Skip to content

    Fewer Than Half of AI Analytics Agent Insights Lead to Action

    AL
    Aria Lin

    November 4, 2026

    Fewer than half of the insights produced by enterprise analytics agents lead to anything that improves the business: on average, 48% do, according to a report from MIT Technology Review Insights, up from 37% a year earlier as recalled by respondents. Behind that average sits a

    Fewer Than Half of AI Analytics Agent Insights Lead to Action

    Fewer than half of the insights produced by enterprise analytics agents lead to anything that improves the business: on average, 48% do, according to a report from MIT Technology Review Insights, up from 37% a year earlier as recalled by respondents. Behind that average sits a small cohort the report calls "insight leaders," where more than 70% of agent insights lead to business-improving action, a gap of at least 22 percentage points that may point to data and context problems more than to the agents themselves. For researchers, developers and product managers shipping agents into analytics stacks, the question is whether anyone can use what the agents produce.

    This piece looks at what the report found, why so many insights go unused, and what the proposed fixes imply.

    What Happened

    The report, from MIT Technology Review Insights, finds that on average 48% of insights generated by analytics agents at the surveyed companies lead to actions that improve the business. Respondents say the figure was 37% one year ago, an improvement of 11 percentage points. That earlier number is recalled by respondents rather than measured, so the trend is best read as perception, not a tracked result.

    The report frames the headline result bluntly: fewer than half of agent-generated insights at the organizations surveyed lead to business improvement. An analytics agent, in plain terms, is an AI system that examines company data on its own and proposes findings, such as a sales dip or a cost spike, for people to act on. The metric here is not whether the agent produced something, but whether the finding changed what the business did, and changed it for the better.

    A provenance caveat matters here. This is sponsored or custom content, produced by Insights, MIT Technology Review's custom content arm, and not by the publication's editorial staff.

    The "insight leaders" are the exception. That is a term the report coins for its top-performing cohort, and in that group more than 70% of agent insights lead to business-improving actions.

    Why It Matters

    An agent that explains what happened is a reporting tool. An agent whose findings drive what a company does next is closer to a decision-maker, so the quality of its output carries real weight. The report names two main reasons for substandard insights: executives cite a lack of business context (55%) and fragmented or inaccessible data (45%).

    A second problem is volume. At 52% of organizations, agents "generate too many insights to be properly evaluated and therefore acted upon." Only 24% of organizations can evaluate all of their agent insights, which the report puts as "less than one-quarter." In other words, at most one organization in four can say it has looked at everything its agents produced.

      • 48%: average share of agent insights that lead to business-improving action
      • 55%: executives blaming missing business context
      • 45%: executives blaming fragmented or inaccessible data
      • 52%: organizations whose agents produce more insights than can be evaluated
      • 24%: organizations able to evaluate every insight

    The findings compound each other. If people must vet each insight and roughly three-quarters of organizations cannot vet them all, then unreviewed output either sits unused or gets acted on unchecked. The first wastes the automation. The second risks acting on a wrong answer.

    Competitive Landscape

    An engineer sits at a desk with a closed laptop, facing a rack of servers with green and amber status lights in a bright room.

    The report does not name vendors, products or benchmarks, so there is no vendor ranking to offer here, and none is invented. Its only competitive axis is internal: the "insight leaders," set against the 48% average. In this group, more than 70% of agent insights lead to actions that improve the business, a lead of at least 22 percentage points.

    The report describes the group only as "a small cohort." That matters for anyone hoping to copy them: without knowing what separates leaders from the average, the number works as a target and not as a recipe.

    Readers should also resist treating the leaders' 70% as proof of what works. A cohort selected for high scores will naturally look good on that score, and the unnamed sponsor, undefined sample and undefined measure leave no way to test the claim from outside.

    The Bigger Picture

    The remedies companies report are mostly about infrastructure. Some 60% are prioritizing stronger AI and data governance (the rules and controls over who can use which data and how AI outputs are checked), and 45% are adding a semantic layer to their data architecture. A semantic layer is a shared dictionary on top of the data that records what business terms and metrics, such as "revenue" or "active customer," actually mean.

    That 45% is a separate question, asked of a separate group, from the 45% of executives who cite fragmented data as a problem. The matching number is a coincidence, and the two should not be read together.

    The logic of the fix still lines up with the 55% context figure: an agent that does not know how a company defines a metric can give a confident answer on the wrong definition. That is our inference, not a finding of the report, which does not test whether semantic layers raise the share of useful insights.

    Neither remedy figure is universal. By implication, 40% of companies are not reported as prioritizing governance and 55% are not reported as adding a semantic layer.

    What's Next

    The sponsored report gives no timelines or forecasts, and none are invented here. It does supply two things worth tracking.

    First, whether governance (60%) and semantic-layer (45%) investments close the distance between the 48% average and the 70%+ leaders. A follow-up that names the cohort's size, definition and tooling would test it.

    Second, whether evaluation can scale. If 52% of organizations already cannot properly evaluate what their agents generate, more capable or more numerous agents will widen the problem unless review needs fewer people.

    The most telling detail is that the likely cure for agent hype is old-fashioned data work: agree on what "revenue" means, and make the data reachable. The irony is that the survey itself lacks the context it prescribes, with no sponsor, sample size or definitions attached to its numbers.

    For business buyers weighing an analytics agent, the translation is this: at the report's average, roughly one in two insights leads nowhere useful, and about 76% of organizations cannot review all of them. Before paying for more agent output, ask any vendor how it measures the share of insights that lead to action, and budget for the metric definitions and review capacity that determine whether the output gets used.

    -- Aria Lin, Enterprise Technology Analyst


    Sources: Mit · Twitterapi · MIT Technology Review

    More on Revuzia