80% of AI Projects Crash and Burn, Billions Wasted Says Rand Report
A RAND report citing estimates that over 80% of AI and machine-learning projects fail has triggered debate over whether this represents extraordinary waste or a normal hit rate for cutting-edge R&D. Commenters highlight recurring causes: executives chasing hype or investor pressure rather than clear business problems, lack of suitable data and MLOps capacity, and attempts to shoehorn LLMs into workflows that don’t need them. Others question the reliability of the 80% figure and argue that even a 20% success rate could be acceptable if the winning projects deliver outsized returns, while warning that many current efforts resemble past bubbles like blockchain.
Link & availability
- Original blog repeatedly returned database errors; several users relied on archived copies.
- Multiple commenters linked directly to the underlying RAND report instead.
Definition and reliability of the “80% fail” figure
- RAND focuses on machine-learning-based projects inside existing organizations, excluding pure “prompt engineering” wrappers around pretrained LLMs.
- The 80% number does not originate in RAND itself; it traces to a business article citing unspecified executive surveys (83–92% failure) without primary data.
- Several commenters flag this as methodologically weak and advise skepticism about the exact percentage.
Reported causes of AI project failure
From RAND summary and discussion:
- Problem selection: stakeholders misidentify or miscommunicate what the AI system is supposed to solve; same root cause as many failed software projects.
- Data: organizations often lack sufficient, appropriate, or high-quality data; many hoard user data instead of generating the domain-specific documentation/expert material LLMs actually need.
- Shiny-object syndrome: teams prioritize “latest tech” and AI branding over solving concrete user problems.
- Infrastructure & MLOps: inadequate data pipelines, deployment infrastructure, and too few data engineers; ML specialists end up maintaining brittle data code.
- Overreach: some projects tackle problems that are currently too hard for AI, or where any misprediction is too costly.
How bad is 80%? Comparisons and ambiguity
- Thread notes claims that “non‑AI IT projects” fail at roughly half that rate, but other sources in the discussion say 60–70% of software projects in general fail.
- Some see a 20% success rate for bleeding-edge tech as quite good; others say it’s worse than mature IT and may still be overstated because many projects haven’t failed yet.
Hype, management behavior, and internal politics
- Many anecdotes of executives demanding “AI everywhere” for optics, often ignoring technical staff and basic ROI analysis.
- Others describe the opposite: AI and LLMs blocked over security/compliance fears, or leaders being ultra‑conservative on spend.
- Several compare this wave to earlier fads (blockchain, NoSQL, microservices, 3D movies), with “insert new tech into everything” mandates and predictable waste.
Data, cost, and feasibility constraints
- Training serious models is described as prohibitively expensive; code is easy, data and compute are hard.
- A common failure mode: AI used to paper over missing process, tooling, or documentation; without those, LLMs add little value.
Is the money “wasted”?
- “Billions wasted” is debated: funds mostly turn into salaries, cloud bills, and hardware.
- From a VC or portfolio view, high failure is acceptable if a few “black swan” wins pay for the rest; for enterprises seeking incremental automation, repeated failures hurt more.
Where AI appears to work
- Coding assistance is repeatedly cited as a genuine productivity booster for some engineers, especially for tedious refactors and boilerplate, though others remain unconvinced or concerned about quality.
- Commenters stress that many so‑called “AI products” are thin chat‑bot or LLM API wrappers with little differentiated value, contributing to the high apparent failure rate.