Why consumer AI agents could be the next big investment shift

THREE years ago, ChatGPT brought artificial intelligence into everyday life. For the first time, people could simply talk to AI. You asked a question, and it gave you an answer.

That was the ChatGPT moment. The next shift could be even more consequential.

AI may soon move from something we talk to, to something we delegate to. That is the promise of the consumer AI agent.

From answers to actions

The distinction matters. Today, you may ask an AI: “What are the best hotels in Tokyo?” and receive a list of recommendations.

An agent could go several steps further. It could know your preferences, check your calendar, search flights, compare hotels, make the booking, reserve a restaurant and eventually complete the payment.

The chatbot gives you information. The agent completes the transaction. That seemingly small difference could have very large economic consequences.

Early consumer interest is already worth watching. Meta’s Muse reportedly reached around 1.8 million downloads in its first 12 days, compared with roughly 1.3 million for ChatGPT in the same period.

It is still very early, and download numbers alone do not tell us which platform will ultimately dominate. But the signal is important.

Once consumers become comfortable saying: “Handle this for me”, the potential market around AI expands dramatically.

From enterprise AI to consumer commerce

Much of AI monetisation today sits within the enterprise and IT spending pool, estimated at roughly US$6.2 trillion.

Consumer agents could increasingly extend AI into much larger pools of economic activity, including around US$6.9 trillion of e-commerce and US$1.3 trillion of advertising.

Some estimates suggest the agentic AI market itself could grow roughly 13 times by 2030, while AI agents could eventually influence or intermediate more than one third of online commerce.

That would amount to a seismic shift. AI would no longer simply be a tool sold to businesses. It could increasingly sit directly in the flow of consumer spending.

For investors, however, this does not necessarily mean trying to identify the dominant consumer agent today.

Competition will be intense. The technology is evolving quickly. The eventual winners may not even be obvious yet. A more useful approach may be to focus on what almost every successful agent will require.

Compute moves beyond GPUs

The first is compute.

AI agents do considerably more than generate an answer. They may run software, open browsers, call application programming interfaces, query databases, use memory and execute multiple tasks in the background.

That changes the nature of compute demand.

The first phase of the generative AI boom was overwhelmingly graphics processing unit (GPU) intensive. Training large models required huge amounts of accelerated computing.

Agentic AI could broaden that demand. A traditional AI server might have roughly one central processing unit (CPU) for every eight GPUs. In more agent-intensive workloads, that ratio could move much closer to one CPU for every two GPUs.

The reason is straightforward. The GPU performs the heavy AI computation. The CPU coordinates everything around it.

As agents move from answering questions to executing millions, and eventually billions, of real world tasks, the infrastructure bottleneck may broaden from GPUs alone to CPUs, foundries, advanced packaging, substrates, memory, networking and other parts of the semiconductor supply chain.

The AI CPU market could grow at around 40 per cent annually in the next five years.

The key investment point is that the AI infrastructure opportunity could become much wider than the first phase of the cycle.

A payment rail for every transaction

The second area to watch is payments. If AI agents increasingly shop, book and transact on our behalf, successful tasks eventually have to move money.

Someone still has to authenticate the customer, process the payment, manage fraud and settle the transaction.

That creates an interesting investment asymmetry. We may not yet know which consumer AI agent becomes dominant. But whichever platforms succeed, the financial infrastructure beneath those transactions could still benefit.

The agent may change. The need to move money does not.

When the middleman meets its middleman

The third implication may be less comfortable for some existing business models.

Travel websites, comparison platforms and other digital intermediaries make money because they own discovery, traffic and customer intent. Consumer agents could challenge that position.

Imagine telling your agent: “Book me the best value hotel in Tokyo.”

The agent may search multiple platforms. It may compare prices automatically. Over time, it may increasingly transact directly with the hotel or service provider.

If that happens, the agent itself becomes the new aggregator. It is when the middleman meets its middleman.

That could shift bargaining power away from some of today’s digital middlemen, particularly businesses whose economics depend heavily on owning the interface through which consumers search and compare.

Not every aggregator will be disrupted. Some may evolve into infrastructure providers themselves.

But the strategic question is becoming harder to ignore: If the consumer no longer starts with a website or app, who owns the customer relationship?

Trust before take-off

There is, however, a considerable distance between asking AI for advice and allowing it to spend your money.

A poor recommendation is annoying. An unauthorised, non-refundable booking is expensive. Consumers will need spending limits, clear approvals, privacy protections and a practical way to undo mistakes.

Trust is therefore not a footnote to the investment thesis. It is a condition for adoption. So are economics: An agent that costs more to operate than the convenience it delivers is not a sustainable business.

This argues for selectivity, rather than indiscriminate enthusiasm.

When AI stops answering and starts acting

The first phase of the AI investment cycle was largely about building the models and the infrastructure required to train them. The next phase may be about what happens, when those models begin acting on our behalf.

That could expand the opportunity from enterprise software into commerce, travel, advertising and payments, while broadening infrastructure demand in the semiconductor stack.

It is still too early to identify the dominant consumer agent. But investors may not need to.

The more durable question may be simpler: What does every successful agent need, and which existing business models become less valuable when the agent sits between the consumer and the transaction?

ChatGPT made AI something we talk to. Consumer agents could make AI something we delegate to.

And if they eventually intermediate trillions of dollars of consumer spending while generating a new wave of compute demand, that could mark the next major chapter of the AI investment cycle.

The writer is regional CIO Southern Asia-Pacific, UBS Global Wealth Management