Many AI initiatives have had massive ROI though. The implementation problems are similar to any pre-AI tech rollout and hugely expensive non-AI tech implementations fail all the time.

After spending nearly 5 years building software which uses AI agents on the back end I've come to the conclusion it's the PC revolution part 2.

Productivity gains won't show up on economic data and companies trying to automate everything will fail.

But the average office worker will end up with a much more pleasant job and will need to know how to use the models, just like who they needed to learn to use a PC.