Why do 85% of AI projects fail?
The famous 85% figure comes from a 2018 Gartner prediction, and similar failure rates keep being reported since. The causes are consistent: solving problems nobody has, bad or insufficient data, no clear success metric, underestimating the human process change, and treating AI as magic rather than software. Small businesses avoid most of this by starting with one boring, measurable problem.
First, the number itself: the 85% figure quoted across the industry traces to a 2018 Gartner prediction that through 2022, 85% of AI projects would deliver erroneous outcomes due to bias in data, algorithms, or the teams managing them. Related figures circulate: VentureBeat reported in 2019 that 87% of data science projects never make it into production, and studies since the generative AI boom (including MIT-linked research in 2025 reporting ~95% of enterprise GenAI pilots showing no measurable P&L impact) keep landing in the same territory. Treat the precise percentage as folklore; the pattern it describes is real.
The failure causes are consistent across these studies. Projects start from the technology rather than a problem: 'we should be using AI', so there's no definition of success to hit. The data needed doesn't exist, is scattered across systems, or is too messy to use, which surfaces only after months of work. The process never changes: a tool gets bought, the team keeps working the old way, and the subscription lapses without anyone noticing.
There's also a measurement failure: many 'failed' projects were never instrumented, so the team can't say whether it worked. There's an expectations failure too: pilots are judged against the magic in the sales deck rather than a realistic baseline of the current process.
For a small business, the lessons compress into a short list. Start with one specific, boring, recurring problem: quote follow-ups that don't happen, enquiries that take days to answer, reports that eat a Sunday. Pick something you can measure before and after. Use the simplest tool that solves it, prove it works, then expand. The businesses getting real value from AI are running dull, measurable automations rather than chasing moonshots.
The same discipline applies to AI in marketing: 'use AI to grow the business' fails; 'answer every enquiry within five minutes' or 'a proper page for every area we serve' succeeds, because success is checkable. If a supplier can't tell you what number their AI project will move, that's the 85% calling.
Useful next steps
Want answers specific to your business?
Book a free call. No jargon, no obligation, real numbers.