Agentic Analytics: Why AI Agents Need Better Decision Foundations

When AI Agents Make Decisions – But Half of Them Are Wrong

The promise is enormous: AI agents that don't just analyze data, but autonomously execute entire workflows, recognize patterns, and initiate actions based on them. Agentic Analytics is considered the next major evolutionary stage of corporate management. Yet a recent study by MIT Technology Review Insights, conducted in partnership with Snowflake, paints a sobering picture of the current maturity level – and provides important guardrails for all companies that have AI agents seriously deployed or in planning.

The central finding is this: only around 48 percent of the insights generated by analytics agents in surveyed companies actually lead to actions that improve the business. A year ago, this figure stood at 37 percent – so development is heading in the right direction, but more than half of agentically generated insights dissipate without any measurable effect. For companies that have invested in AI infrastructure, this represents a significant loss of return on investment.

Descriptive versus Prescriptive – The Critical Quality Gap

Particularly revealing is the differentiation between various insight types. Many companies are quite satisfied with the quality of descriptive insights – that is, the ability of agents to describe states, make patterns visible, and summarize historical developments. But with prescriptive insights – meaning recommendations for action on the basis of which agents are expected to act – a decidedly mixed picture emerges.

This distinction is crucial, because prescriptive insights are the true core of agentic AI: an agent that merely describes what was is essentially an elaborate reporting tool. Only when agents can reliably recommend and act does the transformative potential of the technology unfold. This is precisely where the greatest weakness currently lies.

Missing Context as the Core Problem

The study also clearly identifies the causes of poor insight quality. At the top – cited by 55 percent of the surveyed executives – is the lack of business context. AI agents often lack the organizational knowledge needed to correctly interpret data signals: Which anomaly is seasonal? Which metric is strategically irrelevant? What is considered the internal threshold for escalation?

In second place is the problem of fragmented or difficult-to-access data, cited by 45 percent of respondents. Both factors together explain why even technically sophisticated agents regularly fail due to context gaps – a problem that cannot be solved by better models alone.

"An AI agent is only as smart as the knowledge available to it – and that explicitly includes the implicit organizational knowledge that is often formalized nowhere."

Dr. Maik Bunzel, founder and managing director of mabucon.eu, describes this phenomenon as one of the central implementation challenges in practice: companies systematically underestimate how much contextualization work is required before an AI agent can independently draw actionable conclusions. Building such a contextual framework is not a technical task but, above all, an organizational one.

Insight Overload as an Underestimated Risk

Another finding of the study deserves particular attention: in 52 percent of the surveyed organizations, AI agents generate more insights than can be meaningfully evaluated. Only 24 percent of companies manage to actually evaluate all agent insights – the rest operate on a case-by-case review principle that remains highly labor-intensive.

This phenomenon – let us call it Insight Overload – is paradoxical: agents that are supposed to relieve the burden instead create new evaluation effort through sheer volume. This undermines efficiency gains and causes valuable signals to be lost in the noise. The solution lies not in more agents, but in better-configured agents with clear priority hierarchies and relevance filters.

What "Insight Leaders" Do Differently

The study identifies a small group of pioneering companies referred to as Insight Leaders. Within this group, more than 70 percent of agentically generated insights lead to business-improving measures – significantly above average. What do these companies do differently?

• Stronger AI and Data Governance: 60 percent of all respondents cite this as a priority measure – among Insight Leaders, the maturity level in this area is considerably higher.
• Semantic Layer in Data Architecture: 45 percent of companies are deliberately building a semantic layer that enriches data with meaning and enables agents to interpret terms within the correct business context.
• Fewer, but More Sharply Focused Agents: Rather than broad insight production, pioneers rely on targeted task profiles with a clearly defined scope.

The semantic layer in particular is highlighted by experts as a critical success factor. A Semantic Layer acts as a translating bridge between raw data points and business-relevant context – it defines what a "customer," a "revenue figure," or a "deviation" means within the respective business context, and makes these definitions machine-usable for agents.

Governance as a Foundation, Not a Brake

A common misconception in practice is to view governance as an obstacle to AI agility. The study's data refute this emphatically: companies with strong data and AI governance produce qualitatively superior agent outputs and achieve higher ROI rates. In this context, governance is not a bureaucratic straitjacket, but the infrastructure that enables agents to act in the first place.

Dr. Maik Bunzel from mabucon.eu sees this as a fundamental lesson for building productive automation architectures: "Without robust governance structures, the agent lacks the frame of reference it needs to properly prioritize, interpret, and act. Well-implemented governance is not about controlling AI – it is the prerequisite for AI to be able to act in a controlled manner."

Outlook: What Companies Should Do Now

The study's findings provide important guidance for all organizations that take Agentic Analytics seriously. The path to reliable, actionable agent insights is not a purely technical project – it requires an interplay of data architecture, organizational knowledge, and strategic governance.

Specifically, the following priorities can be derived:

• Systematically formalize business context: Implicit knowledge about thresholds, exception rules, and strategic priorities must be made explicit and machine-readable.
• Introduce a Semantic Layer: A semantic data layer is not a nice-to-have, but a fundamental prerequisite for high-quality prescriptive insights.
• Narrow agent scope: Fewer, clearly focused agents with high precision are preferable to many generalists with insight overload as a side effect.
• Quality over quantity: Success metrics for agents should not be based on the number of insights generated, but on the proportion that leads to measurable improvements.

Agentic Analytics is still in a comparatively early phase – but the learning curve is steep. Companies that invest in context, semantics, and governance today will be among the Insight Leaders of tomorrow. The competitive advantage that results from this is likely to be significant – because more reliable agents mean faster decisions, lower error costs, and an AI-powered capacity to act that translates directly into business outcomes.