The current hype surrounding agentic AI often centers on model capabilities. However, this attention carries the risk of obscuring a more fundamental success factor: the scalability of Agentic AI systems is not primarily determined by model sophistication, but by the maturity of the underlying data foundations. This is precisely where Keyrus operates. As a global consultancy specializing in data, analytics, and AI transformation, Keyrus is positioned not at the surface layer of AI models, but at the foundational layer where it truly matters in determining if Agentic AI can succeed at scale.
Indeed, Agentic AI systems are designed to reason, plan, and act across multiple domains. Unlike traditional predictive models, they do not operate on a single curated dataset. Instead, they rely on multiple structured and unstructured sources of information across multiple operational domains such as ERP’s, CRM platforms, operational databases, data warehouses, data lakes, knowledge repositories, policy documents, real-time signals etc.
As a matter of fact, the challenge, therefore, is not merely connectivity. It is coherence. And Agentic AI amplifies this complexity.
In practice, and although few statistics are available with regards to the number of sources being used by different types of AI agents, the following table gives a high-level overview from both a conservative and a realistic point of view:
Type of agent | Number of sources (Conservative) | Number of sources (Realistic) | Example sources |
|---|---|---|---|
Customer service agents | 2-3 | 5-8 | CRM, knowledge base, contracts, orders, chat history, ERP etc. |
Operational agents and copilots | 4-6 | 10-15 | Emails, calendar, corporate communication tools (ex. Teams), tasks managers, documents etc. |
Research agents | 8+ | 50+ | Company data, knowledge bases, and external data such as RSS feeds, news, social media publications, scientific publications, patents, market analyses etc. |
Data & analytics agents | 3-5 | 3-8 | Data warehouses, lakehouses, SQL databases, flat files, API’s, usually massive. |
The maturity of the underlying data foundations is what is driving the success of Agentic AI at scale
Most industry benchmarks consistently highlight the magnitude of the issue. Estimates from Gartner, IBM, and McKinsey suggest that 60 to 80% of AI project time is devoted to data preparation and integration. At the same time, IDC estimates that 80 to 90% of enterprise data is unstructured, and modern enterprises frequently operate across dozens or even hundreds of SaaS applications.
Technical connectivity without semantic consistency produces fragile autonomy
The structural realities we discussed create fragmentation long before any AI agent is deployed. And this naturally leads to the semantic dimension of the problem.
When core business entities such as “customer,” “product,” “contract,” or “risk exposure” are defined inconsistently across systems, technical integration alone cannot guarantee reliable reasoning. An AI agent may retrieve data successfully, yet misinterpret its meaning. Semantic inconsistency introduces ambiguity into decision-making processes, thereby undermining trust in autonomous outputs, preventing agentic solutions to deliver value.
Technical integration solves access. It does not solve meaning. AI agents retrieving such data can technically connect to every system and still misinterpret relationships. This is not a modelling issue. It is a semantic issue. Autonomous reasoning requires a shared understanding of what entities represent and how they relate to one another.
Consequently, semantic alignment becomes a prerequisite for scalable autonomy. Without it, organizations accumulate what may be described as semantic debt, adding up on top of a too frequent data debt, which is definitely a hidden liability that constrains the reliability of cross-domain reasoning or true agentic solutions.
Structuring intelligence: ontologies enabling contextual AI reasoning
Ontology engineering and knowledge graphs allow AI systems to traverse relationships dynamically, uncover contexts and reason across domains. In essence, an ontology defines the meaning and rules, and a knowledge graph instantiates those definitions with real data and connections.
The lack of very solid practices in those domains explain why many agentic AI initiatives still remain today confined to proof-of-concept environments. Sandbox demonstrations often rely on curated, harmonized datasets. However, once deployed into the complexity of enterprise ecosystems, agents encounter fragmented architectures, unclear semantics and data ownership, incomplete lineage, and inconsistent definitions. The result is not model failure, but systemic friction.
An ontology is a formal description of the concepts in a domain and how they relate to each other. In simple terms: It defines what things exist (customer, product, order, contract etc.). It defines what they mean. It defines how they relate (a customer places an order; an order contains products being shipped; a product belongs to a category). Think of it as a structured business vocabulary that machines can understand. Unlike a glossary, which is descriptive, an ontology is computational. It enables systems to reason over relationships. Without ontology engineering, AI agents operate in an environment where definitions drift and relationships are ambiguous. |
A knowledge graph is a way of organizing data based on relationships rather than tables. While traditional databases store information in rows and columns, knowledge graphs store information as connected entities. This structure allows AI systems to traverse relationships dynamically, uncover context, and reason across domains. |
From data availability to data coherence: a Keyrus differentiator
So, as we see, many organizations can technically connect systems. But Agentic AI introduces a higher bar: coherence across domains.
This is where Keyrus’ capabilities become critical, helping organisation define and standardise their data products to reach consistent semantics and scale their AI successfully. Read more on our services.
The path to scalable Agentic AI is inseparable from data maturity transformation
In conclusion, in order to make Agentic AI scalable, and truly useful, the technical capabilities must be matched by:
Strong data governance
Clear domain accountability
Interoperable architectures
Explicit semantic frameworks
High-quality master data foundations
Agentic AI does not merely require access to data; it requires shared understanding of data. Autonomy depends on contextual reasoning. Context depends on semantics. And semantics depend on disciplined data architecture.
Behind the visible progress in AI models lies a less visible imperative: organizations that wish to move beyond experimentation must first resolve their data and semantic fragmentation. Only then can Agentic AI transition from impressive demonstration to reliable enterprise capability.
In other words: Agentic AI failure is often a data transformation failure. And data transformation is Keyrus’ core business. We’ve been doing this for 30 years.