There is a strange contradiction at the heart of the AI boom. Companies are spending more on artificial intelligence than ever, executives describe adopting it as existential, and yet the rate at which AI projects fail is not falling — it's climbing. The gap between AI ambition and AI results is widening, and the single biggest reason is unglamorous: most organizations' data simply isn't ready.
This article looks hard at that gap. It pulls together the most credible recent research on AI failure rates, examines why data unreadiness is the common thread, and walks through the integration challenges that turn promising pilots into abandoned proofs of concept. The picture it paints is sobering — but it also points to exactly what "ready" looks like.
Start with the failure numbers, because they are worse than most boards realize. Independent research firms studying very different samples keep arriving at the same uncomfortable conclusion.
In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing "poor data quality, inadequate risk controls, escalating costs or unclear business value."
"After last year's hype, executives are impatient to see returns on GenAI investments, yet organizations are struggling to prove and realize value." — Rita Sallam, Distinguished VP Analyst, Gartner
A year later, MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025, built on 150 leader interviews, a survey of 350 employees, and an analysis of 300 public AI deployments. Its headline finding was blunt: roughly 95% of enterprise generative-AI pilots produce no measurable return, while only about 5% drive real revenue acceleration. Crucially, MIT's researchers concluded the problem was not the quality of the models. It was a "learning gap" — flawed enterprise integration, weak workflow fit, and organizations that hadn't done the work to absorb the tools.
RAND, studying the question from an engineering angle through interviews with 65 experienced data scientists, found that more than 80% of AI projects fail — about double the failure rate of comparable IT projects that don't involve AI. And the trend is moving the wrong way: S&P Global Market Intelligence's 2025 survey of more than 1,000 enterprises found that 42% of companies abandoned the majority of their AI initiatives in 2025, up sharply from just 17% the year before, with the average organization scrapping 46% of its AI proofs of concept before they ever reached production.
Read those last two numbers together. Adoption is surging, but abandonment is surging faster. The companies rushing in are not, on the whole, succeeding — they are generating an expensive backlog of dead pilots. Analysts have started calling the mood "AI fatigue."
Dig beneath each of these studies and the same root cause appears again and again. The models are remarkably capable. The infrastructure is available on demand. What's missing is a clean, connected, governed data foundation for the AI to stand on.
Cisco's 2024 AI Readiness Index — based on responses from nearly 8,000 organizations — quantified the gap precisely. Only 13% of companies were fully ready to capture AI's potential, down from 14% the year before, even as 98% reported rising urgency to deploy. On data specifically, just 32% reported being highly ready, and 80% admitted to inconsistencies or shortcomings in how they pre-process and clean data for AI.
When Informatica surveyed chief data officers for its 2025 CDO Insights report, data quality and readiness tied for the single most-cited obstacle to AI success at 43% — level with technical maturity and ahead of the skills shortage. This is the part that should worry leaders: the barrier isn't the AI. It's the decades of data debt the AI is being asked to run on.
Andrew Ng, one of the field's most influential voices, has spent years arguing for a "data-centric" approach for exactly this reason. His framing is memorable:
"Data is food for AI." Feed a model messy, inconsistent, poorly governed data, and the old rule applies without mercy: garbage in, garbage out.
So what, concretely, makes enterprise data "not ready"? It tends to be five overlapping problems.
The single most common technical obstacle RAND identified was data fragmented across disconnected systems. The average enterprise runs hundreds of applications — CRM, ERP, support desks, data warehouses, spreadsheets, SaaS tools — each with its own copy of "the customer," its own identifiers, and its own definition of basic terms. An AI assistant asked a simple question ("what's our churn rate for enterprise accounts?") may find five different answers because five systems compute it differently. Before AI can reason over your business, someone has to reconcile that.
Even within a single system, real-world data is messy: missing fields, duplicate records, outdated values, free-text where structure was needed, and historical gaps where a process changed. Gartner estimates poor data quality costs the average organization $12.9 million a year even before AI enters the picture. Point a model at that data and it doesn't fix the errors — it confidently amplifies them.
AI needs to know what data means, who is allowed to see it, and whether it can be trusted. Most organizations lack a data catalog, consistent metric definitions, lineage tracking, or clear ownership. Without that governance layer, an AI system can't distinguish authoritative data from a stale export, and it can't enforce who should see what — which is how sensitive information ends up in the wrong answer.
The knowledge most companies actually want their AI to use — contracts, policies, tickets, emails, PDFs, wiki pages — is unstructured and scattered. Retrieval-augmented generation (RAG) and modern AI assistants can use it, but only after it's been collected, cleaned, chunked, embedded, and access-controlled. That preparation is real engineering work, and it is exactly the step most stalled pilots skipped.
Connecting an AI model to live business data creates a new and serious attack surface. If permissions aren't enforced at the data layer, an AI assistant becomes a fast, eloquent way to leak whatever it can reach. "Just give it access to everything" is how a productivity tool turns into a breach.
Data unreadiness is the deepest problem, but it is not the only one. The organizations that clean up their data still routinely fail at the integration — the work of fitting AI into how the business actually runs. MIT's "learning gap" lives here.
A pilot that works in a demo often dies on contact with a real workflow. If the AI output lands outside the tools people already use, requires a new habit, or doesn't fit the actual decision being made, adoption collapses. McKinsey's 2025 State of AI research found that while AI adoption is now widespread, only about one-third of organizations have actually scaled it across the enterprise — the gap between "we're using it" and "it's changing how we work" remains enormous. The winning pattern, McKinsey notes, is to "pick workflows with measurable business value, redesign the operating process around them, and keep the system running long enough to compound learning."
Gartner's list of abandonment reasons put "unclear business value" alongside data quality for good reason. Many pilots begin with the technology ("we should use GenAI") rather than a problem worth solving. RAND's researchers put it sharply: "Successful projects are laser-focused on the problem to be solved, not the technology to solve it." Projects without a measurable target have no way to prove they worked — so when budgets tighten, they're the first cut.
The talent gap is real — roughly a third of organizations cite it — but the deeper issue is organizational. RAND found that the leading causes of failure were less technical than cultural: misaligned purpose, miscommunication about intent, and fading executive sponsorship once the novelty wore off. AI changes jobs, and change that isn't managed gets quietly rejected.
S&P Global's respondents named cost and data privacy and security as their top obstacles. The financial burden of building, running, and maintaining AI systems is often underestimated at pilot stage and becomes painful at scale. And in regulated industries — healthcare, finance, government — inadequate governance and risk controls don't just slow projects down; they stop them, as they should.
It would be reasonable to expect failure rates to fall as the technology matures. Instead they're rising — abandonment jumped from 17% to 42% of companies in a single year. The boom itself is part of the explanation.
In 2024, organizations were experimenting. By 2025, boards expected returns. That shift from curiosity to pressure is precisely what exposes weak foundations: teams under pressure to "ship something with AI" skip the unglamorous data work, wire a model to whatever data is handy, demo it, and then discover it can't be trusted in production. The urgency that's supposed to drive success is, paradoxically, driving the shortcuts that cause failure.
The uncomfortable truth: AI doesn't fail because the models are bad. It fails because it's deployed on top of data and processes that were never ready for it — and the rush to adopt makes companies less likely, not more, to do the preparation that works.
The same research that documents the failures also points to what the successful 5% do differently. Readiness is not a model decision; it's a data, governance, and integration discipline. In practice it means:
None of this is exotic. It is the same data engineering, security, and integration discipline that separates durable software from demos — applied to AI. The companies that treat AI as a data-readiness program first and a model project second are the ones quietly landing in the 5%.
This is exactly the work Echo does. Our AI Integration Services start where most projects should: getting your data ready. We prepare and migrate data for AI (RAG, embeddings, knowledge bases), stand up MCP servers so AI assistants can securely read and act on your systems, and build Agentforce, Einstein, and LLM integrations with governance and least-privilege security designed in from day one — the same security-first discipline we bring to every Salesforce and information-security engagement.
If your organization is feeling the pressure to "do something with AI," the most valuable first step isn't a model. It's an honest look at whether your data is ready for one.
Talk to us about AI readiness →
Figures reflect the cited research at time of publication. Where firms have issued updated estimates, the most widely reported figure is used. This article is for general information and is not a guarantee of outcomes.
