---
title: "Why Do AI Projects Stall on Data Rather Than Models?"
url: "https://atlan.com/know/ai-readiness/why-do-ai-projects-stall-on-data-rather-than-models/"
description: "AI projects stall on data rather than models because pilots run on clean extracts. See what breaks in production data, the research, and how to fix it."
author: "Emily Winks"
author_role: "Data Governance Expert"
published: "2026-10-21"
updated: "2026-10-21T00:00:00.000Z"
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

## Why do AI projects stall on data rather than models?

AI projects stall on data rather than models because the model is the one part a team can buy finished. The data has to be found, cleaned, defined and governed inside the company, and most pilots skip that work by running on a hand-picked extract. When the project meets production data spread across hundreds of systems, with no shared definitions and no lineage, the model gives confident wrong answers and nobody will let it act.

## Key takeaways

- The pilot and the production system run on different data, so they are different projects.
- A wrong definition stops being a wrong answer on a screen and becomes a wrong action in a system of record.
- Fixing the inputs is the only path that survives a model upgrade.
- The definitions, lineage and policies built for the first use case are the same ones the second needs.

---

AI projects stall on data rather than models because the model is the one part a team can buy finished. The data has to be found, cleaned, defined and governed inside the company, and most pilots skip that work by running on a hand-picked extract. When the project meets production data spread across hundreds of systems, with no shared definitions and no lineage, the model gives [confident wrong answers](https://atlan.com/know/llm-hallucinations/) and nobody will let it act.

---

## How much of AI failure is a data problem?

Data appears in every major study, and RAND ranks it the second most common root cause. [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk) that through 2026, organizations will abandon 60% of AI projects that lack [AI-ready data](https://atlan.com/know/ai-readiness/ai-ready-data/). Its July 2024 survey of 1,203 data management leaders found that 63% of organizations either do not have, or are unsure they have, the right [data management practices for AI](https://atlan.com/know/ai-readiness/data-management-for-llm-deployments/).

The other large studies point the same way:

| Study | Published | Headline finding | Where data shows up |
|---|---|---|---|
| [Gartner on generative AI](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025) | July 2024 | At least 30% of GenAI projects abandoned after proof of concept by the end of 2025 | [Poor data quality](https://atlan.com/know/data-quality-in-llms/) is the first of four named causes |
| [RAND, Ryseff, De Bruhl and Newberry](https://www.rand.org/pubs/research_reports/RRA2680-1.html) | August 2024 | Interviews with 65 experienced data scientists and engineers produced five root causes | Two are data: too little data to train a good model, and no [infrastructure to manage data](https://atlan.com/know/data-infrastructure-for-ai/) |
| [S&P Global Market Intelligence](https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/) | 2025 | 42% of companies abandoned most AI initiatives, up from 17%; the average firm scrapped 46% of proofs of concept | [Data privacy and security](https://atlan.com/know/data-privacy-for-ai-agents/) rank with cost as top obstacles |
| [Confluent 2026 Data Streaming Report](https://www.ibm.com/think/news/half-ai-projects-stalling-confluent-data-streaming-report) | June 2026 | Nearly half of 4,625 IT leaders report agentic AI projects delayed indefinitely or abandoned | 72% cite weak [real-time data infrastructure](https://atlan.com/know/ai-agent/data-for-ai/real-time-data-for-ai-agents/), up from 61% in 2025 |

Read the numbers with care. The Confluent figure covers agentic projects only, and the [RAND report](https://www.rand.org/pubs/research_reports/RRA2680-1.html) presents its often-quoted 80% failure rate as an outside estimate. What holds across these five sources is the shape of the list. Model capability does not lead any of them, and a data cause appears in every one.

---

## Why does the AI pilot work and production fail?

The pilot and the production system run on different data, so they are different projects. A pilot runs on an extract that someone cleaned by hand, with loose access rules, and a person checks every answer. Production removes all three conditions at once. [IDC research with Lenovo](https://www.cio.com/article/3850763/88-of-ai-pilots-fail-to-reach-production-but-thats-not-all-on-it.html) found that of every 33 AI proofs of concept, only four [reach production](https://atlan.com/know/ai-agent/ai-agent-scaling-in-production/).

| | Pilot | Production |
|---|---|---|
| Data | One curated extract | Hundreds of sources, some stale, some duplicated |
| Meaning | The team agrees what "active customer" means | Regions and departments define it differently |
| Governance | Sandbox access, relaxed rules | [Access controls](https://atlan.com/know/ai-agent-access-control/), approvals and regulation apply to every query |
| Review | A person checks each output | Outputs feed workflows and agents directly |
| Change | Data frozen on demo day | Schemas and business rules keep changing |

Source count alone breaks pilots. In a [2025 survey from a data integration vendor](https://www.fivetran.com/press/fivetran-report-finds-nearly-half-of-enterprise-ai-projects-fail-due-to-poor-data-readiness), 74% of enterprises said they manage, or expect to manage, more than 500 data sources. A model tuned on one clean table has never seen the other 499.

Agents raise the stakes. An [April 2026 analysis of enterprise AI scaling](https://www.ibm.com/think/insights/why-most-enterprise-ai-projects-stall-before-scale) argues that in a pilot, the human reviewer quietly absorbs much of the [gap in context](https://atlan.com/know/closing-the-context-gap/): which table is authoritative, which definition applies, which policy governs use. An agent has no reviewer. A wrong definition stops being a wrong answer on a screen and becomes a wrong action in a [system of record](https://atlan.com/know/ai-agent/data-for-ai/systems-of-record-data-knowledge/).

  Build Your AI Context Stack
  Get the blueprint for implementing context graphs across your enterprise. This guide walks through the four-layer architecture, from metadata foundation to agent orchestration, with practical implementation steps for 2026.
  Get the Stack Guide

---

## What breaks in the data when an AI project stalls?

A stalled project usually has several data gaps at once, and they surface together when real users arrive.

- The data cannot be found. In [BARC's 2026 study of 225 data and AI leaders](https://www.clicdata.com/blog/ai-pilots-stall-in-smbs/), 70% of organizations said less than half of their [unstructured data](https://atlan.com/know/data-for-ai/unstructured-data-for-ai/) is discoverable or usable for AI. Documents, emails and contracts hold much of the context a model needs.
- Nobody agrees what it means. "Revenue" in finance excludes refunds; in sales it might not. A model that reads both columns will pick one, silently.
- Nobody can trace where it came from. The same [BARC study](https://www.clicdata.com/blog/ai-pilots-stall-in-smbs/) found 28% of organizations lack [lineage tracking](https://atlan.com/know/ai-readiness/ai-ready-data-lineage/), so a wrong answer cannot be traced back to the pipeline that caused it.
- Quality is checked after the fact, if at all. Missing values and stale joins reach the model before anyone sees them. See [data quality for AI](https://atlan.com/know/data-for-ai/data-quality-for-ai-agent/) for the checks that matter most.
- Access is not planned. Request [permissions for production data](https://atlan.com/know/ai-agent/how-to-give-ai-agents-access-to-enterprise-data/) at the start of the project, not after the pilot.
- [Freshness](https://atlan.com/know/ai-agent/context-freshness/) lags. In [Confluent's 2026 Data Streaming Report](https://www.ibm.com/think/news/half-ai-projects-stalling-confluent-data-streaming-report), 72% of IT leaders say real-time data infrastructure is a barrier.
- Results cannot be reproduced. Without [versioned training data](https://atlan.com/know/llm-training-data-versioning-strategies/), a team cannot tell whether a model got worse or the data changed, a gap one [engineering guide to AI data bottlenecks](https://lakefs.io/blog/data-bottlenecks-for-ai-projects/) lists among the main bottlenecks.
- Bias controls are thin. [BARC found](https://www.clicdata.com/blog/ai-pilots-stall-in-smbs/) 33% of organizations have inconsistent controls for [data bias](https://atlan.com/know/training-data-bias-detection-methods/), or none.

Each gap alone is fixable. Together, they mean nobody can say why an answer is wrong, so nobody signs off.

---

## Why do teams blame the model first?

Because the model is the easiest part to change. One team can [swap providers](https://atlan.com/know/manage-multiple-llm-providers-scale/) or [rewrite a prompt](https://atlan.com/know/what-is-prompt-engineering/) without waiting for anyone else. Fixing data needs owners in several teams to agree on a definition, and that work produces no demo. [One analysis of BARC's findings](https://www.clicdata.com/blog/ai-pilots-stall-in-smbs/) argues the same: teams blame the model because it is the visible component and the cheapest thing to change.

The model also hides the cause. A language model returns a fluent answer whether its inputs were right or wrong, so a data error reads like a model error. Teams respond by upgrading the model, and the error survives the upgrade, because no model can supply a definition that was never written down.

[Fine-tuning](https://atlan.com/know/fine-tuning-vs-rag/) can make this worse. In a [March 2026 piece on Simple Talk](https://www.red-gate.com/simple-talk/ai/why-most-enterprise-ai-projects-fail-and-how-to-fix-them/), data scientist Animesh Goyal argues that a fine-tuned model captures a static snapshot of business rules and schemas. When either changes, the model is out of date and must be retrained. [Retrieval](https://atlan.com/know/what-is-rag/) keeps the model and the data separate, so a changed definition takes effect without retraining.

In regulated work the cost of mislabelled data is concrete. [Moody's](https://www.moodys.com/web/en/us/creditview/blog/ai-is-here-to-stay-enterprises-must-get-it-right.html) lists fragmented, outdated data and [missing metadata](https://atlan.com/know/ai-readiness/role-of-metadata-management-in-enterprise-ai/) and lineage as causes of errors in credit risk, insurance and KYC screening, where an auditor will ask where every number came from. A new model version does not answer that question. A [lineage graph](https://atlan.com/know/ai-agent/data-for-ai/data-lineage-for-ai/) does.

  Context Gap Calculator
  Estimate how much context your AI use case is missing before the pilot meets production data.
  Calculate Your Context Gap

---

## How do you get a stalled AI project moving again?

Start with the use case, not the whole estate. Gartner warns that organizations which treat AI-ready data as ordinary data management [put their AI efforts at risk](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk). Readiness depends on what the model is asked to do. In the [RAND interviews](https://www.rand.org/pubs/research_reports/RRA2680-1.html), data problems were one of the two causes that more than half of the interviewees named, unprompted, as the main reason projects failed. RAND adds that AI projects need time for teams to explore, understand and [curate data](https://atlan.com/know/ai-agent/data-for-ai/how-to-prepare-enterprise-data-for-ai-agents/) before training, and that [missing domain understanding](https://atlan.com/know/business-context-for-ai/) causes failures. A churn agent and a contract-search assistant need different data, different definitions and different controls.

A practical order of work:

1. Name an owner and a budget for the data work before the model work starts, so the fixes below have someone accountable for them.
2. List the data the use case needs, then mark what is [actually usable](https://atlan.com/know/ai-agent/data-for-ai/what-makes-data-ai-ready/): classified, validated and accessible to the model. The gap between "we have it" and "the model can use it" is usually the stall.
3. [Write the definitions](https://atlan.com/know/ai-agent/how-to-manage-kpi-definitions-and-a-business-glossary-in-a-data-catalog/). Every metric the model will touch needs an owner and a definition in a form a machine can read.
4. Trace lineage for those assets, so a wrong answer leads back to a pipeline and an owner.
5. [Enforce access and policy at the point of use](https://atlan.com/know/data-governance/why-should-data-governance-rules-be-readable-by-ai-agents/), not in a review after launch.
6. [Version and test data](https://atlan.com/know/ai-agent/data-for-ai/data-contracts-for-ai/) the way you version and test code, so you can tell a data change from a model regression.
7. [Measure readiness](https://atlan.com/know/ai-agent/data-for-ai/ai-readiness-assessment/) for that use case, then reuse the work for the next one.

The last step is where most of the payoff sits. The definitions, lineage and policies built for the first use case are the same ones the second needs. Teams that keep them in a shared [context layer between data and models](https://atlan.com/context-layer/) stop rebuilding context for every pilot. Teams that keep them inside each project's prompts start from zero every time.

---

## How does Atlan help AI projects get past the data stall?

Atlan gives models and agents the context that pilots borrow from people: what each asset means, who owns it, where it came from and which policies apply. It holds that context in one [metadata layer](https://atlan.com/know/metadata-layer-for-ai/) across the data stack, and serves it to AI tools through the [Atlan MCP server](https://atlan.com/mcp/), which works with Claude, Cursor, ChatGPT, Gemini and [Microsoft Copilot Studio](https://atlan.com/know/ai-agent/ai-agent-applications/what-are-microsoft-copilot-studio-agents/).

Workday shows the pattern. In [a Re:Govern 2025 session](https://atlan.com/regovern-watch-center/workday-building-semantic-layer/), Joe DosSantos, VP of Enterprise Data and Analytics at Workday, described governance that worked for people: about 1,000 glossary terms and roughly 6 million cataloged assets in one shared vocabulary. AI exposed the gap: machines cannot interpret business logic the way people can, and without clear context, AI tends to invent an answer instead of asking. Workday's response is a machine-readable [semantic layer](https://atlan.com/know/semantic-layer/), co-built in Atlan's AI labs. DosSantos says Atlan captures Workday's shared language so AI can use it through the [MCP server](https://atlan.com/mcp/).

For a deeper look at how a catalog becomes the source of context for AI, see [what an AI data catalog is](https://atlan.com/ai-data-catalog/).

  Atlan in Action: Live Context Layer Demos
  Watch how Atlan turns definitions, lineage, ownership and policy into context that AI models and agents can use.
  Watch the Live Demo

  Book a Demo

---

## FAQs about why AI projects stall on data

### Is bad data the main reason AI projects fail?

It is the most consistent reason. Gartner predicts organizations will abandon 60% of AI projects that lack AI-ready data through 2026, and poor data quality is first in its list of causes for abandoned GenAI pilots. Cost, unclear value and risk controls also appear, but model capability does not lead any major study.

### What is AI-ready data?

AI-ready data is data a model can find, read, trust and use under the right policy for one specific use case. It carries a definition, an owner, lineage, quality checks and access rules. Gartner stresses that its requirements differ from traditional data management, and readiness depends on what the model is asked to do.

### Can a better model fix bad data?

No. A stronger model reasons better over the inputs it gets, but it cannot supply a definition nobody wrote down or a lineage nobody tracked. It will return a fluent answer either way, which makes the data error harder to spot. Fixing the inputs is the only path that survives a model upgrade.

### Why do agentic AI projects stall on data?

Agents act without a person checking each output, so every gap in context becomes an action. Confluent's 2026 survey of 4,625 IT leaders found that nearly half report agentic projects delayed indefinitely or abandoned, and 72% cite real-time data infrastructure as a barrier to AI adoption.

### What should a team check before starting an AI pilot?

Check whether the pilot data matches production data. Ask how many sources the production version will draw from, who defines each metric, whether lineage exists, and how access will be granted. If the answers depend on one person's spreadsheet, the pilot will work and the production rollout will stall.

---

## Sources

1. [Lack of AI-ready data puts AI projects at risk, Gartner (February 26, 2025)](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk)
2. [Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025, Gartner (July 29, 2024)](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025)
3. [The root causes of failure for artificial intelligence projects and how they can succeed, RAND (August 2024)](https://www.rand.org/pubs/research_reports/RRA2680-1.html)
4. [AI project failure rates are on the rise, CIO Dive on S&P Global Market Intelligence (2025)](https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/)
5. [Nearly half of AI projects are stalling due to data problems, new study finds, IBM Think (June 2026)](https://www.ibm.com/think/news/half-ai-projects-stalling-confluent-data-streaming-report)
6. [88% of AI pilots fail to reach production, but that's not all on IT, CIO (IDC research with Lenovo)](https://www.cio.com/article/3850763/88-of-ai-pilots-fail-to-reach-production-but-thats-not-all-on-it.html)
7. [Report on enterprise AI data readiness, data integration vendor press release (2025)](https://www.fivetran.com/press/fivetran-report-finds-nearly-half-of-enterprise-ai-projects-fail-due-to-poor-data-readiness)
8. [Why most enterprise AI projects stall before they scale, IBM Think (April 2026)](https://www.ibm.com/think/insights/why-most-enterprise-ai-projects-stall-before-scale)
9. [AI pilots stall on data, not the model: BARC's 2026 findings, ClicData](https://www.clicdata.com/blog/ai-pilots-stall-in-smbs/)
10. [Data bottlenecks in AI projects: how versioning helps, data versioning vendor blog](https://lakefs.io/blog/data-bottlenecks-for-ai-projects/)
11. [Why most enterprise AI projects fail and how to fix them, Simple Talk (Redgate, March 2026)](https://www.red-gate.com/simple-talk/ai/why-most-enterprise-ai-projects-fail-and-how-to-fix-them/)
12. [AI is here to stay, but enterprises can't afford to get it wrong, Moody's](https://www.moodys.com/web/en/us/creditview/blog/ai-is-here-to-stay-enterprises-must-get-it-right.html)