25 Tech Secrets That Can Change Your Digital Life in 2026
Why Most AI Projects Actually Fail
The 95% failure figure is disputed. The failure itself is not. Here’s the honest data, the four patterns that predict failure, and what the 6% who actually succeed do differently — in exact sequence.
- Between 50–80% of AI projects fail to reach meaningful production — not 95%. That specific number is methodologically contested.
- The successful minority isn’t smarter. They sequence differently: business outcome → data infrastructure → workflow redesign → technology selection. Most teams run this backwards.
- The four root causes are data quality, technology-first thinking, the pilot-to-production gap, and change management failure. They’re predictable and avoidable.
- If you can’t answer the four Phase 1 questions with specific numbers, you’re not behind on AI. You’re ahead on risk avoidance.
Every AI post opens with a scary number. This one starts with a caveat about the scary number — and I think that caveat is the most useful thing you’ll read today.
The “95% of AI projects fail” headline circulating in 2025 comes from a real MIT report. The report is legitimate: 300+ initiative reviews, 52 interviews, 153 surveys. But Wharton Professor Kevin Werbach, after reading the report multiple times, said publicly he cannot find where the 95% figure comes from in the underlying data. That’s a problem worth naming before you cite the stat in a board presentation.
So what does the data actually show? Somewhere between 50% and 80% of AI projects fail to reach meaningful production deployment, depending on how failure is defined and what stage you’re measuring. The honest range is less sexy than 95%. It’s also more useful — because it tells you you’re dealing with a coin-flip problem, not an industry-wide catastrophe, and those require different responses.
01Before the Numbers: The 95% Problem
The figure originates from MIT’s NANDA Project report “The GenAI Divide” (July 2025). The report defines “failure” as anything short of “marked and sustained P&L impact” — a bar high enough to exclude meaningful partial wins. And MIT NANDA’s own research agenda centers on agentic AI protocols, which conveniently makes the 95% failure rate the exact problem their technology claims to solve. Conflict of interest noted. Use with full disclosure. CONTESTED
I’m not saying the directional finding is wrong — it isn’t. Multiple independent sources confirm that most pilots don’t reach full enterprise deployment. I’m saying the specific figure should be treated with the same skepticism you’d apply to a vendor’s benchmark. Cite it as “reported by MIT NANDA, methodology contested.” That’s the honest version.
02What the Verified Data Actually Shows
Here’s each major figure with what it actually measures. These are not interchangeable — they measure different stages of the same failure pattern.
| Statistic | Source | What It Actually Measures | Evidence |
|---|---|---|---|
| Only 6% of orgs qualify as AI high performers with measurable EBIT impact | McKinsey, Nov 2025 (n=1,993) | Full P&L impact — highest bar, most credible | ESTABLISHED |
| Only 25% of C-suite execs say their org is realising significant value from AI | BCG AI Radar, Jan 2025 (n=1,803) | Self-reported value perception — softer bar but broad sample | ESTABLISHED |
| Fewer than 1 in 3 GenAI experiments moved to production | Deloitte Q4 2024 | Experiment-to-production rate — measures the transition gap | ESTABLISHED |
| 42% of companies scrapped most AI initiatives in 2025, up from 17% in 2024 | S&P Global 2025 (n=1,000+) Note: sample skews toward companies already struggling. | Abandonment rate — biased toward failure cases | PROBABLE |
| 46% of AI POCs abandoned before production | S&P Global 2025 | Pilot survival rate — clean, specific metric | ESTABLISHED |
| AI projects fail at roughly double the rate of traditional IT projects | RAND Corporation, Aug 2024 (65 interviews) Note: RAND cites this from prior literature, not original count. | Relative failure rate vs. baseline — directional | PROBABLE |
| 70–85% of GenAI deployments fail to meet expected ROI | NTT DATA, 2024 | ROI attainment — high bar, includes reached-production-but-missed-targets | PROBABLE |
| 70% of digital transformation initiatives fail to meet objectives | Gartner CIO Survey, 2025 (n=3,186) | DX baseline — AI projects are a subset of this | ESTABLISHED |
Two statistics removed from this analysis: The “$2.3 trillion lost annually” attributed to Gartner has no traceable primary source. The “88% of pilots never reach production” attributed to IDC circulates widely but has no accessible primary source. Both were cut. If you can’t verify it, don’t publish it.
03Why These Numbers Seem to Contradict Each Other
Executives regularly see “74% of companies report ROI on AI pilots” sitting next to “80% of projects fail” and assume someone’s data is wrong. It isn’t. They’re measuring different stages of the same lifecycle.
The gap is in the transition. High pilot ROI measures whether a proof-of-concept delivered value in a sandboxed environment — with cherry-picked data, limited scope, and senior attention. Most do. They’re designed to. “We saw great pilot results” is not evidence that you’re in the successful minority. Pilot success is necessary but not sufficient.
Most projects that reach the pilot-success stage still fail at deployment because production data pipelines couldn’t handle volume, integration was harder than scoped, end users didn’t adopt, and the pilot team moved on. This is the honest version.
There’s also a second contradiction worth resolving: the 50–80% failure range itself. Which end applies to you depends on three variables: whether you’re running traditional ML or GenAI (GenAI failure rates appear higher); whether you’re a large enterprise or an SMB (larger organizations fail at production more because production requirements are harder); and whether you’re measuring “didn’t reach production” or “reached production but missed ROI targets” — the latter is a higher bar. Being honest about which scenario matches yours is the first practical step. → AI ROI Measurement
04Why Projects Actually Fail: The Four Root Causes
I’ve spent time reviewing findings from RAND’s 65 practitioner interviews, McKinsey’s 2025 survey, Informatica CDO Insights 2025, and BCG’s research with 1,803 C-suite executives. Four patterns dominate. Every other failure mode is downstream of one of these four.
1. The Data Quality Crisis
| Statistic | Source | Evidence |
|---|---|---|
| 43% cite data quality as top obstacle | Informatica CDO Insights, 2025 | ESTABLISHED |
| 64% name data integrity as their #1 challenge | Precisely, 2025 | ESTABLISHED |
| ~80% of ML work is data preparation, not modelling | RAND synthesis, 2024 | PROBABLE |
The data quality problem existed before GenAI. GenAI just makes it louder and faster. Bad training data produces inaccurate outputs. Bad RAG implementations hallucinate in live customer conversations — at scale, in real time, in front of customers who then screenshot it and post it. The organizations that succeed earmark 50–70% of the timeline and budget for data readiness before any model work begins. Not as bureaucratic prerequisite — as the actual foundation that determines whether the model will be reliable at all. → AI Data Infrastructure
2. Technology-First Thinking
“People said, ‘Step one: we’re going to use LLMs. Step two: What should we use them for?'” — Marina Danilevsky, IBM Research
This is the single most reliable predictor of failure, confirmed independently by RAND, McKinsey, and BCG. The pattern is almost always the same: leadership gets excited about a specific capability (usually after a demo), commits budget to the technology before defining the business problem, and then spends the next six months retrofitting a business case around the tool they already bought.
| Statistic | Source | Evidence |
|---|---|---|
| 60% of companies track no KPIs tied to AI value creation | BCG, Jan 2025 | ESTABLISHED |
| AI high performers are 3× more likely to have fundamentally redesigned workflows before deploying AI | McKinsey, Nov 2025 | ESTABLISHED |
The fix is counterintuitive: start with a specific, quantifiable business pain. Draft AI specifications only after stakeholders can articulate the current manual process cost and what a measurably better outcome looks like in concrete numbers. If they can’t do that, you don’t have a project yet — you have a technology looking for a problem. → Building the AI Business Case
3. The Pilot-to-Production Valley of Death
| Statistic | Source | Evidence |
|---|---|---|
| 46% of POCs scrapped before production | S&P Global, 2025 | ESTABLISHED |
| Fewer than 1 in 3 GenAI experiments moved to production | Deloitte Q4 2024 | ESTABLISHED |
| Median time from prototype to production: 8 months | Gartner, 2024 | PROBABLE |
Most organizations discover production requirements — scale, latency, compliance, monitoring, edge cases — only after declaring the pilot a success and spending the budget. This isn’t stupidity. It’s a structural problem with how pilots are funded and scoped. The pilot team optimizes for demo-readiness. The production team inherits a system that was never built for them. Plan for production from day one. Build observability in the first sprint. Assign product managers to model services with explicit SLOs before the pilot ends.
4. The People Problem Nobody Talks About
| Statistic | Source | Evidence |
|---|---|---|
| 35% cite skills shortage as top obstacle | Informatica CDO Insights, 2025 | ESTABLISHED |
| 75% of organizations have reached or exceeded their change saturation point | Prosci, 2024 | PROBABLE |
The technical team delivers the model. The organization delivers the adoption. Most projects optimize the first and ignore the second completely. I’ve seen this pattern enough times to stop being surprised: a technically elegant system, six months of work, and a user base that routes around it because nobody asked them what they actually needed before building. Top-performing organizations co-own digital delivery between IT and business leaders, and they forecast digital skill needs across the entire organization — not just technical teams — building capability before deployment, not after go-live.
05What Success Actually Looks Like
| Metric | Struggling majority | High performers (6% of orgs) |
|---|---|---|
| Time to production | 8+ months median; many never launch | 90 days to measurable impact |
| Budget allocation | 70% on technology | 50–70% on data infrastructure first |
| KPI tracking | 60% have no defined AI value KPIs | Explicit SLOs defined before development begins |
| Workflow approach | AI bolted onto existing processes | Workflows redesigned before tech selection (3× more likely) |
| Strategic ambition | Cost reduction targets only | Revenue and innovation goals alongside efficiency |
Both McKinsey’s 2025 survey and BCG’s AI Radar 2025 sample organizations that are already using AI — so these figures represent achievable outcomes for committed implementations. Not guaranteed averages. The 90-days-to-measurable-impact figure in particular applies to organizations that had already done the data and workflow work before the clock started. If you’re starting from zero, add time for that.
06The 5-Phase Sequence That Separates the 6%
The pattern isn’t mysterious. High performers follow a sequence that most people skip or shortcut. The sequence matters more than the individual steps — running Phase 4 before Phase 2 is the most expensive mistake in AI project management.
Business-First Scoping
Four questions that need specific answers before any technical work begins: What specific business outcome will this deliver — in measurable numbers? What is the current cost of not solving this problem? Can all stakeholders articulate the manual workflow and its pain points? What would a 10× improvement look like in concrete terms?
If you can’t answer all four with specifics, pause. Every week you proceed without this foundation increases failure probability.
Data Infrastructure — Where Projects Actually Win or Lose
Allocate 50–70% of the total timeline and budget here. Most organizations underinvest in this phase. Weeks 3–4: audit all data sources, quantify gaps and remediation costs. Weeks 5–8: build extraction, normalization, and governance pipelines. Weeks 9–12: deploy observability infrastructure — drift detection, event logging, SRE integration.
This isn’t setup. This is the actual work.
Workflow Redesign Before Technology — The Step Most Teams Skip
Adding AI to a broken process doesn’t fix the process — it accelerates the breakdown. Map the current end-to-end workflow with all stakeholders present. Identify bottlenecks, handoffs, and manual intervention points. Redesign for AI–human collaboration, not replacement. Get sign-off from all process owners before proceeding.
Skipping this step is the most expensive mistake you can make at this stage.
Technology Selection and Pilot
Only now — with clear requirements, clean data, and redesigned workflows — can you make an informed technology decision. Define success metrics, SLOs, and accuracy thresholds before the pilot begins, not after. Build production-ready infrastructure from day one. Include real end users in testing from week one, not just technical teams.
The data consistently favours purchased solutions over internal builds. Treat this as directional, not a rule.
Production Deployment with Product Thinking
The project doesn’t end at launch. Most failures occur here because organizations treat deployment as the final step rather than the starting line. Assign a product manager — not just a data science team. Establish on-call monitoring with defined escalation paths. Tie all success metrics to P&L outcomes, not just accuracy. Build explicit feedback loops from end users into improvement cycles.
07Kill Criteria: When to Stop
The organizations in the successful minority know when to cut. The majority waste 8+ months before acknowledging defeat. Killing a failing initiative at month 1–2 is a success. Three projects killed early outperform one zombie initiative that consumed years of organizational goodwill and budget.
⛔ Month 1 — Stop if any of these are true
- Stakeholders cannot define the business outcome in specific, measurable terms
- Data quality assessment shows more than 40% of required data is missing or unusable
- Remediation costs would exceed 70% of the total project budget
- End users are resistant to workflow redesign conversations before technology is even selected
- Leadership says “just try it and see” without success metrics defined
⛔ Month 3 — Stop if any of these are true
- Data pipeline construction is taking 3× the original estimate with no clear endpoint
- Quality metrics not improving despite active remediation efforts
- Workflow redesign keeps getting deferred by competing priorities
- Technology selection is happening before the data and workflow work is complete
⛔ Month 6 — Stop if any of these are true
- Pilot accuracy below 70% after multiple iterations
- End users bypassing the system to do the task manually
- The technical team cannot explain in plain language how the system makes decisions
- Success metrics are continually adjusted to avoid revealing failure
- No credible path to production deployment exists
This framework is built on survey data, not controlled experiments. Here’s what I’m uncertain about and where the framework might be wrong.
This framework assumes a greenfield project structure
The 5-phase sequence assumes you have the organizational authority to run phases in order. Most real projects don’t — you inherit a partially-built system, a vendor commitment already made, or a pilot already underway. The framework is a diagnostic tool in those situations, not a prescription. SPECULATIVE
The 6% McKinsey figure may undercount value
McKinsey’s definition of “high performer” requires measurable EBIT impact. This excludes organizations running AI in ways that deliver real value but aren’t yet at P&L scale — cost avoidance, risk reduction, process speed. The 6% figure is the right benchmark for aspirational targets, but it may make the situation look bleaker than it is for earlier-stage deployments. PROBABLE
The failure rate may be dropping faster than the data shows
GenAI tooling in 2026 is significantly more production-ready than in 2023–2024. The S&P Global and Deloitte figures were collected during a period of rapid tooling change. It’s plausible that failure rates are declining as platforms mature, vendor support improves, and organizational learning compounds. The honest answer is: we don’t have good 2026 longitudinal data yet. SPECULATIVE
Sample bias in most surveys is significant
The S&P Global sample is explicitly skewed toward companies that have already reported struggling. McKinsey and BCG surveys skew toward large enterprises. SMB failure rates may be different — possibly lower, because production requirements are simpler. My analysis skews toward enterprise. If you’re an SMB: adjust accordingly. PROBABLE
08Frequently Asked Questions
Sources
All linked sources are publicly accessible. Evidence classification reflects verification status as of April 2026.
McKinsey: The State of AI in 2025 — n=1,993 participants, 105 nations, June–July 2025. Key finding: only 6% of organizations qualify as AI high performers with measurable EBIT impact.
BCG: From Potential to Profit: Closing the AI Impact Gap — n=1,803 C-suite executives, 19 markets, January 2025. Key finding: only 25% report meaningful value from AI.
Deloitte: State of Generative AI in the Enterprise — Q4 2024. Key finding: fewer than 1 in 3 GenAI experiments have moved to production.
S&P Global: AI Initiative Survey 2025 — n=1,000+, North America and Europe. Key findings: 42% scrapped most initiatives; 46% of POCs abandoned pre-production. Note: sample skews toward companies already reporting difficulties.
Gartner: CIO and Technology Executive Survey 2025 — n=3,186, conducted October 2024.
RAND Corporation: The Root Causes of Failure for AI Projects — August 2024, 65 practitioner interviews. Note: the “double failure rate” figure is cited from prior literature, not RAND’s own count. RAND’s original contribution is identifying root causes.
Informatica: CDO Insights Survey 2025 · Precisely: Data Integrity Trends Report 2025 · NTT DATA: GenAI Deployment ROI Analysis 2024 · Prosci: Change Management Benchmark Research 2024
MIT NANDA Project: “The GenAI Divide” (July 2025) — The report is authentic; the headline 95% figure is disputed. Wharton Professor Kevin Werbach publicly questioned the methodology. Conflict of interest noted: MIT NANDA researches and promotes agentic AI protocols. Use with full disclosure.
Two statistics removed from analysis: “$2.3 trillion lost annually” (Gartner — no accessible primary source confirmed). “88% of pilots never reach production” (IDC — widely cited but no primary source accessible). If you can’t verify it, don’t publish it.
Related on FutureNow: Building the AI Business Case · AI Data Infrastructure Guide · The K-Shaped Economy and AI Divergence · Measuring AI ROI Without the Hype
Future Side Hustles That Pay More Than a Job
You’re Paying for These Three Google Workspace Tools. You’re Not Using Them.
I Tested 3 MacBook Neos for 30 Days. Here’s the Exact Workflow That Breaks 8GB.
AI Jobs 2026: Why Regional Data Tells a Different Story Than Global Headlines
7 Skills You Need Before AI Replaces You
Weird Scientific Mysteries Scientists Still Don’t Understand
The 2026 Smartwatch That Replaces Your Doctor—Shockingly Accurate
1 thought on “Why Most AI Projects Fail: Common Causes”