Messy timesheet cleanup is exactly the kind of repetitive data work AI handles well.

The task runs weekly on the same CSV schema with the same cleaning rules each time — name normalization, duplicate detection, billable flag review. This is structurally identical every cycle, which is ideal for automation.

Most success criteria are crisp (deduplicated rows, standardized client names, billable flags), but edge cases like 'is this entry a true duplicate or a legitimate re-entry?' require a defined rule or human confirmation. The agent can flag these rather than silently resolve them.

The agent only needs the exported CSV — no live API access, credentials, or external systems required. A canonical client name list would sharpen fuzzy matching but can be derived from the data itself or provided once.

Billing errors — overbilling, underbilling, or miscategorized hours — have direct financial and client-relationship consequences. This is reversible in principle but damaging in practice, so human sign-off before invoicing is non-negotiable.

The cleaning logic is rule-based and the agent doesn't need to understand client relationships or business context to execute it. Flagging ambiguous rows for human review handles the small residual judgment calls cleanly.