Survey finds many companies get confident wrong answers from AI agents

In short: Many companies are rolling out AI agents, but a new survey says the information these agents rely on is often not reliable.

What's going on

VentureBeat Pulse Research surveyed 101 organizations with more than 100 employees about how their AI agents get company information. AI agents are tools that can answer questions or do tasks using a company’s files and data.

The survey’s main finding is a “context gap.” That means the agent sounds sure, but the information it was given is incomplete or inconsistent, like handing a helpful assistant an out of date binder. In the past six months, 57% of organizations said their agents produced confident but wrong answers that were traced back to missing or inconsistent business context. Of those, 31% said it happened more than once.

Many companies rely on something called retrieval-augmented generation, or RAG. RAG is a method where the AI looks up relevant documents before it answers (like searching a filing cabinet, then writing a summary). RAG was the top primary approach in the survey at 38%.

The tools companies use for that lookup are also shifting. “Provider-native” retrieval, meaning search features built into big AI platforms, led usage. 40% reported using OpenAI file search and 38% reported using Google Vertex AI Search.

What to watch

Companies say they want more control, even while they use bundled tools. A plurality, 36%, said they plan to keep separate “best-of-breed” tools, and 57% said they expect to switch or add a retrieval provider within a year. The survey also found 58% are building or running a “governed semantic layer,” which is a shared set of agreed definitions for company data, but only 25% said it is already in production.

Source: Venturebeat