Models from Anthropic, OpenAI and Mistral get better every quarter, yet most AI projects still stall in the same place: the data. In Iron Mountain's 2026 research, 86% of organisations said data problems delayed or derailed AI projects in the past year. A US survey by Further found that only 5% of AI decision-makers believe their data is largely ready for AI. This guide shows what the latest surveys say, what "AI-ready data" actually means for Claude, GPT and Mistral projects, and how a mid-sized company can get its data ready one use case at a time.
The bottleneck moved from models to data
In 2023 the question was whether the models were good enough. In 2026 Claude, GPT and Mistral models handle long documents, call tools through MCP and write production code. What they cannot do is fix a customer master with three versions of every client, a SharePoint full of outdated contracts or an ERP that nobody can query. An AI assistant that answers from the wrong policy version is worse than no assistant at all.
How often data blocks AI: four surveys, one pattern
Different studies measure the problem differently, but they point in the same direction. Iron Mountain's report "The Information Supercycle" puts the share of organisations hit by data-related delays at 86%. Fivetran's survey of 401 data leaders across the US, Europe and Asia-Pacific found that 42% saw more than half of their AI projects delayed, underperform or fail because of data readiness. In Further's 2026 State of Enterprise AI Maturity survey, 37% paused, scaled back or abandoned an underperforming AI initiative in the past 12 months. Gartner predicted in 2025 that through 2026 organisations will abandon 60% of AI projects unsupported by AI-ready data.
Sources: Iron Mountain, "The Information Supercycle" (2026); Gartner (February 2025 prediction); Fivetran AI & Data Readiness Survey (401 data leaders, 2025); Further, State of Enterprise AI Maturity (515 US AI decision-makers, 2026). Different samples and questions.
Ambition is high, readiness is low
The gap is not about intent. In Iron Mountain's data, 95% of respondents say using data to improve business outcomes is a priority, but only 39% have data that is digitised, structured and readily usable for analytics and AI. In Further's survey only 29% have completed a formal data audit with a remediation plan, and just 5% think their data is largely AI-ready. Interestingly, 77% of the same Further respondents still report a positive return on AI: the early, simple use cases pay off, but scaling them runs into the data wall.
Sources: Iron Mountain (2026) for "data is a priority" and "data usable for AI"; Further, State of Enterprise AI Maturity (2026) for "formal data audit" and "data largely AI-ready". Different samples.
Where companies get stuck
The obstacles are remarkably consistent. 70% of organisations in Iron Mountain's research cannot effectively track the lifecycle of data used in their AI models, 56% name data spread across disconnected systems or departmental silos and 53% name legacy formats or physical archives. In Fivetran's survey, regulatory compliance is the top data challenge for AI (59%) and 41% say a lack of real-time data access stops their models from delivering timely insights.
Sources: Iron Mountain (2026): lineage, silos, legacy formats; Fivetran AI & Data Readiness Survey (2025): compliance, real-time access.
The cost shows up in time. Four in ten organisations spend at least 41% of their AI project timelines on manual data preparation (Iron Mountain), and Fivetran found that 67% of centralised enterprises spend more than 80% of their data engineering resources just maintaining pipelines. Every week spent copying spreadsheets is a week not spent on the use case.
What "AI-ready data" means in practice
AI-ready is not a property of a database; it is a property of data for a specific use case. A customer-service assistant and a financial forecasting agent need very different things. We assess readiness on six dimensions:
| Dimension | Question to ask | Typical fix |
|---|---|---|
| Access & integration | Can the AI reach the source system without manual exports? | APIs, MCP connectors, read replicas |
| Quality & freshness | Is the data accurate and current enough for this task? | Deduplication, validation rules, sync schedule |
| Metadata & ownership | Does every source have an owner and a description? | Data catalogue, named data owners |
| Lineage & governance | Can we trace an answer back to its source? | Citations in RAG, logging, retention rules |
| Security & permissions | Does the AI respect the same access rights as the source? | Permission-aware retrieval, SSO, DLP |
| Unstructured content | Are documents current, deduplicated and machine-readable? | Archive cleanup, OCR, chunking, versioning |
Assessment by AI Workshop / Blocksize based on project experience in Croatia and the DACH region, November 2026. Higher is better on every axis.
Why this matters for Claude, GPT and Mistral projects
- Retrieval quality beats model choice. In a RAG system, switching from one frontier model to another rarely fixes wrong answers; removing duplicate and outdated documents usually does.
- Connectors expose permissions. MCP connectors and enterprise connectors in Claude Enterprise or ChatGPT Enterprise make data reachable in minutes. If SharePoint permissions are too broad, the assistant will faithfully surface what employees were never meant to see.
- Long context is not a data strategy. Large context windows let a model read more, but they do not decide which version of a document is valid.
- Structured data needs a semantic layer. Text-to-SQL agents work far better when tables and metrics are documented.
- Compliance needs traceability. The EU AI Act requires data governance for high-risk systems and the GDPR applies to every prompt containing personal data; lineage and logging are the evidence (see our GDPR guide).
A use-case-first readiness plan
Do not wait for a company-wide data platform. The pattern that works in mid-sized companies:
- Weeks 1–2: pick one use case and map its data. List every source the use case needs, who owns it and how it is accessed today.
- Weeks 3–4: score it on the six dimensions. Fix the blockers that matter for this use case only: remove outdated documents, tighten permissions, document the key tables.
- Weeks 5–8: build with citations and logging. Every answer links back to its source, so users and auditors can check it, and feedback shows which data is still wrong.
- Then repeat. Each use case leaves behind cleaned sources, connectors and owners that the next one reuses. Measure the result with a baseline (see measuring AI ROI).
The bottom line
The surveys agree: most AI projects do not fail because Claude, GPT or Mistral are not good enough, but because the data behind them is scattered, outdated, unowned or over-shared. Only a small minority of companies call their data AI-ready. The fastest route is not a multi-year data programme but a use-case-first approach that makes exactly the right data accessible, clean, permissioned and traceable, and then reuses that work for the next use case.
Is your data ready for AI?
AI Workshop runs a focused AI data readiness assessment: we score your priority use case on the six dimensions, fix the blockers and build the first Claude, GPT or Mistral solution with citations, permissions and logging in place.
Book a data readiness check