The State of AI Reliability: What Senior Data & AI Leaders Told Us

Every conversation we have with data and AI leaders right now has the same undercurrent: teams are moving fast, expectations are high, and there’s a lot of pressure to show results from AI investment from across the organization. But beneath this momentum, there’s an uncomfortable, yet pressing, question around whether the data, infrastructure, and even organizational foundations are actually keeping up.

To get to the bottom of this, Monte Carlo partnered with CDO Magazine to survey senior data and AI executives across financial services, healthcare, technology, and other sectors, the majority from large enterprises with revenues of $1B or more. The results are published in our co-authored report, The State of AI Reliability: Perspectives from Data & AI Leaders, published in March 2026.

The headline finding is one we think every data leader needs to sit with: organizations are scaling AI faster than their observability, governance, and measurement capabilities can support. Here’s a look at what the data shows.

AI agent counts are set to explode, but infrastructure isn’t ready

58% of organizations expect their AI agent count to increase substantially, dramatically, or exponentially in the next 12 months. Yet today, only 18% have scaled AI meaningfully across their organizations, and 39% are still in early pilots or not yet in production at all.

Organizations are clearly making headway to start deploying agents, but there’s a clear gap between where they are today and the scale they’re committing to tomorrow. This is happening while foundational issues around data quality, observability, and governance – all necessary aspects of a trusted, scalable productions system – remain unresolved.

The business case is efficiency, yet the verification method is not

Cost savings and efficiency gains are overwhelmingly the primary justification for data and AI investment, cited by 80% of respondents. This is more than double the next option, revenue growth, at 39%.

Something interesting and conflicting is going on at the same time, however. While efficiency is a priority driver, 62% of leaders say they rely on human review before release as their primary method of verifying AI outputs. Only 25% use automated monitoring or guardrails, and 19% have no consistent verification method at all.

You cannot automate your way to efficiency while manually validating everything that automation produces, and the data reveals to us that efficiency in motivation and efficiency in practice are not aligned right now.

Silent failures are far more common than most teams realize

Perhaps one of the most stark findings from the research was just how prevalent silent failures in data and agentic systems are across organizations. 61% of respondents report that in their data incidents, monitored metrics appeared normal while a critical issue was actually occurring — sometimes (47%) or often (14%).

Nearly all respondents (97%) report that data quality issues have impacted business outcomes, with 36% describing the impact as significant or severe. And detection speed makes it worse: among organizations detecting issues days later than acceptable, 72% report significant or severe business impact — compared to 22% among those detecting at an acceptable speed.

In an AI-driven environment where models continuously consume data, every hour of delayed detection compounds downstream errors. AI systems running on undetected bad data can cascade into a myriad of risks and poor outcomes for businesses and, ultimately, their end consumers.

“AI-ready” is a goal most organizations can’t fully define

“AI-ready” is perhaps one of the most popular concepts discussed among data and AI leadership today, but what does it really mean in practice? According to our research, there’s not a clear consensus of that. 71% of organizations say they are actively preparing to be AI-ready, but when asked how AI-ready is actually defined, only 29% report a clear, documented definition with specific readiness criteria. The remaining 71% are preparing for a destination they can’t fully describe, with 42% operating on an informal or partial definition, and 22% using the term without defining it at all.

Preparation without definition is makes it nearly impossible to measure whether readiness initiatives are actually working.

Data governance and AI governance are operating in silos

Despite the inseparable relationship between data quality and AI output quality, only 23% of organizations govern data and AI under a single unified framework. The observability picture is just as fragmented: only 19% have a unified observability approach across both data and AI systems, while 37% have separate approaches with limited integration, and 9% have no observability for either.

The consequence is measurable. Organizations with unified data and AI governance are more than 1.9x as likely to be “very confident” in measuring business impact compared to those with no formal AI governance. When AI outputs are wrong, tracing the failure back to its data origin requires exactly the unified observability that most organizations don’t yet have.

Investment is skewed toward building, not measuring

When asked what they’re planning on investing in in the next 12 months, respondents’ top priorities were AI/ML engineering (48%), automation (33%), and data quality/reliability (29%).

Security, privacy, and legal risk is the single most-cited barrier to scaling AI confidently, named by 45% of respondents, ahead of the talent gap (40%) and data quality (32%). However, the accountability infrastructure — monitoring, observability, measurement — remains underinvested relative to the risk.

What it means

This moment in the data and AI space, like many times of transformation before it, is defined largely by tensions between ambition and the readiness to execute on it. Organizations are speeding ahead with their plans for AI, yet most are not clear on how they define readiness and how they establish secure guardrails. The organizations best positioned for sustainable AI adoption aren’t necessarily the fastest movers, but rather the ones treating reliability and governance as prerequisites, not afterthoughts.

The leaders who win will be the ones who can prove their agentic initiatives are working reliably, and at scale.

We built Monte Carlo to help data and AI teams do exactly that — with end-to-end visibility across the entire stack, from pipelines and warehouses to AI agents and the outputs they produce. If these findings resonate with what you’re seeing inside your own organization, check out the full findings of this research.

Download the report here.

Our promise: we will show you the product.