Evidence, and What the Tool Produces
A Pointer, Not the Thing Itself
The canon of this course, stated for audit: AI output is a pointer to evidence, not evidence. A summary of a contract is not the contract; an extracted clause is a claim that the clause exists; a flagged journal entry is a hypothesis about a journal entry. Each of these is useful in the way an index is useful — it tells you where the evidence might be and roughly what shape it might have. The assurance enters at the moment a person follows the pointer to the source document and finds the thing actually there, and not before. This ordering matters because tools present conclusions in finished form, which quietly reverses it: the summary arrives looking like the end of the work when it is the beginning. Every figure or claim a tool produces is tied back to source by a person before it enters the file.
- A summary of a contract is not the contract; a flag on a journal entry is a hypothesis
- Tool output works like an index: it says where evidence might be, not that it is there
- Assurance enters when a person follows the pointer to source and finds the thing there
- Finished-looking output reverses the ordering: the summary is the beginning of the work, not the end
Relevance and Reliability Still Decide
Nothing in the arrival of AI changes what evidence is. The IAASB's standards on audit evidence — and the PCAOB's, on their side of the water — ask, in substance, the same two questions they have always asked: is this relevant to the assertion, and is it reliable given where it came from and how it was produced? Tool output sits awkwardly against the second question, because it is derived information whose derivation you usually cannot inspect, produced by a process with a known tendency to be fluent when wrong. That does not make it worthless; it makes it low on the reliability ladder, roughly where an unverified analysis prepared by the client would sit. The tie-back to source documents is what moves the underlying fact up the ladder. The output itself never climbs; the evidence it pointed to does.
- The evidence standards ask, in substance, what they always have: relevance and reliability
- Tool output is derived information from a process you usually cannot inspect
- It sits low on the reliability ladder, near an unverified client-prepared analysis
- Tie-back moves the underlying fact up the ladder; the output itself never climbs
Demand the Locator
The practical rule that makes the canon usable at speed: demand a locator, not an assertion. A tool that reports a change-of-control clause should say which document, which page, which passage — and a person opens the document and reads the passage before the clause is treated as existing. A locator is cheap to check and hard to fake convincingly, whereas a source name alone is exactly what a language model invents best. Tools that cannot point are not useless — they can draft, structure and suggest — but their output cannot support anything in the file, and the team should know which kind of tool it is holding before the busy season, not during it. Where the volume makes opening everything unrealistic, the answer is a deliberate sample of the tool's pointers, which is the subject of the last lesson of this module.
- Demand document, page and passage — then a person opens the document and reads
- A locator is cheap to check; a bare source name is what a model invents best
- Tools that cannot point can still draft and structure, but cannot support the file
- Where volume defeats opening everything, sample the pointers deliberately
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.