Cleaning 900 scattered intake records is exactly the kind of messy data job AI handles well.

The core operations — extract fields, normalize text, deduplicate by name+phone, flag incomplete rows — are structurally identical across all 900 records. Once the normalization rules and field mappings are defined, the agent applies them uniformly, which strongly favors automation.

The output format (CSV with standardized fields) and deduplication key (name + phone) are clearly defined, but the full taxonomy of practice area aliases isn't specified upfront and will require enumeration. Edge cases like near-duplicate names with different phones, or partially filled PDFs, need explicit rules or human triage.

Google Forms and Typeform data are exportable to CSV with minimal friction, but PDF intake forms require OCR and structured extraction, which can fail on scanned or non-standard layouts. The agent needs file access, a PDF parser, and a fuzzy-matching library — all available, but setup and access provisioning require human coordination.

Errors are largely reversible since the output is a CSV reviewed before CRM import, not a live system write. However, silently merging two distinct clients or dropping a record without flagging it could cause real downstream problems in a legal context, so a human review gate before migration is essential.

The normalization and deduplication logic is rule-based and doesn't require legal expertise or relationship context. The agent can flag ambiguous cases for human review rather than deciding autonomously, which keeps the judgment burden minimal.