The uncomfortable truth is many businesses do not know what they are doing with AI

AI is not plug-and-play technology and requires dedicated engineering and careful integration

AI tools are often bolted on to legacy IT systems with businesses aiming for small gains in individual productivity. Reuters

When Nvidia delivered yet another blockbuster earnings report this week – $46.73 billion in second-quarter revenue, a 56 per cent increase from a year earlier, slightly above analyst expectations – it looked like a victory lap for the poster child of the artificial intelligence boom.

With its shares up 35 per cent this year, making it the first company to breach a $4 trillion market capitalisation, the US-based chip maker has smashed earnings expectations repeatedly.

But markets, already nervous about the stretched valuations of companies linked to AI, were less convinced. Nvidia’s shares slipped about 2 per cent in early Thursday trading, as uncertainty over its sales in China – amid rising tensions between Washington and Beijing – clouded its growth outlook.

Soft skills to survival skills: How to prepare for the 'job apocalypse' due to AI

Read More

But part of the market anxiety stems from a deeper concern: despite the hundreds of billions of dollars being poured into AI, there is a creeping suspicion it may not yet be delivering real business value.

That suspicion was given further weight by a Massachusetts Institute of Technology report that claimed last week that 95 per cent of organisations have not yet seen any payoff from their investments in generative AI. No boost in productivity. No revenue gains. Not even a noticeable dent in the bottom line.

So where is the disconnect? Why are some companies pouring vast sums into AI infrastructure – Nvidia’s biggest customers, Microsoft, Google, Meta and Amazon are on track to spend $350 billion this year – while so few see meaningful gains?

Nvidia rally tests investors' nerve

01:44

The uncomfortable truth is that many businesses do not know what they are doing with AI. Too often, tools are used haphazardly – bolted on to legacy IT systems or confined to a handful of specialists – with efforts aimed at small gains in individual productivity, rather than improving how the whole organisation runs.

The result is predictable: scattered results, wasted investment and frustrated expectations.

In theory, generative AI should be transformative, driving productivity gains, cutting costs and unlocking new revenue. In practice, it rarely delivers. Yet real success is possible.

The companies seeing meaningful returns from AI today tend to share three key traits. Firstly, they democratise the technology. Not just training top executives or data scientists, but equipping rank-and-file employees with a working understanding of what AI can (and cannot) do. That means treating AI not as a mysterious force to be outsourced to technology vendors, but as a capability to be embedded across an organisation.

Secondly, successful companies define specific, high-value use cases. Not vague aspirations like “improve customer service” or “optimise the supply chain”, but tangible outcomes with dollar signs attached.

One Asian bank I advise, operating in a tightly regulated market, moved third-party models on to its own IT servers so the AI was able to work with sensitive data that could not be shared externally. It then built custom tools around these models – starting with simple applications for general staff, and later expanding to more advanced tools for bankers – while closely tracking the results at each stage.

Thirdly, the firms that thrive give AI the infrastructure it needs. This is not plug-and-play technology. It requires dedicated engineering, careful integration into existing workflows and well-defined governance frameworks to ensure it delivers real value.

You cannot just throw a large language model at a problem and hope it sticks. Companies that do so often end up with demo-friendly prototypes and no commercial payoff.

Contrast this with a pharmaceutical retailer that approached AI with surgical precision. It categorised its efforts into three clear categories: making existing processes more efficient; optimising operations using historical data; and monetising new products and services. That clarity created the conditions for real impact – and financial returns.

Sam Altman's warning of an AI bubble may be true. But the bigger threat is an AI winter

All of this stands in sharp contrast to the surge that has carried US equities to record highs in recent months, powered by investor bets on generative AI. The S&P 500 is up 9 per cent this year, while tech stocks rebounded strongly from their April lows.

But they lost ground last week, as concerns mounted over the stretched valuations of companies linked to AI. The unease deepened after Sam Altman, chief executive of OpenAI, warned that an AI bubble might be forming.

Nvidia, software group Palantir and chip maker Arm were among those caught in the sell-off, which was reinforced by the publication of the MIT report questioning business returns from generative AI. This was a sharp reminder that hype cannot outrun fundamentals forever.

Mr Altman's warning of an AI bubble may be true. But the bigger threat is an AI winter. If executives continue to overpromise and underdeliver, the shift in investor sentiment we are already seeing will only accelerate – not in a sudden crash, but in a long, quiet chill.

What companies should be doing now is recalibrating, moving past the PowerPoint phase and starting to ask tough questions. Does this AI tool fit with our systems? Can our people actually use it? And does it measurably improve something we care about?

For these companies, Nvidia’s earnings are only a weak signal of what is happening on the ground, even if the world’s most valuable public company is treated as a bellwether for the AI boom.

Nvidia's Jensen Huang promises a record year - even without China

01:37

There is a long lag between the chips Nvidia sells and the business results its customers deliver. Its processors power the training of models such as OpenAI’s ChatGPT and Google’s Gemini – tools that then take months to be built into products and workflows.

If Nvidia ever stumbles, it will say less about what is happening in offices today and more about what could unfold six to 12 months down the line.

Which brings us back to the headlines. Are 95 per cent of companies failing with AI? Perhaps. But the failure has less to do with the technology than with how executives choose to use it.

The companies already seeing results are the ones that spread understanding of the tools across their workforce, tie projects to clear business goals – and put the right infrastructure in place to make them work.

There is no shortage of potential. What is missing is strategy. It is not about the chips companies buy from Nvidia, but what they do with them.

Amit Joshi is professor of AI, analytics and marketing strategy at IMD