From Sampling to the Full Population
Sampling Was a Concession
Sampling has always been a concession, and it is worth remembering to what. Auditors never believed thirty items out of thirty thousand were the ideal evidence; they believed reading thirty thousand was impossible at a price anyone would pay, and built a discipline — statistical and otherwise — for saying something defensible about a population from a fraction of it. That discipline is genuine, and the concession was still a concession: every sample leaves most of the population unread, and everyone signing knew it. What has changed is the cost of reading. A tool can now pass over every journal line, every invoice, every contract in the population, at a cost closer to computing than to labour. The honest headline of AI in audit is exactly this — the constraint that made sampling necessary is gone in specific, bounded places — and honest headlines deserve precise handling, which is what the next two slides are for.
- Sampling was never the ideal — it was the defensible answer to reading being unaffordable
- The discipline built around sampling is real, and it still left most of the population unread
- What changed is the cost of reading: whole populations can now be passed over at machine cost
- The headline is honest but bounded, which is why the caveats that follow are part of the claim
What Full Population Actually Buys
Full-population work buys two things, and they are worth naming exactly. First, coverage of the haystack: nothing was excluded by the sample design, so the unusual item in the unexamined ninety-nine per cent — the one a well-designed sample was always statistically likely to miss — has at least been read by something. Risks that live in rare items, unusual postings and related-party traces among them, are exactly where that matters. Second, better targeting: instead of spreading effort evenly across a random selection, follow-up hours concentrate on the items most worth a person's attention, ranked by how anomalous they look. Notice what is absent from this list: neither benefit is a conclusion. Coverage means everything was read, not that everything read was understood; targeting means attention lands better, not that whatever it lands on is a misstatement. The buying is real, and all of it sits upstream of judgement.
- Coverage: nothing excluded by sample design, so rare unusual items have at least been read by something
- Targeting: follow-up hours concentrate where the anomaly ranking points, instead of spreading evenly at random
- Risks living in rare items — unusual postings, related-party traces — are where coverage genuinely matters
- Neither benefit is a conclusion: both sit upstream of the judgement work, which has not moved
The Caveat That Keeps It Honest
Now the caveat that keeps the headline honest, and it is canon for this whole course: more coverage is not automatically more assurance. A shallow pass over everything can be weaker evidence than a deep test of a well-chosen sample, because assurance comes from what the auditor did about what was found, not from how much was scanned. A flagged outlier is a question, not a finding: until someone follows it to the source documents — the invoice, the approval, the contract, the bank entry — it is a statistical observation about the population, and statistical observations do not support opinions. And that follow-up still costs what it always cost, which is why full-population testing that flags ten thousand items has not automated the audit; it has generated ten thousand decisions about where the follow-up budget goes. Coverage claims in planning discussions deserve the same scepticism as any other confident assertion.
- More coverage is not more assurance: a shallow pass over everything can be weaker than a deep sample
- A flagged outlier is a question — it points to evidence only when followed to the source documents
- Follow-up costs what it always cost, and ten thousand flags is ten thousand budgeting decisions
- Treat a coverage claim in planning the way you would treat any confident assertion: ask what was actually done
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.