Often it's the input not the model that makes AI hallucinate in production.

I was sending my call-summary agent the whole call transcript, and it started inventing next steps and action items nobody said on the call.

Managing AI Hallucinations by Arkadiusz Włodarczyk (Packt, 2026), in Chapter 2, says the same. Extra context adds noise, so give the model only what the task needs.

The fix was a data contract, a fixed schema for what goes into the agent and what comes out.

Strip the input to the needed fields and then validate the output against the same schema before anything is written to the CRM/final sink location.

Arkadiusz Włodarczyk, Managing AI Hallucinations, Packt Publishing, 2026. Chapter 2, Preventing Hallucinations (O'Reilly Learning)

Earlier note on data contracts, Your army of agents needs contracts