AI agents should enrich forecasts only with information the model cannot see

Nicolas Vandeput argues that human planners and AI agents should enrich forecasts only when they possess specific information unavailable to the forecasting engine.

The affected decision is whether a baseline forecast should be overridden or enriched. Value requires the agent to identify genuinely new information, connect it to a specific demand event and preserve a measurable audit trail of the adjustment.

The principal failure mode is agents generating confident commentary from the same data already used by the model, increasing overrides without increasing information.