An expert who has been working with artificial intelligence for four decades warned a re/insurance industry audience to be cautious about assuming data-centre risk and to build strong guardrails around the use of AI agents.

Alok Aggarwal, the chief executive of Scry, a company that provides AI-related products and services for industries including insurance, said that AI technology had some way to go before it could deliver the anticipated productivity gains.

In his book, The Fourth Industrial Revolution and 100 Years of AI (1950-2050), Dr Aggarwal estimates that around 400 million jobs will be replaced by AI technology by 2050.

However, in a keynote speech at the PwC Insurance Summit this week, he highlighted the challenges of the AI industry in making its economics work.

“We are all using AI, but how are we using it?” Dr Aggarwal said. “And is it productive? That will be a very important aspect.”

He cited findings of research by Apollo Investments that showed that 493 of the 500 companies in the S&P 500 had produced zero productivity gains over the past two years, a time in which AI tools had come to the fore.

At the same time, there was little evidence of the massive job losses that had been widely predicted.

“Why is this?” Dr Aggarwal said. “Because the cost of these tokens of AI is extremely high. The hardware is very expensive. At this cost, you cannot show return on investment.

“For AI tools to replace jobs, such as at outsourcing companies, they have to be substantially cheaper than Indians and Filipinos and others in lower-wage countries. We are very far from that.”

He referred to comments on an earlier panel at the conference, which had highlighted data centre risk as an historic opportunity for the re/insurance industry to grasp. And he contrasted the AI infrastructure buildout with a 19th-century technological change.

“For a change, I’m actually glad that insurance companies move at a snail's place,” Dr Aggarwal said. “And the reason is, if you talked about data centres as a business, you would say they could go bankrupt.

“Why do I say that? They’re very expensive and not like the railroads — 55 to 60 per cent of the railroads made in the 1860s are still operational today in the UK and US. In data centres, the GPUs and CPUs [specialised microchips] die out in three to five years.

“There are a lot of issues we have to figure out about data centres, cost being one of them when their depreciation is so high.”

Six of the world's largest hyperscalers – huge technology companies ploughing billions into data centres – are projected to grow their combined capital expenditures on AI infrastructure beyond $1.3 trillion by 2027.

Dr Aggarwal, whose experience includes 16 years working in IBM’s research centre between 1984 and 2000, said adequate governance of AI was hugely important.

“We are using a tool without realising what harm the tool we can do,” Dr Aggarwal said.

Strong guardrails were necessary for anyone deploying AI agents, he emphasised, adding that AI had “no values system”. The highly publicised case of Anthropic AI agents going rogue last July, and hacking into the systems of machine-learning firm Hugging Face, was the type of scenario that would occur again without adequate guardrails, he said.