AI Agents Must Verify the Source, Not Just Fact

Multiverse Computing has described a problem with AI agents that connect to external data through the MCP protocol: an agent can give a correct answer while pulling it from the wrong source — or cite the right source while distorting the fact. The authors argue that checking the final answer alone isn't enough; you also need to verify whether the agent actually relied on the correct source.

The issue is that multi-step agents first search several systems for data and only then compose an answer, and at that stage the link between a fact and its origin is easy to lose. If only the end result gets checked, a wrong source choice slips through unnoticed, and trust in the agent ends up resting on an illusion of accuracy.

Ask your AI agent not just 'how much' but 'from where': have it show the exact document, invoice, or record the answer came from. Make a habit of spot-checking sources against originals, especially for figures that feed into financial decisions. Every agent answer in 1Cifer is tied to a specific record in 1C, a CRM, or a bank statement, so the source can be checked with one click.