The 3 Agentic AI Gaps: What CDAOs Must Do Now

Observations from the Gartner CDAO Summit, New York

TL;DR:

Most enterprises are not failing to scale agentic AI because of weak models. They are failing because the underlying data, context, architecture, and governance foundations were never designed for operational AI systems that reason and act across the business.

At the Gartner CDAO Summit, one theme emerged repeatedly: organizations moving successfully from pilots to production are treating context, entity resolution, operational architecture, and runtime governance as shared infrastructure — not application-by-application engineering problems.

This piece outlines the three biggest gaps slowing enterprise AI today and the five moves CDAOs should prioritize now.

The Critical Question Facing Every CDAO Right Now

The conversation at the Gartner CDAO Summit in New York was unusually candid. No one was debating whether to invest in AI. The question — repeated across breakout sessions, boardrooms, and hallway conversations — was simpler and harder:

We’ve run the pilots. Why can’t we scale?

After two days of peer-level discussion, I came away with a clear diagnosis and a clearer prescription. Here it is.

“The organizations investing 4x more in data foundations are not doing it because they love data infrastructure — they are doing it because they understand that AI outcomes are determined by the foundation beneath the model.”

– Ravi Marwaha, Chief Operating Officer and Chief Product & Technology Officer

The Diagnosis: 3 Gaps in Agentic AI 

Most organizations treating scale as a model problem are solving the wrong thing. Gartner’s own data makes this plain:

• Only 17% of organizations have deployed AI agents in production,
  yet 60%+ plan to within two years — the widest ambition-execution gap of any technology in the current survey cycle.
• Organizations with the highest AI-ready data maturity achieve up to 65% greater business outcomes than peers — and invest 4x more in data and analytics foundations.
• 40% of agentic AI deployments are projected to fail by 2027 — not due to model weakness, but due to insufficient governance and interoperability.

The key problem is the foundation beneath the model. Specifically, three compounding gaps:

Gap 1 — The Context Gap. Every AI system being built today reconstructs the business from scratch. Structured data in warehouses. Knowledge in documents. Relationships in CRMs. Operational state in applications. None of it is connected. Agents spend the majority of their runtime — and your team spends the majority of its engineering effort — assembling a picture of the business that should already exist as infrastructure.

Gap 2 — The Architecture Gap. The data stacks most enterprises built for analytics are structurally wrong for agentic AI. Batch pipelines, snapshot-based retrieval, and siloed entity definitions were designed for human-readable reporting. Agents need low-latency operational state, consistent entity resolution across systems, real-time event propagation, and machine-readable semantics. These are different requirements, not incremental ones.

Gap 3 — The Governance Gap. Existing governance frameworks were built for data quality and reporting accuracy. They do not transfer to autonomous systems that act across system boundaries without clean human-in-the-loop checkpoints. When an agent initiates a transaction, modifies a record, or routes a customer, the accountability question — who is responsible, what did the system know, can it be explained — requires entirely new infrastructure.

The Prescription: 5 Moves in Priority Order

Below are the specific actions that separate organizations successfully scaling agentic AI from those cycling through pilots.

Move 1 — Resolve Your Entity Problem 

Before investing further in agent development, audit whether your organization has a single, resolvable view of its core entities: customers, products, counterparties, accounts, risk events. If “customer 4471” means different things in your CRM, your data warehouse, and your risk system, every agent you build will inherit that inconsistency.

Concrete action: Conduct a cross-system entity resolution audit across your top five business entities. Map where definitions diverge. Assign ownership. This is unglamorous work — it is also the unlock for everything that follows.

Quick win: Pick one entity (most organizations should start with “customer”) and establish a canonical definition with a single system of record. Deploy it as a shared service that all AI applications draw from.

Move 2 — Stop Rebuilding Context Per Application

The most expensive and least visible cost in enterprise AI today is context reconstruction — the engineering work required to assemble a picture of the business for each new use case. Every new agent repeats this work. Progress does not compound.

The fix is architectural: Treat context as managed infrastructure, not application logic. Build a governed, continuously updated knowledge layer — encoding relationships, operational state, and provenance — that all AI systems draw from. Gartner now calls this a “universal semantic layer” and classifies it as critical infrastructure on par with data platforms and cybersecurity.

Concrete action: Identify the three pieces of business context most frequently reconstructed across your current AI applications. Build a shared service for each. Measure the engineering hours saved on the next application that uses it.

Quick win: Instrument your existing agent applications to log what context they retrieve and from where. The reconstruction patterns will become immediately visible — and prioritizable.

Watch short explainer video on how live context changes enterprise AI systems and architectures.

Move 3 — Extend Governance to Cover Agent Behavior

The governance frameworks built for data quality do not cover agent action. This is the most underappreciated risk in enterprise AI today. Gartner predicts that by 2030, 50% of agent deployment failures will be due to insufficient runtime governance — not bad models.

Agents require a different governance model: runtime policy enforcement (what actions can an agent take under what conditions), behavioral audit trails (what did the agent do and what did it know when it did it), anomaly detection (when is an agent operating outside its intended parameters), and escalation paths (what triggers human review).

Concrete action: For each agent in or near production, define three things: 

1. the set of actions it is permitted to take autonomously, 
2. the conditions that trigger human escalation, and 
3. the audit log format required to explain its decisions to a regulator or auditor. 

If you cannot define these, the agent is not ready for production.

Quick win: Implement a “shadow mode” for agents approaching production — they generate recommended actions but do not execute them. Human reviewers validate or override. This builds the behavioral baseline needed to calibrate autonomous operation.

Move 4 — Modernize Architecture for Operational Latency, Not Analytical Throughput

Traditional data architecture was optimized for analytical query patterns — high throughput, acceptable latency, batch refresh cycles. Agentic AI requires a different optimization target: sub-second access to operational state, event-driven data propagation, and consistent entity resolution at query time.

This does not mean replacing existing infrastructure. It means adding a layer designed for operational AI: operational data stores, graph-based relationship representations, streaming event buses, and semantic metadata services. The investments that jumped most in Gartner’s 2026 CDAO spending data — AI/ML platforms (55%), data governance tools (40%), MDM and metadata management (29%) — map directly to these architectural gaps.

Concrete action: Identify your three highest-value agentic use cases and map their latency requirements. If any require sub-second access to state that currently lives in batch-refreshed systems, that gap is your architecture roadmap.

Quick win: Identify one high-value entity or relationship that agents currently retrieve from a batch system and move it to an operational store. Measure the latency improvement and the downstream impact on agent reliability.

Move 5 — Reframe Your Mandate

The most significant shift I observed in New York was in how CDAOs were framing their own role. The traditional framing — data leader as enabler, business leader as decision-maker — is dissolving. When agents act autonomously, the CDAO is no longer upstream of the decision. They are co-responsible for it.

This is not a burden. It is leverage. CDAOs who own the context layer, the governance framework, and the architecture that agents run on are positioned to influence every consequential decision the organization makes — not as advisors, but as infrastructure owners.

Concrete action: Map the five most consequential decisions your organization makes today that will likely be influenced or made by AI agents within 24 months. For each, assess whether your current data and governance foundation would support an explainable, auditable agent decision. The gaps are your investment thesis.

Quick win: Present this map to your CEO and CFO. Frame it not as a data problem but as an accountability and competitive positioning problem. The organizations investing 4x more in data foundations are not doing it because they love data infrastructure — they are doing it because they understand that AI outcomes are determined by the foundation beneath the model.

What to Do Monday Morning

If you take one thing from this piece, make it this: the bottleneck in enterprise AI is the absence of managed, governed, continuously updated context that agents can rely on.

The organizations pulling ahead right now are building better foundations — and letting the models do what models do well on top of them.

The five moves above are sequenced by dependency. Start with entity resolution. It unlocks everything else.

Frequently Asked Questions