AI-Assisted Forecasts in Discounted Cash Flow Valuation: Valuation Date, Information Set, and Audit Trail

This paper examines the growing use of generative artificial intelligence based on large language models in business valuation and related decision making. The scientific literature highlights efficiency gains in structuring information and developing scenarios, but also methodological risks when these tools are used to form valuation assumptions. A key concern is temporal leakage, also known as look-ahead bias, in which outputs may be influenced by information that was not available at the valuation date. Another concern is unsupported factual statements that cannot be traced to verifiable sources. These issues matter because, under International Valuation Standards, the conclusion on value must be stated as of a specific valuation date and must rely only on information that was known or knowable at that date. The paper investigates whether simply instructing a model to reason as of the valuation date is sufficient to prevent temporal leakage. It also examines how control over the information set affects the defensibility, reproducibility, and auditability of AI-assisted assumptions. It uses an illustrative discounted cash flow case and focuses on inflation measured by the harmonised index of consumer prices as an input to nominal projections of free cash flow to the firm. Three scenarios are compared: an open information window, an instruction-only regime without a closed information set, and a closed information window regime that restricts the model to a predefined data pack and requires a documented audit trail with source-linked factual claims and explicit reporting of insufficient evidence where support is missing.