An AI answer needs an evidence trail

Give the assistant a bounded job

Start with a task such as extracting revenue, operating profit and reporting currency from a specified annual report. Define the period, required fields and output structure. Ask it to mark missing information explicitly rather than filling a gap with an unsupported number.

Check that the citation supports the claim

A plausible page reference is not enough. Open the cited passage, verify the entity and period, and check units and footnotes. Keep extracted facts separate from interpretation. For a small document collection, record which version of each source was used.

Send arithmetic to a reproducible tool

If operating profit is 18 million and revenue is 120 million, operating margin is 18 / 120 = 15%. Perform that calculation in a spreadsheet or code using checked inputs. The generated explanation should agree with the calculation, including its units and rounding.

Test the workflow before relying on it

Create a compact evaluation set with known answers, ambiguous tables and questions the documents cannot answer. Record extraction errors, unsupported claims and calculation mismatches. The deliverable is a source-checked assistant with a review process, not an assumption that a fluent answer is correct.

A learning example, not a market recommendation. Work through the assumptions before applying any model.