Sixty-eight percent of enterprises traced a confident but wrong AI agent answer to missing or inconsistent business context in the past six months, up from 57% in June, according to a VB Pulse July 2026 survey of 101 qualified enterprises. Recurring failures climbed from 31% to 37% in the same period. The sharpest finding: enterprises running or building a governed context layer report recurring failures at 50%, against 21% for those without one.
The inversion is not a paradox. A governed context layer is what makes failures traceable in the first place. Without a shared reference point, wrong answers get blamed on the model or never investigated. The visibility problem compounds with scale: enterprises above 1,000 employees report recurring failures at 55%, versus 30% for mid-market peers, despite being less likely to have a layer in production, 24% against 37%. Separately, retrieval accuracy ranks third in buying criteria at 15%, behind access control and ease of ingestion at 24% each, meaning enterprises are selecting context infrastructure on governance properties while still grading success on whether the answer is correct at 38%.
The full article is worth reading for the breakdown of how enterprises are actually feeding agents context today: 31% use retrieval over documents as the primary source, 13% load documents directly into the context window, and 5% rely on no structured context at all. Sixty-three percent of enterprises are building or running a governed layer, but only 32% have one in production. That gap is where the budget is going. The piece also quotes Redis AI research leader Srijith Rajamohan on why semantic similarity retrieval fails on sentences like 'Rome is closer than Paris' versus 'Paris is closer than Rome', a precise illustration of why more documents do not fix a definition problem.
[READ ORIGINAL →]