The data behind autonomous agents

The first wave of generative AI sped up tasks people already did: drafting, summarizing, boilerplate code. Useful, but bounded — a copilot bolted onto a broken workflow captures a modest gain and no more. Agentic AI is different in kind. Agents plan, retrieve, act, and run multi-step processes with little human intervention. The value is much larger. So is the demand on the data underneath them.

Why the data breaks first

Point autonomous agents at the standard enterprise — siloed databases, conflicting legacy systems, unstructured documents — and they fail badly. They hallucinate, take wrong actions, or stall on contradictory inputs. The limit is no longer the model's reasoning. It is the gap between the model's general training and your specific, messy data.

This is why agentic ambition and a weak data foundation cannot coexist. A copilot can paper over some of the mess because a human stays in the loop to catch it. An agent acting on its own has no such buffer.

The semantic layer is the fix

The answer is a semantic layer — a corporate ontology that turns scattered raw data into a shared, machine-readable language of business entities. Instead of tables and columns, it defines what a customer, an invoice, a supply-chain event, or a production delay is across the whole organization, and it carries the permitted actions and security rules alongside the data. That context is what lets an agent reason correctly and act safely.

Solve a data bottleneck once in the semantic layer, and later applications reuse it instead of paying the integration cost again. It is the architectural dividend, expressed in architecture.

Underneath it sits discipline across the stack: ingestion as a single standardized product, a platform that unifies access with vector stores so unstructured content is searchable by meaning, data packaged as reusable products with clear ownership, and continuous governance — real-time quality monitoring rather than periodic cleanup, with every agent action logged for audit.

Buy the plumbing, build the meaning

There is a tension worth resolving out loud, because a sharp reader will notice it. Good strategy preaches strict build-buy-partner discipline — protect scarce engineering, never build commodity software you can buy. Yet a semantic layer and an agent architecture are heavily custom. Both are correct, and the line between them is the whole point.

Buy or configure the commodity layers: cloud infrastructure, vector stores, orchestration, foundation models, monitoring. These are solved markets, and building them from scratch burns the capacity you cannot spare. Build only the two things that are genuinely yours — the enterprise ontology, because your definition of a customer or a claim is specific to you and is the source of the dividend, and the differentiating workflows that encode how your business competes. Partner for surge capacity and specialized skills you don't need permanently.

Keep autonomy from becoming risk

At scale, the unit of delivery becomes a small agentic team — two to five people who supervise the AI workflows for an end-to-end outcome, rather than a department of manual processors. The humans direct and handle exceptions; the agents execute.

To keep that from turning into exposure, borrow from software engineering. Just as security checks moved into delivery pipelines, embed control agents into the workflow: agents that challenge the primary agents' outputs for accuracy, enforce policy and brand rules, and check data handling against privacy limits. Every action is logged and explained. This is where the technical architecture meets governance — the controls are not a document filed somewhere, they run inside the system.

None of this is day-one work. Build the minimum viable platform for your first use cases and let demand pull the foundation forward. The semantic layer and the agentic architecture are the horizon a maturing office grows toward — worth understanding now so the board can see where the compounding value comes from, and worth resisting as an excuse to build a grand platform before shipping anything.