thousand arms and eyes.

It is part of the breathless praise, and stretched analogies, that have been heaped on AI agents from some of the tech sector’s most influential voices over the past year.

On Jul 18, after releasing new agentic features in ChatGPT, OpenAI chief executive officer Sam Altman said that watching the tool take on tasks “has been a real ‘feel the AGI (artificial general intelligence)’ moment for me.” (Others have been less than impressed.)

It is worth noting that AGI is yet another overused industry term, loosely defined as AI that is as smart or smarter than a human – but similarly spurs disagreements over what exactly that is.

More than halfway through what has been widely heralded as the year of the AI agent, then, it is worth unpacking how we got here. And perhaps we need to find a better term than agent. These things are not James Bond.

Behind the global push for agents is the hope that this next iteration of AI is what will finally fulfil the promise of making workers more productive (or redundant?), and bring about some returns on the sky-high investments.

A McKinsey report published last month said agents are the key to break out of what it calls the “generative AI (GenAI) paradox” – or data that indicates nearly eight in 10 companies are using GenAI, yet just as many “report no significant bottom-line impact”.

The report urged business leaders to reinvent their entire workflows to centre around agents, to boost operational agility and unlock new revenue opportunities. Still, it also warned that they “introduce a new class of systemic risks”, including the ominous “controlled autonomy”. Not only does launching a billion agents in less than six months seem technically unfeasible, but this also makes it sound perilous.

The reality is that the road to infusing human-like decision-making skills into AI programs is a long one. Cars have been around for well over a hundred years, but autonomous vehicles remain far from mainstream, despite loud predictions over the past decade that they are just around the corner. That does not mean it is all hype.

It is the same for AI. It makes sense that agents built off large language models are good at solving text-based problems, like coding or compiling research papers. They have also shown prowess at mundane and repetitive tasks, where the risk of giving too much agency is much more contained. And recent updates allow certain models to take on multi-step tasks, even if they still require some human oversight.

But the path to fully granting autonomy to agents in the workplace is at least half a decade away. It will require major tech infrastructure upgrades that give these systems greater access to the tools we all use to do our work, and new safeguards for how to govern that sensitive data. It will also demand the establishment of new standards for liability if these tools make a mistake.

Meantime, we should cool off comparing the technology to gods or even humans. It only compounds the hype – and the angst people feel about machines snatching their livelihoods. It would be more useful to embrace AI as normal technology, recognising it can have a major impact on society, but stripping away the idea that it has its own agency.

Instead of anthropomorphising machines, business leaders must focus on how to use AI tools to solve real problems that humans face in the workplace. Only then will it impact productivity and bottom lines. And without a clear definition for what an agent even is, one billion of them can mean everything – and nothing. BLOOMBERG