PwC asked 4,454 CEOs across 95 countries how AI was performing for them. 12% were getting both revenue growth and cost savings. 56% were getting neither.
The 56% is the number worth sitting with. Not because it’s surprising, but because of what usually gets blamed for it.
When AI doesn’t pay off, the assumption is that the technology was the issue. Wrong model, wrong vendor, wrong implementation partner. So firms procure more carefully next time, hire better consultants, run a tighter pilot. And the numbers stay roughly where they are.
It’s rarely the tool
I’ve spent enough time inside organisations running AI initiatives to have a fairly clear picture of what goes wrong. The technology is seldom it.
What I find is some mix of the same few things. Nobody defined which decisions the AI was meant to improve, or how you’d know if it had. The data feeding the model was messier than anyone admitted at the start, because the teams who owned it never had a reason to clean it up. Governance was either missing or theoretical. Someone wrote a policy, nobody reads it, and there’s no clear answer to who’s responsible when the model is wrong. And leadership announced the AI push while doing nothing about the conditions that would decide whether it worked.
None of that shows up on a technology roadmap. None of it comes up in a vendor demo. It only becomes visible when you look at the organisation instead of the tool.
Failure is hard to see in real time
Part of why the number is so high is that AI initiatives rarely die loudly. They produce results nobody can attribute to anything. They get quietly narrowed in scope. They get called a success because a pilot went well, with no follow-up on whether it scaled or changed anything at all.
If you didn’t define success before you started, you can’t tell whether you reached it. That sounds obvious. It happens constantly.
The firms that do get something real from AI tend to have done three unglamorous things first. They agreed what they were trying to change and how they’d know it had changed. They looked at their data honestly, including the embarrassing parts. And they were clear about who owned the output when it was wrong, in a practical sense, not a legal one.
What the number measures
The 56% isn’t measuring whether AI is useful. It’s measuring whether firms are ready to use it. Those are different questions, and most of the market treats them as one.
BCG’s research across 1,000 executives puts around 70% of the AI challenge on people and process, and 10% on the algorithms. That ratio matches what I see on the ground. The tool is the small part. The organisation it lands in is the rest.