There are a few reasons why A.I. projects commonly fail:

① Insufficient or poor quality training data (e.g. better an old algorithm on good data than a cutting edge algorithm on poor data)

② No calibration. Calibration seems to be the best kept secret in ML (people know need good training data but they are lazy or not brave enough to insist on it... I've made that mistake, but https://scikit-learn.org/stable/modules/calibration.html seems to be truly obscure) In the case of a trading strategy or other commercial action you would be calibrating on expected return or possibly something that balances risk and return like Sharpe ratio.

③ Non-stationarity. Distributions are changing all the time on their own, but in markets they get changed by your own actions and those of people following the same strategy of you.

This book has a great study of a hedge fund strategy that burned out the way many algorithmic strategies do

https://www.amazon.com/Hedge-Funds-Perspective-Financial-Eng...

I think it's key that the burnout happens because of greed. A given strategy can absorb a certain amount of capital. When more capital gets attracted to the strategy the returns inevitably go down, market participants can try to make up for this by increasing the leverage but you can see how that goes... So I would still blame too much greed and too much capital because it inevitable that markets fluctuate.

This is what a quant fund bubble looks like.

Stock market bubbles are dramatic because people drive visible stock prices high.

When a quant stategy is overbought the returns on that strategy tend to converge on the T-bill rate at best. This might be hidden for a while because people turn up the leverage to make up for diminishing returns.

In a case like this, for instance, people investing in volatility hedges can drive down volatility.

See

https://www.amazon.com/Hedge-Funds-Perspective-Financial-Eng...