According to the Cisco AI Readiness Index 2025, only 13% of organizations are "Pacesetters" ready to derive real value from AI due to mature data infrastructure. Most large enterprises find that pilot projects for Large Language Models (LLM) or analytical AI fail against the harsh reality: corporate data is siloed, lacks clear ownership, and lacks unified quality standards. Attempting to provide models access to unorganized repositories, for example via RAG (Retrieval-Augmented Generation), leads to unpredictable results and creates serious business risks.
Why AI initiatives stall at the pilot stage: the poor data quality trap
When system architects deploy AI solutions, they often focus on model selection or computing power. However, the main bottleneck is input data quality. If a decision-making system uses outdated, incomplete, or contradictory data from ERP, CRM, and legacy document management systems, the result will be predictable—the classic "garbage in, garbage out" problem.
In enterprise infrastructure, where point-to-point integration prevails, data often exists without context. The same concept may be calculated using different algorithms across departments. Without a clear understanding of the information lifecycle from creation to the analytical lake, the AI model becomes unmanageable. Making decisions based on such data in critical infrastructure or the financial sector is unacceptable.
Data Governance and Data Mesh: shifting responsibility to business domains
To solve the chaos, large organizations are increasingly moving away from monolithic storage. Instead, they are adopting the Data Mesh concept. As experts in this architectural approach (such as Zhamak Dehghani) note, this is not just a technical solution but primarily an organizational model designed for large organizations with many domains. It is based on four principles:
- Domain ownership — transferring responsibility for data directly to business domains.
- Data as a product — each domain has an owner responsible for data quality and relevance.
- Self-service data infrastructure — providing domains with tools for independent data processing.
- Federated computational governance — balancing domain autonomy with the need to adhere to global corporate security and quality standards.
Data Mesh is not a universal "silver bullet" for every business, but for sprawling corporations, it is an effective way to ensure that only verified data enters AI models.
Data lineage as an audit tool: why AI must "know" the origin of every figure
For an AI model to work reliably, a complete map of data movement—data lineage—is required. This is the automated tracking of the entire path: from the primary source (transactions in ERP or records in an electronic document management system), through transformation stages, to the end consumer.
To build such an audit, it is critical to use modern integration patterns. For example, using platforms based on Apache Kafka allows for event storage with replay capabilities. This is necessary for system state reconstruction: if a model produces an erroneous result, engineers can trace the event stream back to identify where the anomaly occurred.
Managing AI risks using the NIST AI RMF 1.0 framework: practical application
To systematically reduce risks, organizations are adopting the voluntary NIST AI RMF 1.0 framework. It structures risk management around four key functions:
- Govern — creating a culture of responsible AI use and defining policies.
- Map — identifying the context of AI use, classifying data sources, and assessing risks.
- Measure — evaluating model reliability and security using metrics.
- Manage — systematic risk response.
For critical infrastructure systems, NIST emphasizes the need to evaluate the context of use, potential harm, reliability, and accountability, rather than just the mathematical accuracy of the model. Data lineage plays a decisive role in the Map and Measure stages: it is impossible to measure model reliability without precise knowledge of data origin.
Architectural basis: how system integration prepares infrastructure for AI
Building a managed data architecture requires moving away from chaotic integrations toward event-driven architecture and a unified semantic metadata layer.
The technological foundation for such tasks can be the low-code platform UnityBase, a joint development of the Intecracy Group technology alliance (where InBase acts as a key, but not the only, developer). Using a unified domain metadata model, the platform combines data, API, and behavior descriptions, allowing the system to evolve as a model-driven solution. To prepare infrastructure for AI projects, UnityBase provides the following mechanisms:
- Built-in audit trail and DataHistory for tracking changes at the platform level, which is the foundation of data lineage.
- Generated REST API for standardized integration with AI services and data buses.
- Flexible access control (RBAC, RLS, and ACL in commercial editions). For high-load projects or increased security requirements, the official platform documentation recommends using Enterprise or Defence editions, which guarantee data protection from unauthorized inclusion in AI training sets.
Large-scale enterprise solutions are built on UnityBase, such as the Megapolis.DocNet electronic document management system, which ensures high performance and access control. Additionally, other Intecracy Group expertise strengthens company readiness for AI implementation. For example, Softengi provides AI consulting and development services and is certified to the ISO/IEC 42001:2023 standard (AI management). The Nectain platform (ECM/DMS class) offers specialized modules for evaluating AI algorithm quality, comparing error rates and accuracy against human-verified benchmark data.
System integration based on modern tools allows for the creation of a reliable Data Governance layer, where access rights are controlled and information origin remains transparent to AI models.
Data infrastructure maturity levels for AI project readiness
| Maturity Level | Infrastructure Characteristics | AI Project Readiness |
|---|---|---|
| Level 1: Ad-hoc | Data is siloed, point-to-point integration, no data lineage, data quality is not business-controlled. | AI projects are impossible or create critical risks due to hallucinations. |
| Level 2: Consolidated | A unified storage (DWH/Lake) is created, but there are no clear owners, quality is checked sporadically. | Training simple models is possible, but there is a high risk of using outdated data. |
| Level 3: Governed | Data Governance is implemented, owners are defined (data as a product), basic data lineage exists. | Ready for local AI initiatives and RAG systems within departments. |
| Level 4: AI-Ready (Federated) | Automated data lineage, federated computational governance, real-time quality control. | Fully ready for scaling critical AI systems with audit guarantees. |
FAQ
How can one build reliable data lineage if a company uses more than 8 different enterprise systems?
In a heterogeneous environment, it is necessary to move away from direct point-to-point integrations in favor of event-driven architecture (e.g., using Apache Kafka). Each system publishes events to a unified registry, which allows for tracking changes. Using platforms with a built-in audit trail mechanism (like UnityBase) significantly simplifies automatic lineage generation and mapping.
Is it mandatory to implement Data Mesh to prepare data for AI, or is a classic DWH sufficient?
Data Mesh is primarily an organizational approach for large companies with many complex business domains. If an organization has a centralized structure, a classic data warehouse (DWH) with well-configured Data Governance processes will be sufficient. The key is quality control, having designated data owners, and transparent lineage, rather than the choice of a specific architectural pattern.
How does the NIST AI RMF 1.0 framework help reduce the risks of hallucinations and errors in corporate AI models?
This voluntary framework offers a systematic approach (Govern, Map, Measure, Manage). At the Map stage, data sources and usage context are analyzed, and at the Measure stage, reliability assessment metrics are implemented. This forces companies to build strict filters for input data and verify the accuracy of AI responses against benchmark data, which minimizes the risk of hallucinations.