MIT–Microsoft's Agent Confidence Index puts average agent trust at 64/100 — with context, not capability, as the ceiling

A survey of 300 technical experts across 12 industries ranks 101 agent tasks, finding that low-scoring workflows fail on missing business context rather than model weakness.

MIT Technology Review Insights and Microsoft published the 2026 Agent Confidence Index on June 29, ranking 101 agentic tasks across AI, data, and cloud workflows at an average trust score of 64 out of 100. Only 30 of those tasks cleared 70. The subtext of the numbers is the actual story: builders don’t think the models are the problem anymore.

The methodology is broad enough to matter. Polling by Forbes contributor Janakiram MSV’s summary confirms 300 technology executives, team leads, and contributors surveyed across 12 industries and 4 regions in February–March 2026, at companies ranging from startups to firms with upward of $10 billion in annual revenue. That’s not a hype-cycle audience. It’s the buyer.

What scores well is legible and self-contained. Automated generation and distribution of business reports lands at 83.5, boilerplate code at 82.5. What scores badly is anything requiring institutional memory: service mesh configuration and troubleshooting at 37.5, disaster recovery testing at 43, database migration planning at 44.5.

The report itself is direct about why. “Where agent readiness drops is largely due to a lack of business context being supplied to agentic systems,” the MIT Technology Review Insights authors write. Microsoft’s Fabric team frames it the same way: “When an agent produces an incomplete or incorrect answer, it is often not because the model lacks capability.” Their prescription is “building the foundation around the models, the right context to train the agents, along with sophisticated observability to monitor their actions.”

The trust data underneath sharpens the picture. 48% of respondents rated accountability their top worry, 47% cited hallucinations, 59% plan to keep humans in the loop, and 53% are monitoring agent decisions closely. This isn’t skepticism about the frontier. It’s operations discipline catching up with procurement.

Which is why the context and knowledge layer has become the vendor race worth watching. Glean crossed $300 million in ARR in May, up from $200M six months earlier, per TechCrunch; Sierra is scaling on adjacent territory; and startups like LemonLime are pitching the same context-plumbing thesis to the SMB tier that Glean built for the enterprise. The index reads as a scoreboard for capability. It’s really a scoreboard for whoever ends up owning the pipes between an agent and the business it’s supposed to understand.

MIT–Microsoft's Agent Confidence Index puts average agent trust at 64/100 — with context, not capability, as the ceiling

Sources