AI agents in analytics still get more than half of insights wrong, study shows

A report published on November 2, 2026 by MIT Technology Review Insights, in partnership with Snowflake, available for download in the original publication, brings a number that should be in every AI budget meeting: on average, only 48% of insights generated by analytics agents at the surveyed companies result in action that improves the business. The other 52% are expensive noise: tokens consumed, dashboards generated, human review time, with no return in revenue, avoided cost, or mitigated risk.

For anyone who founded or runs a company that has already bought (or is pricing) an agentic analytics copilot, this number changes the question that matters. It's no longer "does the agent work?" It's "how much of what it produces do I actually use, and how much am I paying to discard?"

From generative analytics to analytics that acts on its own

The report marks a transition that has been underway since the wave of generative AI applied to data: first, generative AI democratized access to analysis, letting anyone in the company ask something in natural language and get a chart back. The next promise, that of agentic AI, is more ambitious: the system doesn't just describe what happened, it prescribes what to do and, in many designs, executes the action.

It's exactly in this second layer, the prescriptive one, that quality drops. The study itself notes that most of the executives interviewed praise their agents' descriptive insights (what happened), but the assessment of the prescriptive layer (what to do about it) is "decidedly mixed." For the founder, this matters because it's exactly the prescriptive layer that justifies the higher price of the agentic product: if it doesn't deliver, the ROI promised in the contract doesn't materialize.

The one-year leap, and the distance still to go

The picture isn't static. Today's 48% is a significant rise from the 37% that the same organizations reported a year ago, according to the survey. This suggests real maturation, not stagnation. But the metric that separates those profiting from agentic analytics from those merely spending on it is another one: the gap between the market average and a small group the report calls "insight leaders."

The difference between 48% and over 70% isn't a model increment. According to the report itself, it stems from architecture and governance decisions a company makes before the agent runs, not from an LLM upgrade.

Why the agent gets it wrong: context is missing, not intelligence

When insights turn out poorly, the cause cited by 55% of the surveyed executives isn't a limitation of the AI model: it's a lack of business context. The agent knows how to calculate, but it doesn't know what, in that company's culture, counts as a relevant problem or an opportunity worth pursuing. Next comes data fragmentation, cited by 45%: the agent simply doesn't see the whole system it would need to form a reliable judgment.

For whoever sits on the board approving data and AI budgets, this pair of causes (55% context, 45% fragmentation) is a direct message: the agentic analytics bill doesn't end with the platform license. It includes the expensive work, not automatable by AI, of organizing the business's semantics (what "active customer," "churn," and "healthy margin" mean for that specific company) and of unifying the data sources that today live scattered.

Too many insights is also a cost

There's a second problem, less obvious than quality: volume. At 52% of the surveyed organizations, agents generate too many insights to be evaluated, and less than a quarter (24%) manage to evaluate all the insights produced at all. Evaluation, when it happens, is largely manual, case by case.

In short: if the agent produces more than the company can judge, the surplus isn't a benefit, it's a liability. Every unevaluated insight is a compute cycle paid for and not converted, and worse: it's an insight that, if it were prescriptive and wrong, could have triggered a wrong action without anyone noticing in time. The report is explicit about this risk: a mistaken insight that turns into automatic action can cause financial or reputational damage, not just waste.

What companies are prioritizing to unlock this

The surveyed organizations aren't waiting for the next model to solve the problem. They're investing in two concrete pillars:

• AI and data governance, cited by 60% as a priority: rules about who approves what an agent can decide on its own, decision auditing, and clear criteria for when an action requires human review before it goes into effect.
• Semantic layer on the data, cited by 45%: a structure that translates raw tables and metrics into consistent business definitions, so the agent "understands" the company, not just the schema.

The report itself notes that the experts interviewed converge on one point: the semantic layer is a condition, not an option, for high-performing agents. This is consistent with the pattern the data market has been following in recent years, in which data cloud vendors (Snowflake among them, the sponsor of this very study) have been packaging semantic layers as part of their agentic AI offering.

The question for the CTO, and the filter for the founder

For whoever approves the investment, the 48% figure works as a negotiating baseline with any agentic analytics vendor: if the commercial proposal doesn't come with an explicit plan for governance and semantic modeling, the contract is selling only the easy part of the product. The right question for the CTO isn't "does the agent work?" It's "who defined, at this company, what counts as a relevant insight, and who audits the action before it goes into effect?"

A reading caveat is worth noting: the study is content produced by MIT Technology Review's sponsored content arm (Insights), in a commercial partnership with Snowflake, a data cloud vendor with a direct interest in selling semantic layer and governance products.

The aggregate numbers (48%, 37%, 55%, 45%, 52%, 24%, 60%) are useful as a market thermometer, but the report doesn't name the companies surveyed nor detail the sample methodology, which limits how far they can be extrapolated to the specific case of a smaller Brazilian operation. This is a market benchmark, not an independent audit, and should be treated as such when it comes to backing up a board decision.