Half of companies don't trust their own AI agents. Bad data is why.
A survey of 300 executives finds that most organisations are feeding AI agents less than half their company data, and it's costing them speed, accuracy, and confidence.
Key points
- AI agents have access to an average of just 45% of company data across surveyed organisations, falling to 30% or less among the weakest performers.
- Only around half of the 300 surveyed executives trust the decisions their AI agents make.
- Every data leader in the survey trusts their agents' decisions; among everyone else, roughly half do.
- Two-thirds of data laggards say older systems prevent their agents from scaling or making fast decisions; only 8% of data leaders report the same problem.
- Every respondent plans to use AI agents within two years, with 69% expecting to use them widely.
AI agents are software programs that carry out multi-step tasks without a human clicking through each step: checking stock levels, pulling order histories, updating shipping records. Businesses are adopting them fast. The trouble is that many agents are running half-blind.
A report from MIT Technology Review's custom research division, drawing on a survey of 300 data and technology executives, found a stark gap between what agents are promised and what they actually get to work with. Our 5 August story on enterprise AI projects reached a similar conclusion: the unglamorous work of connecting data systems tends to get skipped.
So what's going wrong?
The core problem is old data infrastructure. Legacy systems, even ones updated a few years ago, were built to answer questions, not take actions. Asking an agent to check stock levels, pull a customer's order history, and update a shipping record simultaneously is asking something those systems were never designed for.
On average, agents only reach 45% of company data. At organisations the report labels data laggards, that drops below 30%. Hiring a new employee and locking half the filing cabinets before they start is roughly the equivalent.
Around 66% of data laggards say legacy systems stop agents from scaling across the business. About 68% say those same systems prevent agents from making decisions fast enough to be useful.
Who's getting it right?
A smaller group, the report's data leaders, has cleared most of those hurdles. These organisations give agents access to more than 70% of their data, covering structured figures like sales numbers alongside unstructured material like emails and documents. Every single data leader in the survey trusts the decisions their agents make. Among the rest, only about half do.
That trust gap has a direct operational cost. An agent your team doesn't trust is an agent your team ignores.
Data leaders are also prioritising the automation of data management itself, so that keeping information labelled and accessible doesn't demand a small army of analysts.
What does this mean for ordinary businesses?
Gartner predicts AI agents will augment or automate half of business decisions by 2027. Every executive surveyed plans to be using agents within two years. Getting the data foundation right before then isn't a back-office concern: it's the difference between an agent that helps and one that quietly makes bad calls.
The report's clearest recommendation: widen what data agents can reach, and make sure that data carries enough business context, things like which products are seasonal or which customers have special terms, for the agent to act sensibly.
The gap between leaders and laggards is large but not mysterious. It comes down to whether an organisation has done the connective work before asking software to run on top of it. Based on what we've seen across fifteen AI adoption stories since July, that preparation step is the one most companies still underestimate.
