In case you missed it, MIT’s latest report ‘The GenAI Divide: State of AI in Business 2025’ has revealed that 95% of AI pilots fail to deliver measurable business value.
Markets panicked, headlines screamed “AI bubble”, and stock prices in AI-linked tech companies heavily slid. However, for anyone who has been listening closely, this isn’t news. It’s exactly in line with what Jack Jorgensen, General Manager of Data, AI & Innovation here at Avec, has been saying for years.
The bubble was always going to pop
Off the back of the findings of our latest AI survey, Jack has often reminded us that decision-makers should focus on an outcomes based approach, and avoid decisions made on rhetoric such as: “It did really well in the 8 out of 10 examples I ran.”
While this might sound promising, when you scale this experiment to an industrialised process of, say, 100,000 events, suddenly you’ve got 20,000 mistakes on your hands. At that point? It isn’t innovation, but chaos.
Why failure is a good thing
Our survey revealed that 47.6% of organisations are still at the ‘experimental/pilot’ stage of AI adoption, and that’s where they should be. Pilots are meant to fail and filter.
As Jack puts it:
“95% of AI processes failing to launch is a good thing. The damage if these were to hit production would cause havoc across organisations, not to mention the reputational damage. It means their quality checks are working.”
Think about it: AI speeds things up, whether they’re good or bad processes. So, if your process has faults, AI increases the amount of issues. Fixing AI-driven mistakes often takes longer than if a human had done it from the start.
Innovation doesn’t always look like Steve Jobs
One reason so many pilots flop? Leaders chasing the thrill of innovation theatre. Everyone wants their Steve Jobs moment; the flashy product demo that changes the world.
But as Jack points out:
“The innovation which actually adds value to organisations is far less grandiose.”
The MIT report echoes this: organisations focusing AI on back-end processes and cost efficiency saw far more ROI than those throwing it straight at sales and marketing. Quiet, unglamorous process is where AI pays off.
The real takeaway
Don’t read “95% of pilots fail” as proof that AI doesn’t work. Read it as confirmation of what the experts have been saying all along: the problem isn’t the tech, it’s how organisations try to use it.
The real risk of AI pilot projects failing is when they don’t fail soon enough. Smart companies are the ones testing, breaking, and learning now, so they don’t unleash problems at scale later.
So, what does make an AI project successful?
It’s tempting to want a checklist or a fool-proof formula for AI success, but as Jack points out:
“There isn’t really a one size fits all answer as it depends on the process that is being actioned. I would suggest a process which is robust, reliable and can ultimately fail-over to a human if the system gets something wrong, breaks, or reaches a limit.”
That’s the crux. Success isn’t just about choosing the right model or vendor, it’s about choosing the right process for AI to augment and knowing what could happen when things go wrong.
Key considerations include:
- Failure impact – What happens if the process fails? Is it an inconvenience or a mission-critical risk?
- Capacity check – Can your organisation handle the increased speed AI introduces or will faster output create bottlenecks elsewhere?
- Process suitability – Some workflows are ripe for augmentation and others should stay human. Knowing the difference is half the battle.
As Jack says, “The list of what’s right changes dramatically based on what the process is, but these are the basics.” Successful AI adoption happens when organisations combine sensible process design with partners who bring the expertise, experience, and insights to help them make the right calls.
Need help moving beyond the experimental phase and build AI process that actually deliver? Reach out to our team of experts.
