The productivity boost from artificial intelligence could be transformational or marginal, and investors should pick stocks accordingly, Bank of America said.

If it's transformational, BofA said to buy the providers of the technology. But since it's hard to pick the eventual winners, strategists pointed to AI-themed ETFs. 

The positive impact of AI on productivity could also be limited to marginal gains. That was the case for previously-hyped technologies, like 3D printers, LED lights, and VR headsets, the bank said.

If that's the case, investors should pick stocks that are incorporating AI in their business models, such as firms like Deere, Interactive Brokers, and Humana. That's largely due to four concerns that make more "obvious" AI trades like hardware and services less appealing:

Investors are no longer operating in an era of easy money, where inflation stays close to the Fed's 2% price target and near-zero interest rates facilitate cheap borrowing.

"AI stocks are priced today for a rapidly fading world of 2% bond yields, low costs of capital, easy globalization, and friendly inflation," strategists said. "We expect a structural shift to higher inflation and interest rates over time, from the '2%' world of the past two decades to a '5% world' typical of the last hundred years of market history."

As much as a third of AI firms on the market are unprofitable, Bank of America estimated. That's about double the proportion of unprofitable companies in the Nasdaq 100, wherein just 15% of firms have been unable to pull a net return.

And even successful companies in the AI space may have limited upside. Nvidia stock, for instance, already trades around 40 times its revenue, while C3.ai trades around 16 times its revenue.

The average AI fund is also valued 31 times its forward earnings estimates, the bank said, which is around 1.7 times the ratio seen in the S&P 500 and 1.2 times the ratio seen in the Nasdaq Composite.

Investors have continued to pour their cash into popular AI contenders. That could signal that the sector is overvalued.

Artificial intelligence ETFs saw $3.5 billion of inflows over the past three years, with just $1.1 billion bought below "current breakeven levels," strategists said. Meanwhile, 55% of fund managers saw large-cap tech stocks as the most overcrowded trade, per Bank of America's Fund Manager Survey.

There are obstacles that stand in the way for the improvement of AI technology. Large language models like ChatGPT, for instance, are using more energy and data to train. That could eventually lead to lower-quality data being fed to AI models, the bank said, particularly if firms are stingy with the information they provide to protect company secrets, or governments impose tighter regulation on AI bots. 

That could be worsened by the US's new export curb to China for computer chips and chip manufacturing equipment, the bank added.