AI agents are doing more analysis — but most of their insights still go nowhere

Generative AI has already changed who can use enterprise analytics. Agentic AI promises to go further, automating whole workflows and acting on what agents find. A new report summarized by MIT Technology Review suggests the technology is only partway there: the insights agents produce are often not good enough to act on.

On average, 48% of insights generated by analytics agents at surveyed companies lead to actions that improve the business, according to the report. That is up from 37% a year earlier, a clear improvement — but it still means more than half of agent-generated insights fail to help the business at all.

The gap between describing and deciding

The report draws a sharp distinction between two kinds of output. Companies generally rate their agents' descriptive insights — summaries of what happened — fairly well. Reviews of prescriptive insights, the recommendations on which actions are based, are far more mixed.

A small group of organizations, labelled insight leaders in the report, stands apart. At those companies, more than 70% of agent insights lead to actions that improve the business, and they rate the quality of their agents' output more highly than the survey average. The gap is widest on prescriptive insights, suggesting that the leaders have solved something the rest have not.

Context, not bigger models, is the bottleneck

When agents produce substandard insights, the leading cause is a lack of business context, cited by 55% of surveyed executives, according to the report. Fragmented or inaccessible data is the other major factor, cited by 45%. Fewer insight leaders report struggling with access to context, which points to where their advantage comes from.

A second problem compounds the first: too many insights. At 52% of surveyed organizations, analytics agents generate more insights than teams can properly evaluate and act on. Less than a quarter of companies can evaluate all agent insights; most review them case by case, in a process the report describes as currently human-intensive.

Governance and shared meaning come first

Companies surveyed are responding by investing in the foundations around their agents rather than simply in larger models. Strengthening AI and data governance was cited by 60% as a priority initiative for improving insight quality, and adding a semantic layer to their data architecture was cited by 45%.

Experts interviewed for the report agree that the semantic layer — a shared, consistent definition of what business data means — is crucial to high-performing agents. Without it, an agent can be technically fluent and still misunderstand the business question it is answering.

The lesson for companies rolling out agentic analytics is therefore less about speed and more about discipline: give agents the context, the clean data and the governance to make their recommendations trustworthy, and give people a realistic way to triage what agents produce. Until then, the report suggests, most agent insights will remain interesting summaries rather than decisions worth taking.

Written by the BidAsk Newsroom. Based on reporting from the sources below.