When 6GWorld spoke to Chetan Dube, the Chairman and CEO of Quant, about agentic AI we were not expecting to hear the phrase “Universal Basic Income” or squabble about artificial intelligences competing for jobs. Nevertheless, the possibilities of the technology have some quite far-reaching implications.

But first, some fundamentals…

What is Agentic AI?

For those of us struggling to keep up with the development of Machine Learning, AI, GPT, LLMs and so forth, Agentic AI is a new variation with added capabilities.

“GPT was knowledge retrieval; agentic is a way for you to take it from pure knowledge to action,” Dube summarised.

“If I wanted to redesign the 6GWorld website, a GPT would give me the best available information on how to do that; whereas the agentic would just build me the website.”

How does Agentic AI Function?

Dube is deeply knowledgeable about the technology as he’s been involved in developing it for the past forty years; when asked how it works in detail and whether it’s an evolution of GPT capabilities or something different in kind, he broke it down for us in a straightforward example from the world of CRM where Quant operates.

“One of the largest utilities is right now going towards being serviced through their entire contact centre by agentic agents. One of the simplest calls coming in could be “Why is my bill for September more than my bill for August?”. Anyone who’s used previous chatbots knows what an absolute aggravation that is because it’s going to try and pigeonhole you into one of the seven things it knows inside cookie cutter answers; and the cookie cutter answers only account for 14% of the volume.”

This is quite different from a typical human brain, which will deconstruct the individual question and work out a process to solve it – in this case, opening up the bill for August and the bill for September, comparing them, and trying not just to identify where there are differences but then what that might mean for answering the “why” question.

In other words, for a human to answer that question requires taking specific actions as well as having a certain type of knowledge at your fingertips.

“Agentics are built by having a set of basic atomic “lambda blocks” as actions – in this case let’s say for example get a bill, pay a bill, start service, stop service,” Dube explained.

“You use LLMs and GPT capabilities to be able to select the right set of atomic options that can actually drive this conversation forward towards the goal state…  the agentic framework is about a bunch of actions – not a function or anything, but the atomic actions that can be built up into a function. The GPT is allowing you to select the right atomic models to be able to push the conversation towards the goal.”

This is as close as possible to how humans behave, which means that – when properly trained – they can respond more appropriately to people. In the case of the call centre above there’s a good reason why they’re moving to agentics:

“We are able to get 87% call deflection [that is, answering the problems without being on the phone to a human] when typical chatbots from two years ago struggled to get 23%.”

Risky Business?

It’s easy to imagine things going horribly wrong with this set-up. GPTs and AIs have been recorded making things up, behaving the wrong way or even having what we might think of as psychotic breaks. While this is limited to text or chat interactions the impact can be mitigated, but what happens if it’s taking real-world actions?

Dube has heard this before, that’s for sure – likely from nervous IT and business leaders. He can’t argue that GPT can be a black box coming up with odd or unpredictable outcomes, such as the chatbot that invented a new tariff for a customer.

“When you go LLM-only then you are subject to convolutional neural networks which are inherently probabilistic in nature… And you really don’t have any control over the path it’s going to take because it’s working on 1.1 trillion parameters. In the case of agentics, the good part is that you are able to layer in determinism, because you are only allowing the LLMs to be able to call these specific atomic functions.

You are layering determinism on top of a probabilistic LLM framework to be able to provide banks, insurance companies and so on the guardrails and predictability they need.”

Perhaps the bigger risk, then, is societal.

Regulate This

If one AI can replace a call centre of 6,000 – which is in process at the moment – and operate 24/7, then there are some very rapid shifts taking place. In countries such as the USA and UK the last few decades have seen economic growth accruing to relatively few in society; middle classes have weakened. Agentic AI moves the threat of AI on employment into new areas.

“Productivity will go up by about 38 to 41% because these machines will work nonstop, so you can imagine how the GDP will have a big surge,” Dube commented.

Unlike some of the technology CEOs, the fallout of what he’s creating is something Dube appears to be taking very seriously.

“How do we redistribute that growth, not just have it concentrated in the hands of a few? How do we make sure that that wealth that is being generated by these entities is able to provide for everyone on this planet?” He asked.

Dube also believes he sees the solution.

“I’m a big believer in the minimal viable income, universal basic income that can allow these machines to be able to provide for every person of the human race,” he explained. This would enable people to worry less about subsistence and look both more towards long-term aims and do things that realise their aspirations.

“I have yet to meet anyone who tells me their dream is to answer the telephone to angry people every day,” he commented wryly.

Even in Dube’s own case, he pointed to a love of painting in his youth, but a perceived difficulty in making ends meet pushed him into maths and then AI. Trials of universal basic income seem to have been beneficial in enabling people to pursue education rather than forcing them into work, and to have boosted the prospects of the people involved.

So it is an exciting time, and a pivotal one right now given the pace of adoption of AI systems. Previous labour-saving devices such as the washing machine, kitchen tools, the tractor and so on have not given people more down-time, which may be due to the incentives baked into society. Arguably we need to be looking again at how we manage the systems and incentives to the benefit of the people in those societies, especially when a potent disruption is on the cards.

“The perceived hyperbole should not disguise the fact that the future of the world hangs in the balance here, because when you start to have a change that is so rapid and so can be so far-reaching we need to be able to make sure that we start to do the right things at the onset,” Dube noted.

“I’m a subscriber to that utopian view but the future does hang in balance. People need to know this is coming.”

Image by Pixlr AI: “Call centre robots with flags marching happily to a bright future in front of a rising sun”