Enterprise data infrastructure is chief hurdle to agentic AI, says Harvard Business Review

Ninety-six percent of senior leaders believe agentic AI will be critical to their organisation within two years, yet only 23% say they have the strategy and infrastructure in place to support it today, according to a new Harvard Business Review Analytic Services Pulse Survey sponsored by Cribl, an AI platform for telemetry. The figures lay bare a stark readiness gap as companies across every sector push autonomous AI systems from pilot projects into live production environments.

Agentic AI — software that does not merely summarise or suggest but autonomously plans, decides and acts across live business systems — is being adopted at pace. But the survey, drawn from senior executives across industries, reveals that most enterprises are trying to run this new generation of artificial intelligence on legacy observability and security stacks that were never designed for it. The result, the report warns, is that infrastructure is already buckling under the strain.

When AI agents begin reasoning and executing at machine speed across thousands of parallel tasks, telemetry data volumes can multiply by a factor of ten or more. Dashboards built for human analysts typing in queries simply cannot keep up. For organisations at the leading edge of agent deployment, those legacy systems are not just struggling — they are already failing under the load, the survey finds.

Clint Sharp, co-founder and chief executive of Cribl, said the data growth problem is compounding rapidly. “Data is growing at a 30% CAGR, budgets are not, and now AI agents are multiplying that problem by an order of magnitude,” he said. “The infrastructure most enterprises built for the last decade simply wasn’t designed for the agentic workloads of the next one. This research validates what we see every day: organisations know they need to get ready, and the time to modernise that foundation is now, before the rest of the organisation catches up and demands they move at the speed of AI.”

The financial toll is already evident. Among organisations actively deploying agents, 47% report that infrastructure costs have exceeded expectations. Cribl’s customers are being told they could eliminate entire tier-one security triage teams with agentic AI, only to discover they would need to quintuple their data infrastructure spend to support it. At £10 million a year already, that value proposition disappears fast, the company noted.

More broadly, 82% of all senior leaders surveyed anticipate a significant financial cost to meet agentic AI’s data infrastructure demands. The survey also found that 76% report telemetry data volumes have already increased due to agentic AI, with 31% saying those volumes have doubled or more. Without the right telemetry foundation, the report warns, AI systems become black boxes — ungovernable, unexplainable and ultimately unusable at scale.

Why legacy infrastructure fails agentic AI

The core problem, the research makes clear, is that legacy infrastructure was never architected for the telemetry demands of autonomous agents. Traditional observability and security tools capture what happened — they log events for retrospective human analysis. Agentic AI, by contrast, needs to know why in real time, at scale, across every system it touches. That shift requires a fundamentally different data pipeline.

When AI agents operate across thousands of parallel tasks, each action generates its own stream of telemetry — decision logs, context snapshots, intermediate reasoning outputs, system calls, and performance metrics. Where a human-driven query might produce a few hundred data points per minute, an agentic workflow can produce tens of thousands. The survey finds that existing dashboards, built for human typing speeds, are overwhelmed.

The research briefing adds further depth: legacy networks and data flow setups often lack the low-latency access required for real-time decision-making. Pilot projects mask these issues because they run in controlled, small-scale environments. But when organisations attempt enterprise-wide deployment, performance degrades, latency accumulates, errors propagate across multi-agent systems, and telemetry and middleware coordination break down.

Beyond raw volume, the nature of the data itself changes. Agentic AI requires contextual, temporal, actionable and auditable data. The data must move from simple records to decision contexts — grouping related data, preserving intermediate reasoning inputs and providing a chain of custody for every autonomous action. Legacy stacks, designed for static log aggregation, cannot deliver that.

Ryan Kurt, chief executive of The AI Lab, said the infrastructure ceiling is absolute. “Without the right infrastructure, you’ll hit a ceiling. There is absolutely no way to break through it unless you have the data scaffolding, the governance, and the integrated workflows that you need.”

The challenge is compounded by the fact that agents require access to historical and legacy data as well as real-time streams. Modern data platforms need to offer unified, AI-ready access to comprehensive data sources — a capability most legacy systems cannot provide.

Re-architecting the data layer

The survey identifies a clear pattern among organisations that are pulling ahead. They are not simply buying more AI tools. Instead, they are re-architecting their data layer, treating telemetry as a strategic input rather than an afterthought. They are fusing machine data with human context so agents can reason effectively, and they are choosing open, interoperable platforms that give them flexibility as the AI landscape shifts.

This shift in data understanding is critical. The research briefing emphasises that data must be treated as contextual, temporal, actionable and auditable. Organisations that succeed are moving from data records to decision contexts, grouping related data and preserving intermediate reasoning inputs so that agents can explain their actions and be audited.

The survey also highlights the consequences of failing to prepare. Some 46% of organisations cite unclear ROI and performance metrics as the leading consequence of unprepared infrastructure — the primary reason AI projects stall. Measuring agentic AI’s return is complex, moving beyond simple cost savings to include revenue generation, risk mitigation and innovation, but a persistent gap remains between perceived value and provable outcomes, the research notes.

Adoption barriers are formidable. Privacy risks top the list at 57%, followed by talent shortages (53%), unprepared data architecture (53%) and workflows not configured for agentic AI (52%). Without proper governance, the research warns, organisations risk “shadow AI” and compliance gaps. Overly broad access permissions can allow a compromised agent to cause significant harm, and the complexity of these systems can obscure failures and make it difficult to assign responsibility.

Agentic AI is already being applied across domains from automated research synthesis in pharmaceuticals and law to compliance monitoring, marketing campaign optimisation, IT service desk automation, customer support, regulatory reporting, and finance. But scaling from pilot to production remains the central hurdle. The organisations that are succeeding, according to the survey, are those that have recognised that the infrastructure problem is not an IT issue — it is a strategic one that demands a data re-architecture before the agents can truly run.