Raising the quality of insights in agentic analytics

The headline number is blunt. Across the companies surveyed, fewer than half of the insights produced by analytics agents, 48% on average, led to actions that improved the business. That is up from 37% a year earlier, but it still means most of what these agents surface goes nowhere.

The bigger risk sits in prescriptive insights, the ones that tell a team what to do. Respondents rated those far less kindly than simple descriptions of what happened. When an agent is wired to act on its own advice, a bad recommendation stops being a wasted slide and becomes a wrong action with real money or reputation attached.

We think the most useful line in this report is the diagnosis. When agents underdeliver, 55% of executives blame missing business context and 45% point to fragmented or hard-to-reach data. That squares with what we hear from practitioners: a model that doesn't know how your company defines revenue, churn or an active customer will confidently answer the wrong question.

So the prescription, stronger AI and data governance (60% are prioritizing it) and a semantic layer that encodes shared definitions (45%), strikes us as sensible. It is the boring plumbing that makes the clever part trustworthy. We would rather see companies spend there than chase the next model upgrade and hope accuracy follows.

The finding we'd underline is insight overload. At 52% of organizations, agents produce more insights than people can properly evaluate, and only 24% manage to review all of them. That is a quiet failure mode. If humans can't keep up with the review queue, the pressure will be to let agents act unchecked, and that's exactly where the weak prescriptive scores start to bite.

We'd also read the numbers with care. This is a survey of self-reported impressions, and the small group of 'insight leaders' reporting more than 70% useful insights is a cohort the report itself defines. The direction is believable; the precision is not something we'd bet a budget on.

Where this leads if it scales is a split. Firms that invest in definitions, governance and review capacity will get agents that genuinely help. Firms that skip it will automate guesswork at speed, and their mistakes will look a lot like insight until the quarter closes.

This is not MIT Technology Review newsroom reporting. The page states it was produced by Insights, the publication's custom content arm, and not by its editorial staff. We treat it as sponsored partner content. It works as a survey snapshot, but keep in mind that the fixes it recommends, semantic layers and data governance, match what data platform vendors sell.