AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Raising the quality of insights in agentic analytics

Collected Nov 2, 2026

A sponsored report from MIT Technology Review's Insights group, produced in partnership with Snowflake, examines how organizations rate the output of their analytics agents and what they are doing to improve it. The central finding is a gap between adoption and payoff: across the surveyed companies, an average of 48% of insights generated by analytics agents lead to actions that improve the business. Respondents say that figure was 37% a year earlier, so the trend line points upward, but more than half of today's agent output still fails to help the business. The report also notes that respondents generally rate descriptive insights—the kind that explain what happened—well, while reviews of prescriptive insights, the ones that recommend or trigger actions, are mixed.

The report identifies a high-performing cohort it calls "insight leaders," where more than 70% of agent insights lead to business-improving actions. Those organizations also rate their agents' insight quality more highly than the full respondent pool, and the gap is largest for prescriptive insights. This suggests the difference between average and leading performance is concentrated exactly where the stakes are highest: not in summarizing data, but in recommending what to do next.

When agents produce weak insights, executives point to two dominant causes. Lack of business context is cited by 55%, and fragmented or inaccessible data by 45%. Fewer insight leaders report context problems, which fits the pattern: organizations that supply agents with a clear picture of their business appear to get more usable recommendations.

Volume is a second constraint. At 52% of surveyed organizations, agents generate more insights than teams can properly evaluate, and so act on. Only 24% say they can evaluate all agent insights; most review them case by case, a process the report describes as currently human-intensive. In practice this creates a bottleneck that resembles the alert-fatigue problem familiar from security operations and observability tooling: once output outpaces review capacity, even good insights get ignored, and the review step becomes the real limit on how much automation delivers.

On remediation, the surveyed companies are prioritizing stronger AI and data governance, cited by 60%, and adding a semantic layer to their data architecture, cited by 45%. The experts interviewed for the report describe the semantic layer as crucial for high-performing agents. A semantic layer is a defined model of business concepts—metrics, dimensions, entity relationships—that sits between raw data stores and the tools querying them, so that "revenue" or "active customer" means the same thing to an agent as it does to a finance analyst. For developers building agentic analytics, this implies the work is less about prompt tuning and more about the data and definitional plumbing underneath: consistent metric definitions, retrievable context about business rules, and governance that constrains what an agent can act on.

A likely trade-off is speed of deployment versus depth of investment. Wiring up a semantic layer and governance controls takes time before agents improve, which may explain why the average organization sits at 48% while a smaller group clears 70%. Another open question is how organizations will close the evaluation gap—whether through automated checks on agent output, tighter scoping so agents emit fewer but better recommendations, or continued human review. The report frames the stakes plainly: insights that fail to improve the business undercut the return on agentic analytics investment, and flawed insights that trigger wrong actions can carry reputational or financial cost.

The findings come from a custom content report, researched and written by humans under human oversight, and it is worth reading the numbers as survey-based averages rather than audited results.

Why it matters: Teams deploying analytics agents should treat output quality and review capacity as first-class design constraints, not afterthoughts. The reported correlation between context, semantics and better prescriptive insights suggests investment in a semantic layer and governance is where product and data engineering effort is most likely to pay off, though that causal link is an inference from survey responses rather than a measured result. Organizations that let agent output outrun their ability to evaluate it risk turning a productivity tool into a queue nobody clears.

Read at MIT Technology Review · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

In just a few years, generative AI has changed the face of enterprise analytics, notably by making analytics accessible to users across the organization. Agentic AI promises to do even more, by automating entire workflows and taking actions based on agent insights. But there is a problem: The quality of insights that agents currently generate…