Artificial intelligence can accelerate analysis of large datasets, reduce manual errors and identify trends. In practice, results depend heavily on the quality of a company’s data and infrastructure. Salesforce research shows fewer than half of organisations consider their infrastructure, data and security fully ready for AI. Meanwhile, 84% of analytics leaders believe AI development requires a fundamental overhaul of data-management strategy.
This matters because businesses increasingly use AI for analysis supporting financial and operational decisions, not merely simple tasks. Incomplete, inconsistent or outdated inputs can cause even an advanced model to produce a credible-looking answer that leads to the wrong conclusion.
Salesforce found that 89% of organisations had experienced inaccurate or incorrect AI results. Separately, OneStream research shows 95% of surveyed chief financial officers see risks in using AI for data analysis.
The problem is not always AI itself
One cause of unsuccessful projects is treating AI as a standalone tool that can simply be connected to existing processes. AI systems depend on IT infrastructure, data sources, integration mechanisms and access-management rules.
“Some organisations implement fairly randomly chosen AI tools and models in the hope that they will improve existing processes. Achieving the expected results, however, requires building complete systems,” said Sebastian Jaworski, vice-president of Linux Polska.
He stresses integration with existing IT infrastructure and gradual development of the whole environment. How employees use the tools also matters: some risks arise from inappropriate data use and missing procedures rather than the technology itself.
Bad data can cost millions
Data-quality problems predate generative AI. But AI can amplify them by processing more information and generating recommendations faster.
OneStream found that 47% of CFOs had made a wrong decision in the previous year based on inaccurate or incomplete financial data. In 72% of those cases, reported losses were at least USD 500,000.
The issue is therefore broader than whether a model’s answer is technically correct. If its source information is wrong, automation can also accelerate bad decisions.
Information is scattered across departments
Many businesses keep information in different databases, applications and business systems. Sales may use a CRM, finance an ERP system, logistics its own platform and customer service another environment.
Teams may follow different collection and updating rules, leaving several versions of the same information and some data outside the main systems. According to Salesforce, around 90% of organisational information may be unstructured, including text documents, messages, files, presentations and other material without a uniform schema. Seven in ten analytics leaders believe such data may contain particularly valuable business insights.
AI requires a well-organised data foundation
Before AI can use information, it must be collected, cleaned, integrated and properly described. Traditional database and data-engineering technologies therefore matter alongside the models.
A foundational element is ETL—extract, transform, load—which retrieves information from different sources, transforms it and stores it in a target system.
“Only well-ordered and organised information can be used effectively in analysis with AI-based tools,” said Maciej Wawrzyniak, director of databases and data engineering at Linux Polska.
A company seeking advanced AI analysis may first need to reorganise its data architecture. Deploying a model does not automatically fix incomplete records, incompatible formats or inconsistencies between systems.
Fewer than half report full readiness
Salesforce’s survey illustrates the scale of the gaps. Only 49% of organisations feel fully prepared for AI in terms of infrastructure. Data quality and data security each score 47%, system integration 46%, and data freshness 45%.
Most companies therefore identify at least one area of their data infrastructure that could hinder AI use.
Outdated information increases the risk of wrong answers
Freshness is particularly important. A model analysing information several months old may execute the technical task correctly but draw conclusions that no longer match the business’s circumstances. Systems need access to appropriate sources and clear updating rules.
“Outdated or incomplete data will always result in incorrect analysis. Organisations collect increasing amounts of information and do not always control its quality or security,” said Kacper Szczepaniuk, Observability Specialist at Linux Polska.
More advanced deployments also require infrastructure monitoring to detect interrupted data flows, integration failures or an analytics system drawing on an old source.
Multiple systems remain the norm
Only 19% of companies use inputs from a single system for AI analysis, according to OneStream. For 32%, dispersed databases and a lack of integration are among the main obstacles to achieving expected AI outcomes.
The problem could grow as more AI tools are deployed. If each department adopts its own models, sources and processing methods, the company may create new data silos instead of a shared analytics environment.
More than half develop models using inadequate data
Despite infrastructure gaps, companies continue deploying AI. Salesforce reports that 55% train or develop their own models using data that does not meet required quality standards.
This is particularly significant for forecasting, financial analysis and automated decisions. Models trained on incomplete or inconsistent data can perpetuate its errors.
Generative AI also presents the additional problem of hallucinations: information the model cannot substantiate from available data. Access to appropriate, current sources can reduce the risk, but does not by itself guarantee correct answers.
Data first, AI second
AI’s popularity is changing investment priorities. Companies want to deploy quickly even when they still lack a coherent information-management system.
Yet analytical quality depends largely on what happens before data reaches the model: collection, cleaning, description, integration and updating. AI can accelerate analysis, but cannot replace the infrastructure responsible for data quality. Faster processing of faulty or incomplete information does not necessarily improve decisions; it may simply bring a business to the wrong conclusion sooner.
Sources: Salesforce, State of Data and Analytics – 2nd Edition, 2025; OneStream, Companies Are Scaling AI on Data They Don’t Trust, New Study Finds, 2026. English translation of the ManagerPlus article.
