Why Most AI Projects Fail, and How to Be in the 20%

How often do AI projects actually fail?

By some estimates, more than 80% of AI projects fail to deliver value, roughly twice the failure rate of conventional IT projects (RAND, 2024). And the trend is worsening: the share of companies abandoning most of their AI initiatives before production jumped from 17% to 42% in a single year (S&P Global, 2025).

The generative-AI numbers are starker still. MIT's 2025 NANDA study, The GenAI Divide, found that around 95% of enterprise generative-AI pilots deliver no measurable impact on the P&L, only about 5% drive real value. Adoption is near-universal (78% of organisations now use AI, per McKinsey, 2025), yet 74% of companies struggle to achieve and scale value from it (BCG, 2024). The gap between using AI and getting value from it has never been wider.

Why AI projects fail: people and process, not technology

The single most useful finding in the research is BCG's 10-20-70 rule: success with AI is roughly 10% algorithms, 20% technology and data, and 70% people and process (BCG, 2024). Most organisations invert it (months evaluating models, weeks on change management) and then wonder why adoption stalls.

This reframes the whole question. It isn't "what can we automate?" It's "what outcome matters, and what's actually blocking it?" The companies that get this right pull ahead measurably: BCG's "future-built" firms achieve about 1.7× the revenue growth and 1.6× the margin of laggards (BCG, 2025). They aren't playing a different game: they're playing the same game with better foundations.

The real root causes

Failures cluster around six causes, and only one of them is really about technology. The rest are about data, process, people and money.

Two of these deserve a number. Poor data quality is the most-cited blocker: 43% of organisations name data quality or readiness as their single biggest barrier to AI success (Informatica, 2025). And the money rarely behaves: around 85% of organisations misestimate their AI costs by more than 10%, with nearly a quarter off by 50% or more (Mavvrik/Benchmarkit, 2025). Gartner has predicted that roughly 30% of generative-AI projects will be abandoned after proof of concept (Gartner, 2024).

The pitfalls that sink projects

Beyond the root causes, five recurring traps derail otherwise promising initiatives. Each has a simple antidote, none of which is "buy more technology".

• AI theatre. Chasing visible, impressive-looking AI instead of the boring back-office work that actually pays back. (Amazon's "Just Walk Out" checkout, marketed as fully autonomous, reportedly leaned on around 1,000 workers in India to label and validate transactions, and was pulled from Amazon Fresh stores in 2024. Amazon disputes the "watching" framing.) Antidote: solve dull problems that cost real money.
• Technology-first thinking. Starting from "we need AI" and hunting for a use case. Antidote: before any initiative, answer three questions: can we measure success? do we have clean, owned data for this? have we fixed the underlying process? If not all three are "yes", you're not ready.
• Underinvesting in data. Spending on models while the data foundation rots. Antidote: expect data readiness to take the majority of the effort, not a footnote.
• Neglecting change management. 45% of CEOs say most employees are resistant or hostile to AI (Kyndryl, 2025), and around 75% of workers worry AI could make their jobs obsolete (EY). A tool nobody trusts doesn't get used. Antidote: involve people early, communicate honestly, train, and embed AI in the tools they already use.
• Unrealistic timelines. Expecting quarterly returns from a multi-year change. Antidote: set board expectations for one-to-three-year value, with faster wins on narrow use cases.

How to be in the 20%

The organisations that succeed don't have better algorithms: they do the groundwork first. This is the sequence that separates the 20% from the rest.

1. Start with a workflow that matters. Not the most complex or most political one. Pick something contained that costs real time or money and whose outcome you can measure: invoice processing, lead qualification, document review.
2. Map it end-to-end as it actually works. Not the official version. Who touches it, which systems are involved, where it gets stuck, where people email spreadsheets because the systems don't talk.
3. Fix before you automate. Clean up the workflow first. Once it's clear, the AI part becomes almost obvious: you can see where automation helps and where human judgement still matters.
4. Define success metrics and a baseline. What does success look like in numbers, what's the current baseline, and what timeline is realistic? If the answer is "efficiency" with no number, you're not ready.
5. Run a controlled pilot. Small team, short window, measure everything, gather feedback, iterate. Expand only when results warrant it.
6. Scale systematically. Let operational reality (not a vendor timeline) set the pace.

How AI colleagues change the odds

If 70% of the problem is people and process, the tool you choose should be one that fits how people already work, not another system to adopt, learn and babysit. That's the design idea behind Frntir's AI Synths.

A Synth is a named AI colleague with persistent memory that works inside the tools your team already uses, operates with bounded, explainable autonomy under a human boss, and keeps a full audit trail. That directly targets the failure modes above: it lowers the change-management barrier (it lives in your existing email, chat and docs rather than being a new destination), it accumulates institutional knowledge rather than losing it, and its actions stay visible and correctable. It doesn't remove the need to fix your data and processes (nothing does), but it attacks the 70% instead of adding to it. If you want to see what that looks like for a specific job in your business, book a call.