Organizations are moving from testing AI agents to embedding them in business operations, but a new MIT Technology Review Insights report argues that many enterprises have not prepared the data systems those agents need. The report, published August 12 and based on a survey of 300 data and technology executives, identifies limited access to enterprise information and legacy infrastructure as major barriers to trustworthy, fast decisions.

The findings suggest that the next obstacle for agentic AI may be less about model capability than about whether agents can reach the right data, understand its business context, and interact safely with the systems where work happens. That matters for product teams and enterprise buyers planning deployments that go beyond answering questions to taking operational action.

What the survey found

According to the report, AI systems can access an average of 45% of company data across the surveyed organizations. Access falls to 30% or less among the group classified as “data laggards,” while “data leaders” provide access to more than 70% of their data.

The report links that difference to confidence in AI agents. Only about half of all respondents said they trusted their agents’ decisions to be accurate and relevant. Among the data leaders, however, the report says 100% expressed that trust.

The report also identifies a gap in scale and speed. Sixty-six percent of data laggards said legacy data systems limit their ability to scale AI agents, while 68% said those systems prevent agents from making decisions quickly. Among data leaders, only 8% reported either constraint.

Those figures portray data access as an operational requirement rather than a background infrastructure issue. An agent that cannot retrieve current supply-chain records, point-of-sale information, or human-resources data may produce a plausible answer but still be unable to complete a workflow reliably.

Why legacy systems matter

Traditional enterprise data environments were generally designed for applications, reports, and human users with defined access paths. AI agents create a different demand: they may need to combine structured records with unstructured documents, interpret information in context, and make decisions while connected to live operational systems.

The report argues that even relatively recent legacy systems can create friction when agents need this type of access. Data may be distributed across incompatible platforms, governed by inconsistent permissions, or stored without the business definitions an agent needs to distinguish an authoritative record from an outdated or incomplete one.

That distinction is important for enterprise AI. Giving an agent broader access without improving data quality, permissions, lineage, and context could increase the number of actions it can take without making those actions more dependable. For builders, the problem is therefore not simply connecting a model to more sources. It is creating controlled access to information that is current, interpretable, and appropriate for a particular task.

The report names improved access to structured and unstructured data as the leading initiative for scaling agents. It also highlights stronger data governance that includes business context, along with greater automation of data management.

Evidence and limits of the claims

The central evidence comes from a survey of 300 data and technology executives described in the MIT Technology Review Insights report. The categories of data leaders and data laggards, as well as the reported percentages for access, trust, and system constraints, are therefore survey findings rather than independently audited measurements of enterprise deployments.

The report also cites a Gartner prediction that AI agents could augment or automate 50% of business decisions by 2027. That is a forecast, not an established result, and the material provided does not identify the methodology or assumptions behind it.

The publication carries an important disclosure: the content was produced by Insights, the custom-content arm of MIT Technology Review, rather than by MIT Technology Review’s editorial staff. The report says its survey and writing were produced by human researchers, writers, editors, analysts, and illustrators, with any AI tools limited to secondary production processes subject to human review.

That disclosure does not invalidate the survey, but it means readers should treat the strongest comparisons and conclusions as research-backed content associated with a custom report. The source provides no independent case studies, technical audits, or named enterprise deployments that would verify whether the data leaders achieved the reported outcomes in specific workflows.

Implications for builders and enterprise buyers

For AI builders, the findings shift attention toward the systems surrounding a model. An agent used for procurement, customer operations, finance, or workforce management needs more than retrieval. It needs access policies, fresh records, clear source priority, error handling, and a way to explain which data shaped a decision.

Product teams should also define what an agent is allowed to do when information is missing or conflicting. A system that pauses for human review may be slower, but it can be more dependable than one that fills gaps with an unsupported assumption. Monitoring should cover not only model responses but also tool calls, data freshness, permission failures, and downstream effects.

For enterprise buyers, the report is a reminder to evaluate data readiness before expanding an agent pilot. Questions should include whether the proposed system can connect to operational systems without creating uncontrolled copies of sensitive data, whether access can be limited by role and task, and whether business users can audit an agent’s decisions.

This also creates a competitive opening for infrastructure vendors. Tools that unify data access, automate metadata management, enforce governance, and connect agents to operational systems may become as important to deployment economics as model selection. But broader connectivity alone will not guarantee better outcomes; reliability depends on the quality and meaning of the information being exposed.

What to watch next

The first signal will be whether enterprises move from isolated pilots to agents that operate across multiple systems. Wider deployment should reveal whether data access improves in practice or whether organizations continue to restrict agents to narrow, low-risk tasks.

A second signal is the emergence of independent benchmarks for agent data readiness. Useful measures could include the share of relevant data an agent can access, data freshness, permission accuracy, recovery from missing information, and the rate of actions requiring human correction.

Buyers should also watch how vendors address governance and auditability. Claims about trusted agents will be more meaningful when vendors disclose failure rates, escalation behavior, and performance across real workflows rather than relying only on model benchmarks or adoption surveys.

Finally, the Gartner forecast cited by the report will need to be tested against actual business-decision automation. If adoption expands without corresponding improvements in trust and controls, enterprises may use agents widely while keeping them away from the decisions where their value would be highest.

Creati.ai perspective

The report’s core message is credible as a deployment warning: agents cannot act reliably on information they cannot reach or interpret. But its figures should be read as survey signals, not proof that data leaders have solved enterprise AI. The more useful takeaway for builders is to treat data access, governance, and system integration as part of the agent product itself.

The market will likely separate vendors that merely make agents easier to launch from those that make their actions measurable, reversible, and accountable. For enterprise teams, the practical test is simple: before asking whether an agent is intelligent enough, determine whether it has the trusted data and controlled permissions required to act.