The perils of performative AI

By Peter Springett · 3 March 2026

In the past six months or so, we've all seen research papers revealing that anything between just 5% and 15% of AI projects deliver ROI.

That many? Quite honestly I'm surprised that we're hitting double figures given the scatter gun approach taken by so many businesses:

- Giving Claude to every team and let them vibe code apps without IT oversight or strategic coordination.

- Installing CoPilot on every laptop and giving employees half a day's training.

- Attempting to implement sophisticated LLMs while the company’s actual data remains trapped in siloed, unorganized spreadsheets or legacy systems that AI can't access.

All these, and many more, are examples of performative AI: The implementation of artificial intelligence "to maintain a progressive corporate image, rather than achieving meaningful operational efficiency or solving specific business problems".

I've already seen cases where management has uninstalled LLMs and vibe coding tools until a) there's a clear, measurable target or b) The business has consilidated fragmented pilots into meaningful strategy.

The result: fewer, better quality implementations. My guess is that business user numbers will flat line or even stall in the first half of this year while organizations mop up pilot slop, and put in place more solid foundations going forward.

(Thanks to Gemini and Nano Banana 2 for the copy and image for this post).