More companies are getting money back from AI, and almost as many are learning that the work around the AI is harder than the AI. 64% of business leaders now count their AI spending as positive, up from 54% a year earlier. But 21% have canceled a major AI project after it failed, and more than a third say their AI strategy came about by accident.
The computer systems are the half companies know how to fix
Many leaders point at their technology. Two thirds of CIOs think their foundations need significant work before AI can be used widely, and that is probably right. It is also the kind of problem a company knows how to handle: you buy, migrate, patch and schedule.
Ismail Amla, who leads consulting at Kyndryl, the IT services company behind the research, argues the harder problem is how work changes: whether the people who use AI agents are trained for it and rewarded for making it work. That part has no purchase order. It is a decision about how a team spends its day.
The electric factory had the same two halves
Before electricity, a factory was built around its steam engine. The engine turned one long shaft, belts carried the power to each machine, and so the machines were packed close to the shaft in tall, narrow buildings. Moving a machine meant moving a belt.
Electric power can be split into small pieces and sent anywhere a wire reaches, and that was the real change. But the first buyers used it as a cleaner steam engine. Only when each machine got its own motor could a factory be rebuilt around the order of the work. Economists who studied it found the delay was not about the motors. It was about redesigning the factory, changing the workflow, teaching workers new skills and changing how managers managed.
What it looks like in an accounts payable team
Take a finance team of four that processes 1,500 supplier invoices a month. They add an AI that reads each invoice and fills in the entry. A clerk still checks every one, a manager still signs off, and payment still goes out in one weekly run. If typing an invoice took two minutes, the team gets back about 50 hours a month, and the process is otherwise the same.
Now rebuild the work around the tool. The AI matches each invoice against the purchase order and the delivery record. Suppose about nine in ten match. Those go straight into the next payment run. People look only at the ones that do not match, and the manager's sign-off applies to those. If a check took three minutes, skipping it for about 1,350 invoices gives back another 70 hours a month, more than the AI saved on typing.
The software is the same in both cases. The difference is that someone had to agree that matched invoices need no human look, and that a clerk is now judged on how well exceptions are resolved, not on how many invoices crossed the desk. That is a change in incentives and training, which is exactly the part the leaders find hard.
Ask who owns the process
The person who can redesign the work is the owner of the process, not the server team. Before the next AI tool goes live, list the steps in the process and mark the ones that exist only because the work used to be slow or hard to trust. If you remove none of them, expect the typing to be saved and little else.