Why agents stall in pilot

Ninety percent of decision-makers say their organization uses autonomous AI agents. Among decision-makers at those organizations, 85% have hit a roadblock moving agents from pilot into full production in the last twelve months, which works out to 76% among all decision-makers.

The leading obstacle among those who use autonomous AI agents, named by 41%, is legacy systems and fragmented data architecture making agent integration difficult, the same six-to-ten-platform reality every enterprise runs on, showing up as friction the moment an agent has to act across it rather than simply query it.

Close behind, 34% cite an inability to verify or govern the unstructured data and context their agents retrieve. That points to the unstructured data problem underneath much of this report: the 80% of enterprise data that lives in contracts, tickets and email threads is exactly the data agents need, and largely the data most governance stacks were never built to reach. It’s also where the pattern Anthropic has called context rot tends to appear first, where more raw, ungoverned information leaves an agent less able to find the right answer, not more.

The remaining obstacles are more organizational than technical: 31% say competing priorities pull budget or leadership attention away before agents reach production, 28% cite silos between data infrastructure and AI development teams, 25% have no clear process for deciding when an agent is production-ready, and 23% point to unclear executive ownership of AI data risk, echoing the ownership gap described earlier.

These are not model problems. They’re context and control problems, surfacing at the two points where agents most often fail: the data they reason from and the accountability structure around them.

Source: Collibra/Harris Poll survey, p.6 (Q4)