AI workflow for auto-naming and filing Drive documents

Why this workflow works (and where it breaks)

• Email is the intake point: documents show up as attachments before anyone has time to file them.
• A ledger prevents “AI guessing”: the model shouldn’t invent folder paths; it should look them up.
• Confidence + review queue keeps it safe: fully automate only when matching is clear.
• A log makes it reversible: if someone can’t find a file, you can trace it.

AI file organization: the system architecture

1. Trigger: new email arrives to an “AI inbox” (or a shared mailbox).
2. Extract: pull identifiers from the email + attachments.
3. Lookup: search the transaction ledger for likely matches.
4. Score: compute a confidence score based on how well it matches.
5. Decide:
• If score ≥ threshold: rename + move file in Drive.
• If score < threshold: send to human review with “Approve / Reject” options.
6. Log everything: append a row to an audit log (even for “no action taken”).

Step 1: Build the transaction ledger (Google Sheets)

Suggested schema (columns)

• Record ID (unique)
• Created time
• Lead / Transaction ID (from your CRM)
• Client name
• Client emails (comma-separated or JSON list)
• Assigned agent (name + email)
• Shared drive ID (or shared drive name)
• Target folder ID (the folder where docs should land)
• Transaction type (buyer / seller / other)
• Property address (if applicable)
• Status (appointment set / under contract / closed)
• External transaction inbox / routing email (if you use one)
• Notes

Rules

• Humans can view the ledger, but only the automation should write to the “system” tabs.
• Keep “configuration” (templates, mappings, thresholds) separate from “events/logs” to reduce accidental edits.

Step 2: Extract identifiers from inbound emails

• From / To / CC emails
• Subject line
• Body text (plain text is fine)
• Attachment filenames
• Attachment text (PDF/OCR if needed)

What to extract (examples)

• Client email(s) (match against ledger “Client emails”)
• Property address (even partial)
• Agent name/email
• Transaction / deal ID (if included)
• “System addresses” (for example, transaction-specific routing emails)

Step 3: Matching logic (deterministic first, AI second)

A) Exact matches (highest confidence)

• Any inbound email address exactly equals one of the ledger’s Client emails
• A known transaction routing email is present
• A CRM lead/transaction ID is present

B) Strong fuzzy matches (medium confidence)

• Property address matches after normalization (remove punctuation, normalize abbreviations)
• Client name matches + at least one other signal (agent, city/zip, etc.)

C) Weak matches (low confidence)

• Only a name match with no other signals
• Only a partial address match

Normalization tips

• Normalize addresses into tokens (street number, street name, suffix, unit, zip)
• Normalize emails (lowercase)
• Strip formatting from phone numbers if present

Step 4: Confidence scoring + thresholds

• Exact client email match: +60
• Exact transaction routing email match: +60
• Address token match (street number + street name): +30
• Zip match: +15
• Agent match: +10
• Attachment content contains address or client name: +10

Suggested thresholds

• 95–100: auto-rename + auto-move
• 80–94: queue for human review (recommended)
• < 80: do not file automatically; request clarification

Step 5: Human review queue (the “safety valve”)

• Original attachment name
• Proposed new name
• Proposed destination folder
• Top 3 ledger matches with confidence
• Approve / Reject buttons (or a simple “reply YES/NO” pattern)

What to do on rejection

• Log the rejection
• Store the file in a “Needs triage” folder
• Optionally: ask the reviewer to pick the correct transaction

Step 6: Drive renaming + filing (safe operations)

• Copy or move?
• Start with copy during testing to reduce risk.
• Switch to move once your audit trail and approvals are reliable.
• Rename convention (example)
• YYYY-MM-DD — {Client Last Name} — {Address/Unit} — {Document Type}
• Always keep the original file ID (Drive file IDs are stable even when names change).

Step 7: Audit trail (non-negotiable)

• Timestamp
• Email message ID
• File ID
• Original filename
• New filename
• Source folder ID
• Destination folder ID
• Matching record ID (ledger)
• Confidence score
• Decision (auto / approved / rejected)
• Approver (if any)
• Error details (if any)

Error handling checklist

Rollout plan (recommended)

1. Week 1: “suggest-only mode” (no renames/moves; just proposals + logging)
2. Week 2: auto for ≥ 95, review for 80–94
3. Week 3+: expand extraction rules (document types, deeper PDF parsing)

Get help building your Google Drive AI filing system