All articles
AI StrategySMB GuideResearch

Why 95% of AI Projects Fail —
And How to Be in the 5%

The MIT stat everyone's quoting is real. The conclusion everyone draws from it is wrong. A cross-source breakdown of 8 independent studies, the 3 actual failure modes, and a 3-question diagnostic you can run on your business today.

11 min read
By Iu Ayala, Gradient Insight
95%
AI projects → zero ROI
MIT NANDA, 2025
6%
see significant financial impact
McKinsey, 1,993 executives
42%
abandoned AI in 2025
was 17% in 2024 — S&P Global
2–4y
when ROI actually appears
Deloitte longitudinal study

Prefer to watch? The full video is above. The post goes deeper on the data and case studies.

What MIT Actually Found

Before reacting to the headline number, read what the study actually measured.

The MIT NANDA "State of AI in Business 2025" report is the source behind the 95% figure everyone's citing. I read it in full when it came out — and the first thing worth knowing is what it actually studied.

MIT NANDA — Study Methodology

52
structured interviews
153
surveys
300+
AI initiatives disclosed
Directional
not statistically precise

The report's own disclaimer: the data is "directionally accurate" — it tells you which way the wind blows, not a precise number to four decimal places. That caveat matters, but direction is what matters most here.

The researchers are upfront: the sample isn't statistically representative. A Wharton professor noted there's limited support for how exactly they derived the 95% figure. That's a legitimate point. But here's the thing — when you stack this against seven other independent studies, the direction doesn't change.

80% tried GenAI
95% of those saw zero ROI
5% of custom AI tools
ever reach production
~69% of orgs
have employees bypassing official AI tools — using personal Claude or ChatGPT instead (Gartner, 2025)
>50% of AI budget
goes to sales & marketing — the lowest-ROI category by far

The hidden adoption story

The enterprise spent millions on a sanctioned AI platform. Most employees are ignoring it and using their personal Claude or ChatGPT account instead. Value is happening — it's just not the kind anyone bought, measured, or reported back.

"The 95% figure has been contested. That's a legitimate point about methodology. But when you stack this against seven other independent studies, the direction doesn't change. Something is genuinely broken."

8 Studies. Same Direction.

MIT's methodology is contested. The cross-source pattern is not.

McKinsey surveyed nearly 2,000 organizations. Only one in twenty reported significant financial impact from AI. Eighty-eight percent use it. Six percent win from it.

BCG ran their study across a thousand executives in nearly sixty countries. Three quarters say they struggle to scale AI value. Only four percent — four — have what BCG calls genuine, cutting-edge AI capability.

IBM interviewed two thousand CEOs directly. Three quarters said their AI initiatives failed to deliver expected ROI within three years.

Source
Sample
Key finding
Signal
MIT NANDA
300+ initiatives
95% of GenAI: zero ROI
Confirms
McKinsey
1,993 executives
Only 6% see significant impact
Confirms
BCG
1,000+ CxOs, 59 countries
74% struggle to scale
Confirms
IBM
2,000 CEOs
75% failed expected ROI in 3 years
Confirms
S&P Global
Enterprise tracking
42% abandoned AI in 2025
Confirms
RAND
Large-scale survey
>80% report underperformance
Confirms
Accenture
Large enterprises
74% met or exceeded expectations
Partial ⚡
Deloitte
Multi-year tracking
66% see productivity gains
Mixed

S&P Global — AI Project Abandonment Rate

The failure isn't slowing down. It's accelerating.

2024 17% abandoned most AI initiatives
17%
2025 42% abandoned most AI initiatives
42%

2.5× increase in one year. More companies run AI experiments without a clear framework — and walk away faster when they don't see results in the first six months.

Why some studies contradict each other

Productivity ≠ revenue. Meeting expectations ≠ positive ROI. The studies that show "AI is working" and the studies that show "AI is failing" are measuring entirely different things. The ones measuring financial impact all point the same direction.

The 3 Real Failure Modes

Not abstractions — patterns I've watched happen in real projects with real clients.

01
Wrong Timeframe
Measuring at 6 months when value arrives in 2–4 years.
02
Wrong Use Case
>50% of AI budget goes to the lowest-ROI category.
03
Wrong Approach
Broad 15-use-case rollouts instead of one deep integration.

Failure 1 — Wrong Timeframe

Deloitte's longitudinal research found that significant ROI from AI materialises in two to four years. The MIT study measured at six months. When you measure at six months, you'll always see 95% failure — because you're measuring before the return exists.

Measuring AI ROI at six months is like measuring whether a new hire was worth it after their first week. Nobody manages a business like that — except when it's AI.

Month 1–3
Month 6–12
Year 2–4

Efficiency Metrics

Time saved per task
Error rate reduction
Volume processed

Operational Metrics

Cost per transaction
Headcount impact
Process throughput

Financial Metrics

Revenue enabled
Cost base reduction
Compounding returns

I've seen this pattern kill good projects. One client was three months into a document automation system that was genuinely transforming their back-office workflow — leadership wanted to shut it down because the ROI report didn't look right. The efficiency gains were real and measurable. The financial returns hadn't compounded yet. They almost pulled it.

Failure 2 — Wrong Use Case

More than 50% of enterprise AI budget flows to sales and marketing. Which, according to the same MIT data, delivers the lowest return on investment of any use case. Meanwhile, back-office automation — replacing repetitive, high-volume document processing — consistently delivers the highest returns.

Where the budget goes vs. where the ROI is

Sales & Marketing AI >50% of budget · Lowest ROI
55%+
Back-office Automation Underinvested · Highest ROI
~20%

Here's a real example. DG Financial came to us needing to search across 80,000 documents spread across multiple cloud platforms — multiple formats, scanned PDFs, handwritten notes. We built a pipeline that made all of it searchable in under a second. Their team was spending hours every week manually hunting through files. That time disappeared almost immediately.

DG Financial — Document Search Pipeline

S3
OneDrive
SharePoint
sources
OCR / STT
all formats
Elasticsearch
80,000 docs
<1 second
results

"Nobody puts document search on a conference slide. Everyone wants the AI that writes personalised emails. But the document search is what makes money. Back-office, not front-office. High-frequency, repetitive, boring — and that's exactly why the ROI is there."

Failure 3 — Wrong Approach

The MIT finding here is explicit: pick one pain point, execute well, partner smartly. Only one in five companies actually redesigns their workflows around AI. The rest layer AI on top of existing processes and wonder why nothing changes.

The 95% Approach

5,000 users, 15 use cases
Shallow depth on each
AI layered on existing process
No one owns any single workflow

The 5% Approach

1 use case, deeply integrated
Workflow redesigned around AI
Measurable outcome defined upfront
One team owns it end to end

Angel Protection came to us with one specific problem: weapon detection across hundreds of live camera feeds simultaneously. Not "AI for security." One detection task, one hardware constraint, one latency target. That system now runs at sub-300ms latency and is part of their core product. Narrow scope. Deep integration. Measurable outcome.

The 5% Blueprint

Three questions to answer before you start — or before you kill — any AI initiative.

The 5% Blueprint — 3 Questions

1

What is the one workflow where a manual task takes 2+ hours and happens 3+ times per week?

Not 'where can AI help?' — that question gets you 15 answers and no decisions. Name one specific, recurring, painful task.

2

Are we measuring the right thing in the right timeframe?

Month 1–3: efficiency metrics. Year 2–4: financial metrics. Budget for the longer horizon and set expectations accordingly.

3

Build, buy, or partner — and does our team have the capacity to own whichever path we choose?

Vendor solutions succeed at roughly 2× the rate of internal builds. Be honest about your team's capacity to maintain what gets built.

If you can't name one specific workflow in 30 seconds...

...you are not ready to start an AI project yet. That's not an insult — it's the most useful thing I can tell you. The first step isn't technology. It's clarity about the problem.

Question 3 expanded: Build, Buy, or Partner?

Build

High risk

Custom requirements + your team can own and maintain it long-term

If the key developer leaves, the system becomes a liability
Highest success rate

Buy

Low risk

Generic use case — a vendor has already solved it well

Fastest path, lowest maintenance burden, proven at scale

Partner

Medium risk

Proprietary data + custom workflows where off-the-shelf can't go

Ask: is something they built 2 years ago still running in production?

The SMB Advantage

The enterprise failure story gets one thing wrong about your company.

The companies in those stats are mostly huge: broad rollouts, committee-approved tools, year-long deployment timelines. Bureaucracy stops them from doing exactly what the 5% actually does. SMBs map directly onto the success playbook.

Path to ROI — Enterprise vs. SMB

Time →ROI ↑break-evenEnterpriseSMB

Same destination. SMBs get there faster because they can go narrow, deep, and iterate without committee approval.

Speed

One use case and a working prototype in weeks — not months of vendor approval and procurement theatre.

Depth

A small team with AI embedded in their real workflow outperforms 1,000 people ignoring a sanctioned tool.

Iteration

Feedback comes directly from the person doing the work — not filtered through five layers of management.

The stat the failure reports missed

69% of organizations report employees using unauthorized AI tools at work.

Bypassing whatever official tool the company paid for. The grassroots adoption story isn't in the failure statistics — but it tells you exactly how much demand exists once the right use case is in front of the right person.

DG Financial, Angel Protection, Salibo — all small-team builds by enterprise standards. No procurement theatre. One problem, one focused build, fast iteration. The 5% aren't the biggest companies. They're the most focused.

The whole game — in three lines

1
Narrow use case. One specific, recurring, painful workflow — not fifteen.
2
Right measurement timeframe. Efficiency at month 3. Finance at year 2. Don't confuse them.
3
Smart implementation approach. Build, buy, or partner — based on honest capacity, not aspiration.

That MIT number — the one everyone's using to say AI is broken — isn't a reason to wait. It's a blueprint. The useful question isn't "will this AI project succeed?" It's: are we doing the three things that separate the 5% from the 95%?

Gradient Insight

AI project not delivering ROI yet?

We help SMBs identify the one high-ROI workflow, build it right, and measure it correctly. Average client ROI: 300%+ in the first year. Free discovery call, no commitment.