Four out of five organizations that increased their AI investment this year cannot demonstrate measurable return on it. That figure — surfaced again Monday morning at the opening of the Gartner Data & Analytics Summit Sydney 2026 — is not new, but the explanation being pressed on enterprise leaders this week is: the problem is not the models. It is the data infrastructure beneath them, and specifically the architectural mismatch between the batch-processing pipelines that most enterprises already own and the continuous, real-time data demands that autonomous AI agents cannot function without.
The two-day summit opened today at the International Convention Centre Sydney, the penultimate stop on Gartner’s 2026 global circuit ahead of Mumbai in September. The event draws data, analytics, and AI leaders from across the Asia-Pacific region and lands at a moment when the APAC enterprise market is grappling with the same inflection point that shaped conversations in Orlando, London, and Tokyo earlier this year: how to extract value from AI investments that, in most organizations, are already substantial and largely unproductive.
The 80% Problem Has a Specific Engineering Cause
The opening keynote, delivered by Gartner Director Analyst Georgia O’Callaghan and VP Analyst Jorg Heizenberg, framed the summit’s central challenge as one of pace misread as strategy. The message has been consistent across this year’s circuit: accelerating AI deployment does not solve the ROI problem. Gartner’s research shows it does not.
What the research shows instead is a data-layer failure. According to figures Gartner has presented across its 2026 summit series, 60% of AI projects that lack AI-ready data will be abandoned before they deliver business value. The leading cause is not model underperformance — a separate Gartner survey found that only 23% of failed AI initiatives trace their failure to the model itself. The remaining failures trace to strategy, governance, and — critically — data architecture that was never built to serve autonomous systems.
Gartner issued a sharpened warning on this front on May 26, predicting that by 2027, 40% of enterprises will decommission or demote autonomous AI agents specifically because governance gaps were discovered only after production incidents had already occurred. “Enterprises are treating AI agent governance as binary — either locked down or fully trusted — and that is the root cause of failure,” said Shiva Varma, Senior Director Analyst at Gartner.
Why the Data Lakehouse Cannot Serve an AI Agent
The most technically specific argument being advanced at the Sydney summit concerns what engineers call the “agent-ready data architecture” problem — and it begins with a constraint built into the most widely deployed enterprise data platform of the last five years.
The data lakehouse, the hybrid architecture that combines the cost-flexible storage of a data lake with the management and query performance of a traditional warehouse, was purpose-built for human analytics. It uses open columnar file formats like Apache Parquet, table formats like Apache Iceberg or Delta Lake, and organizes data through a medallion architecture: raw data arrives in a Bronze layer, gets cleaned and standardized in a Silver layer, and is assembled into analytics-ready datasets in a Gold layer. Each transition between layers happens on a schedule — daily, hourly, or at most near-real-time for demanding pipelines. A human analyst queries the Gold layer and reconciles any stale or ambiguous values manually before drawing conclusions.
An AI agent cannot do that. As Gartner analysts and independent researchers have documented this year, business intelligence workloads tolerate stale metadata and inconsistent data definitions because a human can correct for ambiguity at the moment of analysis. Autonomous agents cannot reconcile contradictions in real time — they act on whatever context the data provides at the moment of inference, propagating any staleness or inconsistency downstream through every decision they make from that point forward. This is not a configuration fix. It is an architectural incompatibility: a lakehouse optimized for scheduled batch processing at reporting cadences runs at the wrong tempo for agents that must act continuously.
Google Cloud made this incompatibility explicit in an April 22 engineering post announcing its AI-native lakehouse architecture: traditional lakehouses were designed for the era of reporting and must evolve into an AI-native foundation — one that replaces batch processing with continuous feedback loops and live data streams, giving agents the reliable context they need to transform raw data into action. The post described an architecture built on fully managed Apache Iceberg storage, cross-cloud interoperability, and continuously refreshed context — specifically the data freshness that batch pipelines do not provide.
The practical implication for enterprises deploying agents: the data stack that was modern three years ago is now the primary obstacle to making agents work in production. Only 14% of organizations currently have data architecture rated as ready for agentic AI deployment, according to a 2025 Deloitte survey.
What Failing Fast Actually Costs
The scale of the gap between AI ambition and AI output has grown precise enough that analysts are now publishing financial estimates. McKinsey’s April 2026 analysis of organizations that have attempted to scale agentic AI found that eight in ten companies cite data limitations as the primary obstacle. The organizations that are successfully scaling share one consistent prior investment: they fixed the data infrastructure layer before they built the agent layer.
The financial consequences of not doing so are specific. More than half of IT leaders globally report concerns about AI cost overruns. Gartner projects that task-specific agents will be embedded in 40% of enterprise applications by the end of this year — up from less than 5% in 2025 — and that the agentic AI market will grow from roughly $7.8 billion today to more than $52 billion by 2030. Organizations paying for that infrastructure without the governance and data foundations to make it work are funding experiments their data estate is not equipped to support.
The AI FinOps gap compounds this. Fewer than half of organizations have implemented financial guardrails or cost-management practices for AI workloads, according to Gartner figures cited across the 2026 summit series — a finding that the Sydney agenda is directly addressing through sessions on AI cost quantification and value measurement.
Governance Is an Engineering Problem, Not a Policy One
The Sydney summit’s session led by Ashwin Rajeeva on agent-ready data architecture advances an argument that has consolidated across this year’s Gartner circuit: governance for agentic AI cannot be retrofitted onto the policy frameworks that governed classical analytics. It must be engineered into the data layer from the start.
This distinction matters architecturally. Classical data governance validates data quality and access rights at the point of pipeline ingestion — typically through scheduled checks at Bronze-to-Silver and Silver-to-Gold transitions in the medallion stack. By the time an agent queries data, it has already passed through those checkpoints and been certified clean. But agents operate across systems simultaneously, query data across multiple pipelines mid-inference, and may encounter data that was valid at ingestion but has since degraded in quality or relevance. Runtime governance — enforcement of data access policies and quality standards at the moment an agent acts, not at the moment data enters the pipeline — requires a fundamentally different technical mechanism: continuous observability, federated access control spanning hybrid cloud environments, and governance that is active during inference rather than passive at ingestion.
The Gartner May 2026 guidance on proportional governance adds a second dimension to this problem. Applying identical controls to all AI agents regardless of their autonomy level is itself a failure mode: an agent that reads documents needs different governance from one that writes to production databases. Gartner’s four-tier framework classifies agents by what they can actually do — Observe (read-only), Advise (recommendations only), Act with Approval (human-reviewed execution), and Act Autonomously (unsupervised execution) — and requires governance controls proportionate to each tier. Enterprises treating all agents identically will encounter either over-restriction that slows delivery or under-restriction that creates production incidents.
APAC’s Signal to the Rest of the Field
The Sydney summit, the fifth stop on Gartner’s 2026 global circuit, carries a significance that extends beyond its regional audience. The questions dominating the agenda — on data architecture readiness, runtime governance, unstructured data management, and the specific failure modes of lakehouse architectures in agentic environments — are the same questions that occupied London in May and will arrive in Mumbai in September.
The regional dimension adds a layer specific to APAC: digital sovereignty. Data residency requirements and restrictions on cross-border data flows have become board-level concerns across the region, creating constraints on the federated access architectures that agent-ready data systems require. An enterprise in Australia, Japan, or Singapore deploying agents across hybrid cloud environments must satisfy local data residency laws while building the cross-system, real-time access that agents demand — a structural tension that shapes how APAC enterprises approach agent-ready infrastructure differently from their North American and European counterparts.
The Gartner message, delivered with consistency across six cities this year, is that the organizations pulling ahead on AI are not those with the largest model budgets or the most ambitious agent roadmaps. They are the ones doing the architectural work that does not show up in product demos: replacing batch pipelines with continuous data streams, building runtime governance into the data layer, and treating AI-ready data infrastructure as a precondition for deployment rather than a follow-on investment. Day two sessions continue June 17, covering cloud strategy, data architecture maturity, and the evolving role of the Chief Data and Analytics Officer.
Frequently Asked Questions
Why do 4 in 5 enterprises fail to show ROI on AI investment?
Gartner’s 2026 research finds that AI ROI failure is primarily a data problem, not a model problem. Only 23% of failed AI initiatives trace their failure to model underperformance. The majority fail because organizations lack AI-ready data infrastructure — data that is continuously governed, actively quality-assured, and accessible to agents without human reconciliation at query time. The data lakehouse architecture that most enterprises use was designed for scheduled batch analytics, not the continuous, real-time data access that autonomous agents require.
What makes a data lakehouse incompatible with agentic AI?
Data lakehouses use a medallion architecture — raw, cleaned, and analytics-ready layers — where data is processed on a schedule and quality checks happen at pipeline ingestion points. A human analyst can tolerate stale or inconsistent data because they reconcile it manually. An AI agent cannot: it acts on whatever data context it receives at the moment of inference and propagates any inconsistency forward through every subsequent decision. The incompatibility is architectural — agents need continuous data streams and runtime governance enforcement, while lakehouses deliver scheduled batches and pipeline-time governance.
What is runtime governance for AI agents, and why does it matter?
Runtime governance means enforcing data access policies and quality standards at the moment an AI agent acts, not at the moment data enters a pipeline. Classical data governance validates data when it is ingested. But agents operate mid-inference across multiple systems simultaneously, and may encounter data that has changed since it was last validated. Without runtime governance, an agent with broad enterprise access can act on stale, inconsistent, or unauthorized data without any checkpoint triggering. Gartner’s May 2026 research found that applying the same governance controls to all agents — regardless of their autonomy level — is itself a failure mode that will cause 40% of enterprises to decommission agents by 2027.
What should a data leader do before deploying AI agents at scale?
Gartner and McKinsey research converge on a single prerequisite: fix the data infrastructure before building the agents. Specifically, organizations should audit data sources for freshness and governance coverage, assess whether their current architecture delivers data at reporting cadences or continuously, classify planned agent deployments by autonomy level and apply proportional governance controls, and implement AI FinOps practices to track cost per deployment before scaling. McKinsey found that organizations that completed this infrastructure work before deployment are the ones scaling successfully in 2026.