Why 80% of AI projects in companies fail, an analysis of data from 1,300+ deployments, yesfor

A first note, before we begin

Most articles about AI project failures use a single number (usually "73% Gartner" or "95% MIT") and build an entire narrative on it. That is intellectually dishonest, because each of these numbers has its own methodology, its own sample, and its own definition of "failure".

This article does it differently. I show all six studies, their differences, their critics, and their convergences. The conclusion is surprisingly consistent, but you reach it through context, not through selective quoting.

Six studies, one conclusion

| Study | Year | Sample | Definition of "failure" | Result | |---|---|---|---|---| | RAND Corporation | 2024/2025 | 65 ML engineers | Failure to reach production | 80%+ | | MIT Project NANDA | 2025 | 300+ GenAI deployments | No P&L impact | 95% of pilots | | S&P Global | 2025 | 1000+ US/EU firms | Abandoning most initiatives | 42% of firms | | BCG "Widening AI Value Gap" | Q3 2025 | 1250 respondents | No material value | 60% of firms | | McKinsey Global AI Survey | 11/2025 | Global enterprise | No EBIT impact | 80%+ of organizations | | Gartner I&O Survey | Q1 2026 | 782 I&O leaders | Failure to achieve ROI | 72% of projects |

What these numbers do NOT mean

They do NOT mean that AI does not work technically. The models work. It is the organizations that are broken.
They do NOT mean that 80% of non-AI IT projects have better results, traditional IT projects have about 40% failures, and large IT projects (>USD 10M) have a 98% failure rate per the CHAOS report from the Standish Group.
They do NOT mean that "AI is a bubble". Firms in the 5-20% success group make real money on AI (Lumen Technologies: USD 50M/year in savings).

What they really mean

AI is a "hard"-class project and needs to be treated as a business transformation, not an IT deployment.

A critical analysis of the methodologies, which numbers are credible

RAND 80%, the most solid source

The report Ryseff, De Bruhl, Newberry 2024 is based on 65 structured interviews with experienced ML engineers and data scientists. The 80% figure comes from synthesizing data on hundreds of projects these people were involved in. The definition of failure: no "meaningful production deployment".

Strengths: interviews, not a survey. Experienced respondents. Concrete case files.

Weaknesses: the sample is limited to the US, a possible selection bias (those who saw failures are more willing to talk).

Verdict: The most solid number available. Repeated in 2025 with a similar result (Why AI Projects Fail, RAND PTA2680-1).

MIT 95%, loud, but controversial

The report The GenAI Divide: State of AI in Business 2025, 300+ GenAI deployments, 150 interviews. The 95% figure refers only to GenAI pilots, not to AI projects in general. The definition of success: a sustained, measurable P&L impact confirmed by users and the leadership team.

Critique Sean Goedecke 2025:

"Looking at MIT's methodology, 60% of firms started considering task-specific AI, 20% built a pilot, 5% built a deployment with measurable impact. That is an 8.3% success rate, not 5%, because 40% of firms did not try at all."

Verdict: The number is probably inflated. The real figure should be about 88-92%. But the conclusion stands: most GenAI pilots do not scale to production.

S&P Global 42%, solid, but about something else

A study of 1000+ firms in the US and Europe. The 42% figure refers to firms that abandoned most of their AI initiatives in 2025 (vs 17% in 2024). The average organization scrapped 46% of its pilots.

Verdict: Credible, but it measures the abandonment rate (giving up on projects), not the failure rate (not achieving value). Those two numbers are different, though related.

McKinsey, BCG, Gartner, corroborating evidence

All three measure a lack of EBIT/material value impact, not technical failure. Consistent results in the 60-80% range regardless of methodology confirm the main thesis.

Verdict: Consistency of results across different methodologies and samples = high confidence in the conclusion.

Seven real causes of failure, ordered by frequency

A synthesis of data from RAND (Ryseff 2024), MIT NANDA, McKinsey 2025, Gartner I&O 2026, and Informatica CDO 2025. The percentages show how often a given cause appears in failed projects (one project can have several).

Cause 1: No success metrics defined before the start, 73% of projects

The project kicks off because the board decided "we need AI". No one wrote down what exactly is supposed to change in the numbers after 6, 12, 24 months. After the first quarter the CFO asks "what did we get for PLN 2M?", and no one knows how to answer.

The failure mechanism:

Board decision: "we are deploying AI"
The CTO selects the technology (usually under pressure from a vendor)
The deployment starts with no baseline KPI
After 6 months: the CFO demands metrics
Metrics are reconstructed post-hoc, not credible
The CEO loses confidence, the project dies quietly

The cheapest cause to fix. One strategy workshop before the start (4-8 hours) is enough to settle: what the KPI is today, what it will be in 12 months, who measures, who reports, who to escalate to.

Cause 2: The data is not AI-ready, 43% of responses

Gartner defined "AI-ready data" as data meeting five conditions: a match to the use case, ownership at the asset level, automated pipelines with quality gates, live metadata, continuous quality control.

In Polish enterprises, typically:

The ERP was deployed 4-7 years ago, 30% of key fields empty or "n/a"
The CRM does not integrate with operational systems (everyone has their own version of the client)
The sales team keeps key data in spreadsheets (shadow IT)
The data warehouse was last refactored in 2021, 60% of tables unused
GDPR compliance "exists", but no one knows what exactly the client consented to

The consequence: AI trained on such data returns garbage. Just faster than a human. And with more confidence, which is even worse (hallucinations).

Cause 3: Loss of C-level sponsorship in the first 6 months, 56% of projects

RAND 2024 shows: projects with sustained CEO sponsorship have a 68% success rate. Projects where sponsorship withdraws after the first phase, 11%. A sixfold difference.

Sponsorship does not mean "the CEO said yes at the board meeting". Sponsorship means:

A specific person in the C-level has the project in their portfolio
They have KPIs tied to the project's success in their review
They take part in monthly project reviews (not quarterly)
They defend the budget against the CFO while the project is still pre-results
They intervene when operational departments sabotage the deployment

In most Polish companies, sponsorship ends at the signing of the contract with the provider.

Cause 4: Treating AI as an IT project, 61% of projects

The pattern: the board delegates "this AI thing" to the CTO. The CTO selects the technology. The operational department is handed a finished tool with an order to "use it". No one redesigned the processes. No one changed the KPIs. No one updated the compensation system.

McKinsey 2025 shows this unambiguously: organizations reporting "significant financial returns" are 2x more likely to redesign end-to-end business processes before choosing an AI model.

Put differently: AI without a process redesign = the automation of chaos. The chaos becomes faster and cheaper to produce.

Cause 5: Expectations too ambitious, the horizon too short, 57% of responses

I&O leaders in the Gartner Q1 2026 survey say it plainly: projects fell apart because "we expected too much, too fast". They assumed AI would instantly automate complex processes. After 90 days with no effect, a loss of confidence.

The real horizon for AI ROI in an enterprise, according to McKinsey:

Quick wins (automating 1-2 processes): 6-12 months
A measurable P&L improvement for a whole department: 18-24 months
An enterprise-wide transformation: 2-4 years

Vendors promise 6-12 months for everything. Boards sign, expecting 6-12. After a year, failure.

Cause 6: No co-design with end users, 50%+ of projects

This is a finding from RAND that most vendors ignore, because it hurts sales. The most common non-technical cause of failure is misaligned incentives + the absence of end-user co-design.

What this means in practice:

A tool designed by IT without consulting the people who will use it
No mechanism to gather user feedback after deployment
No iteration of the solution based on real usage
Employee bonuses still reward the "old way of working"

The consequence: end users (per the Writer 2025 study) deliberately sabotage AI deployments in 31% of cases. They do not use them. They enter bad data. They slow projects down.

Cause 7: The consulting dependency trap, appears in 40%+ of enterprise projects

Global spending on GenAI consulting in 2024: USD 3.75B (3x more than in 2023, National CIO Review 2025). The client pays the consultant for strategy. The consultant leaves. The capability leaves with them.

After 12 months the company realizes that:

It has no in-house people who understand the deployed solution
Every change requires re-ordering the consultant
The maintenance cost = 30-50% of the deployment cost per year
The vendor lock-in is so strong that migrating would cost a second deployment

That is why capability transfer, deliberately building competencies into the organization, is a key condition for success. A yesfor.ai audit always ends with a capability transfer plan.

Industry failure rates

Not all industries fare the same. Data from RAND and Pertama Partners 2026 show significant differences:

| Industry | Failure rate | Main cause in the industry | |---|---|---| | Financial Services | 82% | Compliance + legacy systems | | Healthcare | 79% | Data privacy + clinician resistance | | Manufacturing | 76% | Legacy OT/IT separation | | Retail | 74% | No unified customer data | | Professional Services | 69% | Workflow integration | | Construction | no separate data, estimate about 75-80% | Fragmentation of processes across sites |

The Polish context:

In Poland there is no dedicated study on the AI failure rate per industry. It can be estimated based on:

A lower level of investment per firm (Horvath 2026: 0.35% of turnover for EU mid-size vs a 0.5% global average)
A higher share of legacy ERP (60-70% of mid-sized firms on SAP/Comarch from before 2020)
A lower level of data literacy in the C-level (PARP 2024)

A conservative estimate of the failure rate in Polish mid-sized firms: 78-85%.

What the 5-20% who succeed do

All the serious studies (RAND, McKinsey 2025, BCG 2025, the Trullion AI Survivability Matrix) agree on four patterns of success:

Pattern 1: Workflow-First Design

Firms with measurable ROI are 2x more likely to redesign processes before choosing AI. The order: business → process → KPI → only then technology.

Pattern 2: Data Integration Priority

Firms with strong data integration achieve 10.3x ROI vs 3.7x for firms with weak integration (Integrate.io 2024). That is almost a threefold difference, on the same deployment budget.

Pattern 3: Realistic Scoping

The 5% that deliver rapid revenue acceleration in MIT NANDA do not attempt an enterprise-wide transformation in 12 months. They solve specific, well-defined problems with measurable outcomes, and expand from where it worked.

Pattern 4: Sustained C-level Involvement

68% success with sustained sponsorship vs 11% with its loss (RAND). Sponsorship does not end at the signing of the contract. It runs through the entire project cycle, including maintenance.

These 80% of failures are not the result of board stupidity. They result from AI being sold like software while requiring management like a business transformation. Boards that understand the difference are in the 20%. The rest invest in software and buy themselves problems.

— Arleta Marczyńska, Founder, yesfor.ai

What this means concretely for Polish companies

If you are on the board of a Polish company of 500+ people and are considering an AI deployment, the statistical probability says that your project will not deliver the promised value. This is not pessimism. It is the data.

Three things that genuinely change that probability:

1. A pre-deployment audit instead of jumping into the deep end

The audit diagnoses whether the company is ready before you spend money on deployment. The cost of the audit (PLN 25-200k) is 2-5% of a typical deployment budget. The audit does not guarantee success, but it eliminates 4 of the 7 most common causes of failure.

The full guide to the pre-deployment AI audit →

2. Success metrics defined BEFORE the start

One strategy workshop before the contract with a provider eliminates cause #1 (73% of projects). Time: 4-8 hours. Cost: the distance between the board and operations. Return: disproportionately high.

3. Capability transfer from day 1

Do not deploy AI without a plan for how you will maintain the system after the consultant leaves. Vendor lock-in and consulting dependency are more expensive than the project itself.

Self-test: are you in the 80% or the 20%?

Answer 7 questions honestly. The more "yes", the higher the probability that you are in the 20%.

Does the board have it in writing which KPIs are supposed to change as a result of the AI deployment?
Does the data for the planned use case meet Gartner's 5 dimensions of AI-ready data?
Does a specific person on the board have this project's success written into their annual review?
Have you designed (or do you plan to design) new business processes before choosing an AI tool?
Is the expected horizon for measurable ROI 18-36 months (and not 6-12)?
Do you have a plan for how to involve end users in the design of the solution?
Do you have a capability transfer plan, who in the organization will maintain the system after the consultants leave?

0-2 yes: Readiness is not in place. Deploying now = high statistical risk. We suggest an audit before continuing.

3-4 yes: Partial readiness. You can deploy quick wins. Large projects need refinement.

5-7 yes: High readiness. You are in the minority, a transformation strategy makes sense.

A more precise assessment comes from the AI Readiness Self-Assessment, 30 minutes, 36 questions, a measurable score of 0-100.

FAQ

The next step

The statistics are merciless, but deterministic, you can change them once you know the causes. Three paths:

AI Readiness Self-Assessment, check whether your company is in the 80% or the 20%. 30 minutes, free.
The full guide to the pre-deployment AI audit, what to do to avoid joining the statistic.
Discovery call, 30 minutes of conversation about your specific context.

All sources cited in the article

Primary research:

Ryseff, J., De Bruhl, B. F., Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation, RRA2680-1. Full report
Ryseff, J., Narayanan, A. (2025). Why AI Projects Fail. RAND Corporation, PTA2680-1. Webinar
Challapally, A. et al. (2025). The GenAI Divide: State of AI in Business 2025. MIT Project NANDA.
S&P Global Market Intelligence (2025). AI Initiative Abandonment Survey.
BCG (2025). The Widening AI Value Gap (September 2025 update).
McKinsey & Company (2025). Global AI Survey: November 2025.
Gartner (2026). AI Projects in I&O Stall Ahead of Meaningful ROI Returns.
Informatica (2025). CDO Insights Survey 2025.
Writer / Workplace Intelligence (2025). Generative AI in the Enterprise.
Horvath (2026). Mittelstand AI Investment Study.

Methodological critiques and analyses:

Goedecke, S. (2025). Is it worrying that 95% of AI enterprise projects fail? Link
Pertama Partners (2026). AI Project Failure Rate 2026: 80% Fail.
Talyx (2026). Why 90% of Enterprise AI Implementations Fail.