Imagine the VP of finance at a large retailer. He asks the company’s new AI analytics agent a simple question: “How much was our income last quarter?” The answer comes back in seconds.
Confident.
Clean.
Wrong.
💜 of EU technology
The latest rumblings from the EU tech scene, a story from our wise founder Boris and some questionable AI art. Free in your inbox every week. Register now!
This exact scenario happens more often than many organizations would like to admit. AtScale, which enables organizations to implement managed analytics environments and semantic consistency, found that simply increasing model parameterization alone could not solve the problem. AI management and contextual issues facing businesses.
When AI systems query inconsistent or unmanageable data, adding more model complexity doesn’t contain the problem, it compounds it. Organizations across industries have moved quickly to advance agent AI, deploying systems that analyze data, generate insights and run automated workflows. In response to this trend, AI models have adapted to respond quickly through larger model parameters, increased computing power, and additional features. The underlying assumption is that if the model is large enough, the final result will be reliable.
However, there are indications that this assumption may not hold true. End TDWI study It found that nearly half of respondents characterized their AI management initiatives as either immature or very immature. This may have more to do with the data lineage and the business definitions on which these models are based than with the capabilities of the models.
Why don’t the larger models solve the steering?
The AI industry tends to operate on an unexamined assumption about what leads to better performance: that as we build more advanced models, they will somehow correct their performance errors on their own. In enterprise analytics, this assumption can quickly unravel.
While scale can improve the breadth of reasoning in a model, it does not automatically enforce which definition of gross margin a business is willing to use. It doesn’t solve the metric inconsistencies that have been living in separate dashboards for years. Nor does it produce a traceable lineage by itself.
Management problems are not solved at scale. Business rules hidden in individual tools, inconsistent definitions across teams, and results with no audit trail are structural problems that a larger model structure cannot fix. It just makes invalid answers more fluent.
At AtScale, there’s a consistent theme among our customers: Problems don’t end there when inconsistent data definitions drive organizations into the AI layer. They usually spread forward at a higher rate and with less transparency than the previous layer offered.
Performance and responsibility are separate business. A model causes. The control layer defines what the model does, constrains how it applies business logic, and ensures that results are returned to the record source. One cannot replace the other.
Real risk: Unrestricted agents in an enterprise environment
The problem with AI agents is rarely the model itself. This is what the model does and if anyone can see what it is doing.
In a general context, AI agents can read data differently in different systems. In large enterprises, even small differences in definitions can lead to different results. Structural risks usually arise from four main causes:
- Agents draw from sources where the same metric can mean different things to different teams, making data definitions less clear.
- Indicators of different departments disagreeing – two agents give two answers, but it is not clear which is correct.
- Implicit reasoning produces results without a clear lineage of how the decision was made.
- Audit gaps: When results cannot be traced back to a controlled record source, there is no reliable way to catch errors, assign accountability, or correct course.
These are not signs that the AI is not working. They show that the infrastructure around artificial intelligence is not keeping up.
What guardrails actually mean in AI analytics
Railings are often seen as limitations. However, in many cases, the guardrails are conditions that allow AI agents to operate with greater confidence.
Blocks can help align AI-generated results with defined business logic. They also create a structure within which autonomous agents can operate; thus, as autonomy increases, so does reliability. In analytics, protective bars usually available in several specific formats:
- Shared data definitions: A single definition of terms such as profit, loss or margin shared across all systems.
- Business logic constraints: Rules governing how computations are to be performed, independent of the instruments or agents performing those computations.
- Generational vision: Ability to identify where any output is coming from.
- Access controls: Permissions that define what data the agent can query.
- Standardization of metrics: Consistent definitions applied across departments and platforms.
The goal is not to prevent AI from functioning. This is to offer the AI a base it can stand on.
The role of the semantic layer as a constraint framework
The semantic layer sits between data and applications and the AI agents that use it, defining business concepts, implementing logical processes, and providing a common framework of terms for all applications and AI agents to use.
A semantic layer does not manipulate or duplicate data; defines what the data represents. Instead of generating inferences by asking questions about a more controlled semantic layer than the underlying table, AI agents can generate inferences based on business-defined logic. This difference in output becomes especially important when multiple AI agents on multiple systems need to produce similar results.
From AtScale’s perspective, the semantic layer serves as a context boundary that can help AI agents interpret data according to shared business definitions. The semantic layer is more like a common language than a buffer that ensures all systems work with a common understanding.
Governance is an architectural question, not a pattern question
Business organizations understand this AI management it is less about building the largest model and more about creating an environment in which the selected model can work well. A well-designed and managed architecture (with common definitions for concepts, traceable logic, and shared context across all systems) will produce better, more reliable results than a larger model operating in an unmanaged data environment.
Scaling models without improving semantic clarity tends to add rather than reduce complexity. As each additional tool, system, or workflow is added to an uncontrolled environment, the opportunity for variation increases.
In this sense, responsible AI is an infrastructure problem. Organizations with successful AI deployments consider the meaning of their data as a design decision before a model is chosen.
Economic and operational results
Management spaces do not remain abstract for long. They tend to appear on the budget.
Uncertainty in the sense of data can increase operational friction, with agents producing inconsistent results requiring human review, reconciliation cycles, and rework linked across teams and tools. Audits cost more when generation is unclear. Improving controls after deployment is usually more expensive than building the right architecture from the start.
In complex enterprise settings, costs can appear in predictable ways: unnecessary validation when results don’t match across systems, overcalculation caused by ambiguous queries, and slower analysis as teams pause to figure out which answer is actually valid. Clear semantic constraints can mean fewer validation cycles, and this operational cost is easier to measure.
The way forward: Limited autonomy
AI agents are not in the future, they are already in use. It’s still the infrastructure around them. Without clear context and constraints, agents tend to act outside of what the organization can actually control. This gap does not close on its own.
AtScale argues that the distinguishing model in enterprise AI will not be scale, it will be clarity of models of the environment in which it operates. As agents become more pervasive in workflows, how well the semantic layer is defined may become more important than how large the model is.
This shift towards a controlled context and limited autonomy is explored in more detail AtScale’s State of the Semantic Layer to 2026 reporthow open standards are researched, interactionand semantic management are shaping the next phase of enterprise intelligence.
