The honeymoon is over

There was a time when progress in AI felt straightforward: throw more compute at larger models and watch capability climb. That pattern produced astonishing leaps in language, vision, and multimodal reasoning. Now a new class of systems—agentic AIs that perceive, plan, and execute across systems—is promising to transform how companies operate. They will draft contracts, route invoices, manage inventories, triage customer issues, and orchestrate workflows across dozens of services.

But a hard truth is settling in: the world these agents must operate in is messy, inconsistent, and siloed. Talent and capital flow to models and GPUs because they are visible and exciting. The less glamorous work—taming data—gets deferred. Yet in production, the model is only as useful as the data it can ground itself in. The models may be omniscient in isolation, but when they meet enterprise reality they’re often starved for coherent, timely, and trustworthy data.

What agentic AI needs from data

Agentic systems are not static prediction engines. They are continuous control systems with four interdependent needs from data:

1. Contextual grounding. Agents must reliably link abstractions in language to the concrete entities in a company: customer accounts, procurement contracts, product SKUs, SLAs, regulatory constraints.
2. Freshness and timeliness. An agent’s decision about whether to fulfil an order or open a support ticket can hinge on data that changed minutes ago. Latency kills trust.
3. Lineage and provenance. For an agent to act autonomously—such as committing spend or changing supply routes—stakeholders need an auditable trail: where did this assertion come from, who validated it, and how confident is the mapping?
4. Operational discoverability. Agents need a routing map of data: what sources exist, what they contain, what access patterns they support. Without this, agents cannot know which canonical source to trust.

None of these are solved by a larger model. They are properties of data infrastructure, metadata, and systems design.

Why the model/compute narrative is incomplete

It is seductive—and sometimes true—to point to model improvements as the primary lever for better agentic behavior. But consider three dynamics that show why the narrative is incomplete:

• Diminishing returns at the interface. Bigger models improve reasoning but only if the inputs they receive are accurate and actionable. No matter how fluent a model, if it retrieves a stale contract version or mismatches customer IDs, the agent will err.
• Hardness of integration. Agents must operate across APIs, databases, document stores, and human workflows. The friction is rarely in model inference latency; it is in authentication, schema mismatch, incomplete entity resolution, and access policies.
• Trust and auditability. Organizations will not let agents run unconstrained decisions when there is no traceable chain from data sources to actions. This is a compliance and governance problem more than a compute problem.

In short: models enable potential; data infrastructure enables reliability.

Common enterprise data pathologies that cripple agents

Enterprises typically accumulate technical debt in their data layers. When agents attempt to act in this environment, specific failure modes appear:

• Siloed canonical sources. Different teams maintain their own notion of a customer, product, or contract. An agent that reconciles these in real time faces an identity crisis.
• Inconsistent schemas and semantics. Column names, units, and business logic drift across systems. Parsing a numeric field can be a minefield of implicit transformations.
• Poor metadata and discoverability. Datasets exist, but no one knows their recency, quality, or sensitivity. Agents will either avoid them or make dangerous guesses.
• Fragmented access controls. Authentication and authorization rules differ wildly. Agents often need elevated privileges to act but policies are scattered and manual.
• Unversioned and mutable records. Without versioning of documents and records, agents cannot reliably reproduce decisions or understand when an earlier conclusion was invalidated.
• Latent data quality issues. Missing values, OCR errors in scanned invoices, and legacy encodings propagate into embedding indexes and knowledge graphs, polluting agent reasoning.

These issues are not exotic. They are endemic—and invisible until an agent makes a costly mistake.

How data failures manifest in agentic behaviors

Failure modes aren’t abstract. They look like this:

• An agent reorders stock based on a dataset that omits backordered items because that table was maintained by a different ERP instance.
• A customer support agent promises a refund using an old policy captured in a PDF scanned three years ago; the policy changed and the customer escalates.
• An autonomous contract-negotiation agent signs attachments from a shadow repository that contain non-compliant clauses because the repository was outside legal’s governance.

Each of these can have financial, legal, or reputational impact. The model may have been fine; the agent failed because the data pipeline was not fit for use.

Reorienting investment: from models to data flywheels

Scaling agentic AI requires a shift in priorities. Instead of treating data as a secondary concern, enterprises should build data flywheels that create compounding value for agents. Key components of that flywheel include:

1. Data discovery and semantic cataloging. A living catalog that ties datasets to business concepts, shows freshness, and encodes usage constraints. This reduces cold-start friction for any agent.
2. Reliable canonicalization and identity resolution. Unifying entity identity across sources so agents can reference the same customer, order, or SKU regardless of where the evidence lives.
3. Embeddings and retrieval architecture engineered for realities. Precomputing and versioning embeddings, using hybrid retrieval strategies (dense + sparse), and co-locating vector stores near operational systems to cut retrieval latency.
4. Provenance, lineage, and immutable checkpoints. Every assertion the agent makes should be traceable to a canonical source and a particular timestamped view of that source.
5. Observability and feedback loops. Telemetry on agent actions, data quality, and downstream impacts that feed back into dataset remediation and model prompt updates.
6. Traffic shaping and policy orchestration. Dynamic policies that determine which agent actions require human oversight versus full autonomy, based on confidence, impact, and compliance rules.

These are not academic desiderata; they are engineering priorities that directly reduce incidents, increase agent autonomy envelopes, and accelerate ROI.

Practical steps to unblock agents today

For teams building or deploying agentic systems, here are pragmatic steps that pay back quickly:

• Map the decision surface. Identify the specific decisions agents will make and enumerate the exact data points required for each decision. Begin with the smallest useful slice.
• Create canonical read paths. For each key entity, choose a canonical source and implement a stable API or materialized view. Agents should not have to query five systems to know a contract state.
• Adopt metadata-first ingestion. Capture provenance, schema, and recency at ingestion time. Treat metadata as first-class data so agents can reason about trust.
• Version everything an agent relies on. Datasets, embeddings, retrieval indices, and transformation logic must be versioned so actions can be reproduced and rolled back.
• Instrument human-in-the-loop gates. Start with conservative autonomy and empower agents gradually as observability and data quality improve.
• Operationalize data quality as code. Implement automated checks, lineage tests, and alerting that tie anomalies back to specific pipelines and owners.
• Design for graceful degradation. Agents should have fallbacks: ask for clarification, raise a human task, or switch to a restricted set of actions when data confidence is low.

These steps emphasize making data predictable and legible—two qualities agents need more than raw predictive power.

Architectural patterns that matter

Some architectures have intrinsic advantages for agentic workloads:

• Semantic layer. A runtime layer that maps business concepts to physical data stores and exposes unified query semantics to agents.
• Hybrid retrieval fabric. Combining vector search for unstructured context with structured queries for authoritative facts reduces hallucination risk.
• Event-driven synchronization. Event streams and change data capture ensure agents see near-real-time state without hammering transactional systems.
• Policy-as-code and capability gating. Encoding who or what can act, under what conditions, and with what approval paths keeps agent autonomy aligned with governance.
• Embeddings registry. A system that versions and stores embeddings alongside the data they represent, so retrieval results are traceable.

When these patterns are combined, agents become resilient: they can reconcile ambiguity, explain actions, and be progressively granted more autonomy.

Beyond technology: cultural and organizational shifts

Fixing data is not purely a technical exercise. It requires a reframing of incentives and processes across an organization:

• Measure data-level outcomes. Track metrics not just for model accuracy but for dataset freshness, discoverability, and resolution rates of agent-initiated exceptions.
• Reward upstream quality. Teams that produce canonical data must be measured on the downstream cost of corrections they create. This aligns incentives away from optimizations that only benefit local owners.
• Institutionalize remediation workflows. When agents flag bad data, fixes should propagate quickly through prioritized pipelines, not into a backlog that never closes.

These shifts create a virtuous cycle: better data makes agents more useful, which generates more agent-derived telemetry and value to invest back into data improvements.

The ROI of getting data right

Investments in data infrastructure have a multiplier effect for agentic AI. Dirty data forces repeated manual intervention, constrains what agents can do autonomously, and requires expensive guardrails. In contrast, a disciplined data posture yields:

• Faster time-to-autonomy for agents and reduced human oversight cost.
• Lower incident and legal exposure from auditable, versioned actions.
• Improved user trust as agents make reliable, explainable decisions.
• Greater leverage of the same models and compute budget, because cleaner input yields fewer errors and less need for conservative prompting.

Put bluntly: for the same spend on models and GPUs, companies with better data will extract far more business value from agentic systems.

Conclusion: data as the infrastructure of agency

The AI story of the next decade will not be written solely by model size or flops. It will be written by organizations that recognize agency requires a new kind of infrastructure—one that makes data controllable, comprehensible, and accessible.

Agentic AI is a systems problem. It composes models, orchestration, interfaces, and people. But the wire that binds these components together is data. If that wiring is frayed, agents will misfire. If it is robust, agents will outperform expectations and unlock new workflows that were previously unthinkable.

Every enterprise rushing to deploy agents should ask a simple question before scaling: have we made our data fit for agency? If the answer is no, the path forward is not to buy a bigger model; it is to invest in the plumbing, standards, and operational rigor that will let agents act safely, reliably, and profitably.

Clean data is not a boring prerequisite. It is the stage on which agentic intelligence will either shine or stumble.