The Economist: Why your company is struggling to scale up generative AI

For investors concerned that America’s tech giants are making recklessly large bets on generative artificial intelligence, big tech’s latest quarterly results have offered some reassurance.

The growth in demand from companies for the cloud services of Amazon, Microsoft and Google was red hot.

Andy Jassy, boss of Amazon, put it most strikingly.

He said that AI revenue for Amazon Web Services was growing at triple-digit rates — three times faster than AWS itself grew in the early years after it pioneered cloud computing in 2006.

Dig deeper, though, and the situation is more nuanced.

Generative AI appears to be one of those innovations, such as email or smartphones, whose most eager early adopters are individuals.

Companies are being far more tentative.

In the two years since OpenAI launched ChatGPT, generative AI has had a faster rate of adoption than personal computers or the internet.

Fully 39 per cent of Americans now say they use it, according to a study by Alexander Bick of the Federal Reserve Bank of St Louis and co-authors; 28 per cent say they use it for work and 11 per cent that they do so every day.

Many employees seem to be secret cyborgs, using generative AI in their work even as their employers go slow.

Only 5 per cent of American businesses say they are using the technology to produce goods or services, according to a survey by the US Census Bureau.

Many companies seem to be suffering from an acute case of pilotitis, toying with pilot projects rather than implementing the technology on a large scale.

In a recent survey conducted across 14 countries by Deloitte, a professional-services firm, only 8 per cent of firms said they had deployed more than half of their generative-AI experiments.

As a consequence, revenue from AI services remains limited.

Although Mr Jassy said AWS now generates “multi-billion” dollars of revenue from AI, that is a smidgen of the $110 billion of annual revenue for its cloud business as a whole.

Accenture, a consulting giant that recently announced it would train 30,000 staff to better help companies adopt generative AI, said in September that it had booked $3b worth of work related to the technology over the past 12 months, a ten-fold increase year on year.

But compared with the company’s total sales of more than $81b, that too is small beer.

Why are many bosses hesitating to adopt generative AI?

One reason appears to be worries about the risks posed by the technology.

Listen to the tech giants and they will tell you — as Sundar Pichai, the boss of Alphabet, said in July — that “the risk of under-investing is dramatically greater than the risk of over-investing.”

Alphabet, Amazon, Microsoft and Meta are expected to pour at least $200b into AI-related capital expenditures this year.

Bosses in other industries are more circumspect.

At a recent closed-door discussion, the head of a big American business group spoke of two types of fears chief executives have about generative AI.

One was being left behind if they adopted it too slowly.

Yet they also worried about being embarrassed if they moved too quickly and damaged the reputation of their firms.

Legal and regulatory risks loom large.

Lawsuits related to privacy, bias and copyright violations are making their way through the courts.

In August the European Union’s AI Act came into law.

AI bills have been introduced in at least 40 American states this year.

Bosses in heavily regulated industries, such as health care and financial services, are especially wary.

Although they recognise the potential of generative AI to transform their businesses, say by speeding up drug discovery or fraud detection, they are also keenly aware of the threats to privacy and security if their customers’ medical or financial data are compromised.

Another problem is that the benefits of adopting generative AI can be uncertain.

Using large language models is expensive, whether via a company’s own servers, which may be safer, or via cloud-service providers, which may be simpler.

Full-scale implementation of generative AI may increase revenues and reduce costs, but the payoff is not immediate, raising concerns about returns on investment.

In its most recent survey Deloitte found that the share of senior executives with a “high” or “very high” level of interest in generative AI had fallen to 63 per cent, down from 74 per cent in the first quarter of the year, suggesting that the “new-technology shine” may be wearing off.

A business leader sums up the scepticism by recounting the story of a chief information officer whose boss told him to stop promising 20 per cent productivity improvements unless he was first prepared to cut his own department’s headcount by a fifth.

Even when companies are eager to scale up their use of generative AI, though, they can quickly find it tricky in practice.

To reap the full rewards of the technology, companies have to first get their data, systems and workforce into shape, says Lan Guan, Accenture’s head of AI.

She says companies’ readiness for generative AI is many times less than it was for previous technology waves such as the internet or cloud computing.

One problem is messy data, scattered in different formats across various departments and handled by a variety of software companies, such as Salesforce for customer information and SAP for internal processes.

Ms Guan gives the example of a telecoms firm that wanted to train a call-centre AI assistant by feeding it PDFs, manuals, call logs and more.

It found that instead of one standard operating procedure — what she calls “a single source of truth” — the company had 37, accumulated over decades.

A failure to organise data before using it to train a bot increases the risk of hallucinations and mistakes, she says.

Another problem is legacy IT systems stitched together in many cases over decades.

Creaky old software is often full of bugs, or “technical debt”, that can make it tricky to plug in LLMs without causing further problems.

What is more, integrating semi-autonomous AI agents into IT systems built for human interactions could create security vulnerabilities.

Above all, there is the problem of skills.

Many companies are still struggling to get their hands on enough AI specialists.

According to Lightcast, a firm of labour-market analysts, AI-related job postings in America have surged by 122 per cent so far this year, compared with an 18 per cent rise in 2023.

Elizabeth Crofoot, an economist at Lightcast, says that this increase is mostly explained by generative AI, with job descriptions mentioning ChatGPT, prompt engineering and large language modelling on the rise.

Some companies are also on the hunt for workers in other roles who know how to apply generative AI to their jobs, and are willing to pay a premium for them.

A sales rep with AI skills can earn $45,000 a year more than one who lacks them, says Ms Crofoot.

No wonder, then, that even as some bosses prevaricate about scaling up generative AI, their employees are all in.

If there is one way to keep their jobs safe, it is to stay ahead of the curve.