The problem with agentic AI: no one is actually using it

How do you illustrate an article about AI? It’s a quandary picture desks and art editors have been wrangling with a lot over the last few years, because there are only so many spinning orbs and moodily lit microchips out there to choose from. Far better to use Metal Mickey (below), the iconic TV robot who, during the 1980s, epitomised technology for a generation.

But there are bigger problems for those writing about agentic AI than which robotic cliche to reach for in the image library. Because nobody, not even the people developing it, is quite sure what it actually does.

Generative AI, essentially a chatbot fed on a library’s worth of data, was relatively comfortable as a notion because we were all using a variant of it by the time it became truly mainstream. Agentic, best defined as the ability of AI systems to communicate with each other and carry out end-to-end transactions, feels pretty alien.

I’ve been writing a lot about agentic recently because its potential underpins a lot of the economic shifts we’re seeing in industries from marketing to tech, as well as predictions about the future labour market. But if you look for live examples of agentic, they’re mostly theoretical: AI ‘personal shoppers’ that know what we’re looking for and buy it on our behalf, or an AI ‘digital twin’ that goes to meetings with us, takes notes and assigns actions.

That sounds plausible, but it’s not happening right now. As the Financial Times found when it compared the performance of mainstream AI agents on a variety of office tasks, they offer a useful facsimile of a human worker but there are too many errors and inefficiencies to be useful or reliable.

I spoke to Rob McCargow, PwC’s HR technology guru, as part of a recent Raconteur article on agentic. “When we’re doing big surveys, on the face of it some of the levels of adoption of AI look quite high until you tease out the detail of who’s actually moved beyond the basics into true AI,” he told me. “Agentic is another level deeper from that. Every conference you go to talks about it, but on the ground it’s still very early days.”

Agentic does offer some tantalising possibilities. You can imagine, for example, an AI agent issuing an invoice to another business’s agent, then reconciling it automatically. You could also feasibly tell an agent where you wanted to go on holiday and let it find a suitable flight and hotel room, cutting online booking sites out of the loop. But in neither scenario would it be an end-to-end automated process: you’d need a human check before money left your business’s account, just like you’d want to see the itinerary and the costs of your trip in detail. Talk of leaving everything to the agents is a flight of fancy.

It isn’t fantastical, however, to suggest that one day, as Accenture claims, more AI agents than humans will be using ERP software. Or that agents will account for the majority of traffic to e-commerce sites. But even Neil Patel, a notable digital marketing optimist, told me in a Campaign Asia interview that he doesn’t see that happening in the next five years. There’s a lot of words to be written, and doubts to be aired, about agentic before we reach that point.