Most enterprises are handing AI agents half the picture and hoping for good decisions. A new report from MIT Tech Review puts a number on the gap: across 300 data and technology executives surveyed, AI agents have access to just 45% of company data on average. In the weakest organizations, that figure drops to 30% or less.
This is the quiet problem behind the agentic AI hype. Companies are racing to deploy agents while starving them of the information those agents need to be useful.
What the researchers looked at
MIT Tech Review surveyed 300 data and technology leaders about how their legacy systems affect AI agent performance. The framing question comes from Gartner, which predicts AI agents will augment or automate 50% of business decisions by 2027. If that holds, the data pipes feeding those agents become the bottleneck that decides who wins.
The report splits respondents into two camps: “data leaders” who give agents access to over 70% of their data, and “data laggards” stuck at 30% or below. The contrast between the two groups is where the real story sits.
The numbers that matter
The split is stark. Here’s how the two groups compare:
- Data access: Leaders give agents 70%+ of company data. Laggards sit at 30% or less. The overall average is 45%.
- Trust in agent decisions: 100% of data leaders trust that their agents make accurate, relevant decisions. Across all organizations, only about half do.
- Scaling blocked by legacy systems: 66% of laggards say old data systems limit agent scaling. Among leaders, just 8% report the same.
- Speed blocked by legacy systems: 68% of laggards say legacy systems stop agents from deciding at speed. For leaders, again only 8%.
What stands out here is the trust figure. Every single data leader trusts its agents. That’s not a coincidence or better prompting. It’s a direct payoff from feeding agents clean, complete, well-governed data. Reliable AI needs a reliable foundation, and the survey shows it plainly.
Why this matters for practitioners
The pressure isn’t theoretical. According to MIT Tech Review, 100% of respondents plan to use agentic AI within two years, and 69% expect to use it widely. Nobody is opting out. The only variable is whether your data estate is ready when the agents arrive.
The report is blunt about the risk: without removing data system constraints, agentic AI won’t deliver the speed and efficiency it promises. You can buy the best agent on the market and still get mediocre results if it can only reach a third of your data.
So what can you actually do with this? The leaders point to a clear playbook:
- Widen data access first. The top priority across all respondents was improving agent access to both structured and unstructured data. Structured tables aren’t enough. Agents need the documents, emails, and messy text too.
- Add business context to governance. Enhancing data and AI governance with business context ranked high. Governance here isn’t red tape. It’s the layer that tells an agent what a number means and whether it can trust it.
- Automate data management. Leaders lean heavily on automation to keep their data estates agent-ready without drowning in manual cleanup.
My take: the companies treating this as a data problem, not an AI problem, are the ones pulling ahead. Buying agents is easy. Making your data worthy of them is the hard part, and it’s the part that separates the 8% from the 66%.
The limitation worth noting
This is survey data, self-reported by executives, not an audit of actual systems. Trust in agent decisions is a perception, and 300 respondents is a snapshot rather than the whole market. The correlation between data access and agent success is strong, but the report describes leaders and laggards, it doesn’t prove one causes the other. Still, the pattern is consistent enough to act on.
The direction is set. Agents are coming to nearly every enterprise within two years. The organizations that spend the next stretch fixing data access, governance, and automation will be the ones whose agents actually earn their trust. Full findings are available in the original MIT Tech Review report.