A compliance review may examine hundreds of files and still answer the wrong question. If the sample excludes customers who abandoned onboarding, a clean result says little about whether the process is fair or complete. If all files come from one branch, the review cannot confidently describe the entire bank.

The Basel Committee's compliance principles support an independent, adequately resourced compliance function, subject to internal audit review. UK firms should separately establish which SYSC 6 provisions apply to their entity and activities; detailed investment-business monitoring rules are not a universal sampling prescription for every bank process. The method below is my practical assurance design.

State the decision before selecting the sample

Write a one-sentence test objective. For example: “Determine whether newly opened high-risk business accounts received the required approval before activation.” This is more precise than “review onboarding compliance”. Define the applicable control, period, entities and evidence that would demonstrate success or failure.

Decide what the result will support. A diagnostic review identifies failure mechanisms. A statistical estimate supports a population-level defect rate, subject to assumptions. A targeted review examines deliberately selected high-risk cases. Mixing these purposes creates misleading confidence: a sample chosen because cases looked problematic cannot estimate the ordinary failure rate without an appropriate design.

Reconcile the population first

Obtain the population from a controlled source and compare its totals with an independent operational record. Include rejected, cancelled, manually processed and exception cases where relevant to the objective. Capture extraction date, query version and exclusions. If the population cannot be reconciled, report that limitation before interpreting the test results.

Preserve the original selection. Replacing an unavailable file with an easy one can systematically remove the weakest records. Treat unavailable evidence as a finding or a documented limitation, depending on the test criteria. Do not silently assume that a missing record means the underlying control operated correctly.

Combine methods without confusing their conclusions

A sensible programme can contain a random component and a separately labelled risk-targeted component. Stratify by factors likely to change control performance: product, jurisdiction, channel, manual override, system version or processing team. Give rare but consequential scenarios explicit coverage even when they are unlikely to appear in a simple random sample.

Where statistical precision matters, agree the confidence level, tolerated error, expected failure rate and sampling assumptions before testing. There is no credible universal rule that 30 files prove compliance. Independence, clustering and population size matter. Where a defensible estimate is unnecessary, a smaller diagnostic review may be more useful than a large sample whose results cannot change a decision.

Classify failures by consequence

In a hypothetical test, five of 100 randomly selected account openings lack evidence of approval. Two more defects appear in a separate targeted review of 20 overrides. Report those populations separately. A combined seven-out-of-120 rate would obscure the different selection methods.

Investigate whether “missing evidence” reflects absent approval, lost records or a faulty extraction. Each has different implications. Then assess customer exposure, duration and whether the weakness affects other products. An isolated formatting error and a control that permits prohibited activity should not share a severity merely because both count as one exception.

Require a useful closure test

Agree the remediation evidence when the finding is raised. Policy updates can address design; they do not establish sustained operation. A closure review should test cases created after the repair and consider whether historical customers need correction. Where deployment occurred in stages, ensure every relevant system version has been covered.

Maintain a review record containing the objective, population, selection method, test steps, exceptions, root causes and conclusion boundaries. The report should distinguish what is known from what is inferred. It should also name the decision required: expand testing, restrict a process, repair data or accept a tightly bounded residual uncertainty. That is how compliance sampling becomes useful assurance rather than a count of files inspected.