An auditor samples only 25 permitting chatbot chats at random and ignores stratified high-risk permit types. What sampling improvement should the auditor make?

Designing risk-based stratified sampling that overweights high-risk municipal permit types works because random convenience samples of chatbot chats can miss hazardous-use, environmental, or life-safety permitting scenarios where wrong guidance is most damaging. Stratification aligns testing effort with civic risk, improving the chance of detecting material failure modes in the permitting GenAI. Auditors should define strata from permit risk tiers, not from chat volume alone. Simply increasing the random sample to 200 chats with equal probability selection fails because it still treats every chat alike and leaves high-risk permit types to chance, so hazardous-use scenarios can remain underrepresented. Adopting systematic interval sampling across the full annual chat population fails conceptually: fixed-interval selection still gives every chat the same weight and does not concentrate testing on high-risk municipal permit types where errors cause the most harm. Extending sampling coverage across twelve months using proportional allocation by chat volume fails because it allocates by volume rather than risk, so low-volume but high-harm permit types stay underrepresented and ordinary weekday traffic can drown out consequential cases. Exam caveat: overweighting is not ignoring low-risk strata entirely—maintain some coverage of common permits while concentrating power on high-harm types. Operational check: build a stratified sample plan keyed to permit risk categories, document oversampling rules for hazardous types, execute tests across strata, and report findings by stratum so leaders see where the permitting bot fails under consequential scenarios.