Unlock the Editor’s Digest for free

The writer is head of AI and Data Science at HSBC

Markets love drama. Few tropes have spread faster this year than the claim that artificial intelligence is an asset bubble in search of a pin. Coverage that explicitly mentions an “AI bubble” has proliferated.

The narrative we hear is that trillions are being committed to AI investment, but little is showing up in revenues or productivity gains. It’s a neat narrative but it’s also increasingly wrong.

A recently published academic study offers unusually clean evidence that generative AI can drive increased revenue, and thus improve productivity, inside a large retail operation. This is important as it shows that AI implementation can help to increase revenue in the real economy. In other words, it’s not just companies selling AI that can benefit from the AI theme.

The headline result of the study from researchers at Zhejiang and Columbia universities — Generative AI and Firm Productivity: Field Experiments in Online Retail — was simple.

Most of their GenAI experiments increased revenue, in some cases by a large margin. The largest gain was recorded when the platform added an AI assistant before the point of purchase — sales rose by 16.3 per cent and the conversion rate, the share of visitors who become buyers, increased by 21.7 per cent.

An even tougher comparison, which pitted a hybrid AI system that escalated complex issues to humans against a team of human agents, produced an 11.5 per cent sales increase.

This dovetails with what we’ve seen at HSBC in experiments in which we pit AI against humans in investment research. The results suggest AI is best used to augment human analysts, rather than trying to replace them. This is not something which is likely to change as AI models improve — high-stakes decisions, in particular, are likely to remain the purview of humans for the foreseeable future.

Some content could not load. Check your internet connection or browser settings.

But the distribution of benefits found by the Zhejiang and Columbia study is telling. Smaller and newer sellers, along with less experienced buyers, saw a disproportionate lift. If this pattern repeats itself across sectors, it hints at a narrowing of capability gaps. That would be awkward for the simple “winner takes all” story and would force investors to look beyond the usual list of obvious beneficiaries.

Not every GenAI experiment in the study succeeded. When AI was used to generate product titles for Google adverts, the result was slightly negative for sales, albeit not to a statistically significant degree. However, the number of times the adverts were seen or clicked on declined significantly. The mechanism for this failure is unclear. Perhaps the GenAI models simply needed to be fine-tuned on an advert-specific dataset so they could produce more compelling advertising copy.

Alternatively, perhaps the Google algorithms identified the adverts as being AI-generated, marked them as less relevant and thus showed them to fewer people.

There are two practical implications from this research. First, the value from GenAI is already showing up in ordinary places: customer service, query refinement, translation, and more efficient matching of customers to products. These sorts of GenAI use cases will be repeatable in many areas of the economy and thus small changes in best practice could have large effects.

Second, the winners of the AI theme will not be limited to those involved in the AI infrastructure build-out. Some of the strongest results may come from companies that implement AI into their operations to achieve lower costs, increased revenue, or both. Indeed, we have already seen that the share prices of US companies which have implemented AI into their operations have outperformed their peers that have not.

Recommended

Thus far, this has been the market pricing in expectations of continued successes from these AI projects. The likelihood of these AI-generated benefits has been called into question by the AI bubble narrative, but this recent evidence suggests these expectations are more likely to be fulfilled.

The AI bubble narrative will not disappear. It is too tidy and ties in too neatly with the memories from the dotcom bubble. However, there are important differences. Most of the spending on AI comes from already large incumbent companies which have vast resources to invest, rather than speculative start-ups. Those worried about an AI bubble can point to S&P 500 capital expenditure as a fraction of GDP being higher than the levels seen during the dotcom bubble.

Yet, the equivalent capex numbers are only about 40 per cent of operating cash flow — far lower than the over 70 per cent levels seen during the dotcom mania.

The returns from GenAI have been clearly visible in the earnings of the companies associated with the AI infrastructure build-out for some time now. What’s now becoming clear is that the benefits of AI are accruing to the rest of the economy too.