There is a question every senior finance leader has fielded from an audit committee, usually delivered mildly: where did this number come from? For most of corporate history it had a reassuring answer. A named analyst, a known spreadsheet, a review chain of people who could each defend their step. The answer was slow, but it was an answer.
That answer is dissolving. Board packs and investor materials now contain figures a model computed, narratives an AI drafted, and summaries a copilot assembled from systems no single person has personally inspected. One in four executives now report audit-detected AI errors reaching boards or external audiences, drawn from data most of them do not trust for AI use in the first place (the verification gap). The volume of machine-touched content in board reporting is rising; the ability to defend it, mostly, is not.
Why the old assurance model broke
The traditional defence of a board number was human familiarity. The pack was assembled by hand, so each figure had a person behind it, and the controller's confidence was a summary of the team's confidence. The model scaled poorly but honestly.
Automation broke it in three places at once. Volume. A modern reporting cycle assembles hundreds of figures from dozens of systems; nobody personally re-derives them. Opacity. A figure that passed through a model or an AI summariser carries transformations its presenter cannot narrate from memory. And variability. Content drafted by generative tools can differ between runs, so what was checked and what was published are not provably the same artefact. Adding more review meetings to this pipeline is the instinctive response and the wrong one: review without lineage is confidence theatre performed on a larger stage.
The defensible pack
What replaces familiarity is lineage, the property that every number in the pack can produce, on demand, the account of itself that the committee question implies. Concretely, four requirements:
- Traceable sources. The figure names the systems and data it came from, as a record generated by the assembly process itself, not a reconstruction performed after the question.
- Deterministic recomputation. Same inputs, same number, every time. The published figure can be regenerated exactly, which is the difference between an error that can be diagnosed and one that can only be apologised for.
- Model provenance where models contributed. Which model version, trained on what, approved by whom, performing how, with the record attached to the output rather than filed in a different department.
- One governed path. Every figure that reaches the board travels through the same lineage-preserving pipeline. The side channel, the number pasted from a spreadsheet the night before, is where defensibility goes to die, and closing it is a design decision, not a policy memo.
"Where did this number come from?" should be a query with an answer in seconds, not a project with an answer in weeks.
The offensive version of the case matters as much. Reporting teams spend a remarkable share of each cycle on assembly and reconciling versions of numbers, time that lineage infrastructure returns as thinking time. And a finance function that can answer provenance questions instantly changes its relationship with the committee: scrutiny becomes cheap to satisfy, which is what trust looks like operationally.
The near-term forcing function is simply that the question is now being asked. A committee that has had one AI-flavoured error surface in front of it, and per the benchmark data a quarter of executives report exactly that, does not return to not asking.
Try the question on your own last board pack: pick any figure and time how long it takes to produce its full account. If the answer embarrasses you, that is the gap Prophesee's Reporting module closes, with lineage, deterministic recomputation and model passports behind every figure. Make the pack defensible.