The agent isn't the problem

Most enterprise data platforms weren't designed for AI agents. They were built for storage, query performance and pipeline reliability. Business meaning was added later through documentation, tribal knowledge and data dictionaries that are rarely complete or current.

When agents operate in these environments, they inherit the same structural confusion that has long frustrated human analysts.

Consider a straightforward business question: "Which of our subscribers is at highest risk for churn this week?" The agent must reconcile data scattered across legacy and acquisition-era systems, inconsistent identifiers and relationships that were never formally encoded. The result is often answers that look plausible but are semantically wrong -- or, in some cases, there's no answer at all.

The core issue isn't tooling -- it's orientation. Traditional data systems were built around technical logic rather than business logic. What enterprises need is an architecture that is built around business semantics that organizes data according to how the business operates.

Business semantics as first-class architecture

The central idea is straightforward: Anchor all data assets to master data. These are the canonical definitions of your core business entities, such as your customers, products, contracts and routes.

New data sets are automatically linked to these entities through an ID graph that resolves real-world identity challenges. A single customer might appear as an anonymous browser ID in analytics, an account number in billing and a different key in a CRM. The ID graph creates a unified reference point that every downstream consumer -- human or agent — can rely on.

This approach is far more than data hygiene. It's the prerequisite for reliable agentic AI. A knowledge graph built on this master-data foundation supplies the semantic grounding agents need. When an agent generates SQL queries or retrieves data, it understands what things mean, not just where they're stored. As new data sets link to business entities, the agent's effective knowledge base expands naturally, reducing reliance on manual prompt engineering or one-off context injection.

This diagram shows how an architecture built around business semantics works:

Making outputs trustworthy

As organizations move from experimentation to production automation, trustworthiness is critical. A business semantics-centric system addresses this issue with a clear pattern:

• A generation agent translates natural language requests into executable outputs, such as SQL queries, API calls or workflows, along with a step-by-step reasoning trace.
• These outputs route to multiple verification agents that independently assess technical correctness and semantic alignment with the knowledge graph, each producing a confidence score.
• When the collective confidence exceeds a defined threshold, the result proceeds. Below the threshold, the system escalates the output to a human expert. The knowledge graph helps route the question to the person with relevant domain expertise.

All outputs, confidence scores, function calls and human feedback are logged in a persistent outcome database. This database delivers full traceability, supports anomaly detection and meets audit requirements in regulated industries. To control cost and latency, the architecture favors smaller, specialized models for well-defined tasks, such as SQL generation, and uses parallel evaluation where possible.

Workflow insights most teams miss

Many data teams apply uniform standards to exploratory analysis and production pipelines, resulting in work that's either too slow for exploration or insufficiently rigorous for production. A business semantics-centric approach deliberately separates the analysis and pipelines:

• Exploration workflow prioritizes speed. Data is lightly processed, linked to business entities via the ID graph, minimally cleaned and labeled as exploration-only. Analysts and data scientists can start working immediately.
• Production workflow applies full quality assurance, standardized metrics and BI-compatible transformations once the value is proven.

The powerful advantage is domain knowledge and SQL logic generated during exploration transfer directly into production. As a result, timelines are compressed: Teams that once needed six or more weeks to acquire and prepare data can often reduce that time frame to about two weeks.

The human factor: Fewer people, higher stakes

Business semantics doesn't eliminate people, rather it elevates their role. Agents handle routine tasks, such as query writing, data classification and basic pipeline logic, shifting the human effort to deep semantic judgment.

For business leaders, the key question is … whether their data systems speak the same language as the business.

The result is a smaller but more impactful team. Instead of large groups organized around tools and tasks, organizations have domain stewards aligned to core business entities in the same knowledge graph that agents use. When an agent's confidence score drops, escalation routes to the right expert -- someone who can say, "This join is technically correct but semantically wrong because of how we define a subscriber post-acquisition."

These expert roles demand a deeper understanding of the business and carry higher stakes. They're intellectually demanding and harder to fill in the talent market. But they're exactly what makes agentic systems safe and valuable at scale.

What this means for enterprise AI strategy

Agentic AI is advancing rapidly toward systems that plan, execute and adapt across multistep workflows. Yet capability is no longer the primary bottleneck, business meaning is.

Businesses extracting the greatest value have data infrastructures semantically coherent enough for agents to reason correctly. A business semantics-centric foundation delivers that coherence because it has master data as the anchor, ID graphs for identity resolution, knowledge graphs for operational context and domain experts as the final validators.

For business leaders, the key question isn't whether models are ready. It's whether their data systems speak the same language as the business and whether their agents will truly understand what their answers mean.

Rashmi Choudhary is a data scientist specializing in large-scale AI systems for routing, navigation and operational intelligence. An IEEE senior member and inventor on multiple patents in transportation AI, she focuses on building reliable and accountable AI systems for safety-critical environments. She writes about AI governance, infrastructure intelligence and production system reliability.