UpShaqo
Intelligence desk
Enterprise Agents Source-backed analysis

Agentic Analytics Buyers Are Shopping for Context, Not Compute

A new survey finds only 48% of analytics-agent insights drive business improvement, and the fix isn't a bigger model—it's a semantic layer. That reframes who gets paid in enterprise AI.

UpShaqo Editorial IntelligenceNovember 2, 20265 min read
Intelligence standard

Independent UpShaqo analysis built from fresh, attributed sources. We explain the impact instead of repeating the announcement.

Read for leverage: focus on the workflow change, the customer problem, and the next action—not only the product announcement.

A year ago, analytics agents at surveyed companies turned insights into business-improving action 37% of the time. Today that figure sits at 48%, according to MIT Technology Review Insights research conducted in partnership with Snowflake. That's real progress. It's also a confession: more than half of what analytics agents generate still doesn't help the business. For a technology category sold on the promise of autonomous decision-making, a coin-flip success rate is an uncomfortable number to publish—even in a sponsored report.

That discomfort is the real story here, not the uplift. Buyers evaluating agentic analytics platforms right now are effectively being told, by the vendor ecosystem itself, that the bottleneck isn't model horsepower. It's context. That reallocates where enterprise budgets—and vendor pricing power—should be heading next.

What the Survey Actually Found

The headline numbers are worth sitting with before drawing conclusions. Descriptive insights—the "what happened" layer of analytics—get good marks from most respondents. Prescriptive insights, the ones that recommend or trigger an action, get decidedly mixed reviews. That gap matters because prescriptive output is the entire value proposition of "agentic" analytics versus traditional dashboards. An agent that describes a sales dip accurately but recommends the wrong fix hasn't automated a workflow; it's generated a plausible-sounding liability.

A small cohort the report calls "insight leaders" breaks from the pack: more than 70% of their agents' insights lead to business-improving action, well above the 48% average. The report attributes the gap to two causes, and the ranking is instructive—55% of executives blame insufficient business context, while 45% point to fragmented or inaccessible data. Context, not data plumbing, is the larger complaint, and insight leaders report struggling with it far less often than the average organization.

Why Context Beats Compute

Here's the analytical leap worth making explicit: if the dominant failure mode were model capability, bigger or newer LLMs would meaningfully close the gap. Instead, the surveyed companies are responding by investing in governance (60%) and semantic layers (45%)—infrastructure, not inference. A semantic layer is essentially a translation layer that tells an agent what "revenue," "active customer," or "churn" actually means inside a specific company's data, rather than letting the model infer definitions from whatever table it happens to query. Without it, an agent can run a technically correct SQL query and still produce a business-wrong answer, because it guessed at a metric definition that the finance team quietly redefined two quarters ago.

This is a useful mini-framework for any operator evaluating analytics-agent vendors: ask not "which model powers this?" but "how does this system know what our numbers mean?" A vendor that can't answer with specifics about metric governance, access controls, and business-rule encoding is selling a demo, not a deployment.

The Insight Overload Problem Nobody Prices For

A quieter but arguably more consequential finding: a majority of organizations, 52%, say their analytics agents generate too many insights to be properly evaluated and acted upon. Only 24% can evaluate every insight their agents produce; the rest triage case by case, and that triage is still human-intensive. In other words, the bottleneck has shifted from insight generation to insight review—exactly the kind of labor-intensive workflow agentic AI was supposed to eliminate.

This is the underserved wedge in the market right now. Vendors have poured effort into generating more insights and, increasingly, into grounding those insights with semantic context. Almost nobody has built the triage layer: tooling that ranks, deduplicates, and routes agent output by business impact before a human ever sees it. A company that solves insight-prioritization—essentially a relevance and risk-scoring layer sitting between the agent and the human reviewer—would be selling into a pain point that 52% of the market already admits having, with very little direct competition. It's a smaller, less glamorous product than "autonomous analytics," but it addresses the actual operational bottleneck the report surfaces.

Competitive Positioning and Who Captures the Margin

The fact that this research was produced in partnership with Snowflake is itself a data point about where pricing power is migrating. Snowflake doesn't primarily sell language models; it sells data infrastructure and, increasingly, the governance and semantic tooling the report identifies as the fix. A survey showing that governance and semantic layers—not bigger models—are what separate insight leaders from laggards is also, not coincidentally, an argument for spending more on the infrastructure layer where Snowflake and its data-platform peers compete. That doesn't make the findings wrong, but buyers should read the recommendations with that incentive in mind, and weigh the data against their own pilot results rather than taking the framing at face value.

Strategically, this suggests the near-term competitive battle in agentic analytics won't be won on model selection—most enterprise buyers can swap underlying LLMs with modest friction. It will be won by whoever owns the semantic and governance layer sitting between the model and the company's actual business logic. That's a stickier, higher-switching-cost position than being "the agent orchestration layer,” and it's where durable pricing power is likely to concentrate over the next product cycle.

What to Do With This Before the Next Budget Cycle

For operators currently running or evaluating analytics agents, three concrete next actions follow directly from the findings:

  • Audit prescriptive outputs separately from descriptive ones before renewing or expanding an agent deployment—the report's own data shows these are judged very differently, and lumping them together hides the real failure rate.
  • Ask vendors to show their semantic-layer and governance roadmap, not just model benchmarks, since that's what the surveyed insight leaders say separates them from the average.
  • Build or buy a triage mechanism for agent output now, while insight-overload remains an underserved problem rather than a commodity feature bundled into every platform.

The broader lesson is unglamorous but important: in agentic analytics, the limiting factor isn't how smart the model is. It's how well the system understands the business it's supposed to be reasoning about.

Sources

#agentic analytics#semantic layer#data governance#Snowflake#enterprise AI#AI procurement

Two doors. Pick one.

Hire the team.
Or become it.