Setup Guide
Estimated setup time: 30–45 minutes. If you skip optional phases (accounting routing, Gmail labels), 15–20 minutes.
Step 1 - Activate the template
- Open the Phacet template gallery
- Select "AI Inbox — Multiple Subsidiaries"
- Click Create a new project from the template
Your project is now pre-configured with all seven processing stages, column definitions, and agent instructions. No need to build anything from scratch.
Step 2 - Configure your entity catalog
This is the most important setup step. Matching accuracy depends directly on the quality of your reference data here.
- Open the subsidiary matching configuration (Stage 3)
- Delete the placeholder entities
- For each of your subsidiaries, add an entry with:
- Official company name, exactly as it appears on your documents
- Country
- Registration number (SIREN, Company No., EIN, HRB...)
- Company-specific email address or domain
- Double-check registration numbers and email domains, these are the strongest matching signals. A typo here means missed matches in production.
Step 3 (optional) - Configure accounting system routing
Skip this phase if you don't use Pennylane or an ERP with an email import feature.
This step maps each subsidiary to its dedicated supplier email address in your accounting system, so validated documents can be forwarded automatically to the right entity's books.
- Open the integration preparation configuration (Stage 6)
- For each subsidiary in your entity catalog, enter the corresponding supplier email address:
- In Pennylane: find this under Settings → Supplier Inbox for each entity
- In other ERPs: use the email import address or API endpoint your system provides
- Verify each address by sending a test email manually: a typo here means documents won't reach your accounting system
Step 4 (optional) - Configure Gmail labels
Skip this phase if you don't use Gmail or don't want automated email labeling.
This step maps each subsidiary and document type to a Gmail label, so processed emails get organized automatically in your inbox.
- In Gmail, create the labels you need - one per subsidiary, one per document type
- Retrieve the Gmail label ID for each label → [See integration guide for how to find label IDs]
- Open the integration preparation configuration (Stage 6) in Phacet
- Map each subsidiary and document type to its label ID
Step 5 - Review fraud screening settings
The fraud detection stage works out of the box. This step is a review, not a rebuild.
- Open the fraud screening configuration (Stage 4)
- Confirm the default indicators are active:
- Sender name/address inconsistency
- Urgent or threatening tone
- Suspicious links (display text ≠ actual URL)
- Unusual attachment types
- Brand or domain spoofing
- Requests for sensitive information
- Adjust only if needed - e.g., if a known supplier regularly uses urgent language, note that a single pattern signal alone is unlikely to trigger SUSPICIOUS. However, one confirmed web finding (domain flagged in threat databases) will, this is by design.
Most teams keep all defaults active and move on.
Step 6 - Run a test batch
- Upload or push your 10–20 test emails into the project
- Let the agent process the full batch
- Review results against this checklist:
- ✅ Document types classified correctly?
- ✅ Supplier names extracted accurately?
- ✅ Subsidiaries matched to the right entity?
- ✅ Legitimate emails all marked SAFE?
- ✅ "To Review" flags only on genuinely ambiguous items?
- Open 5–6 results in the Detail interface and check the citations:
- Does the agent point to the correct PDF section for supplier name?
- Does the entity match reference the right identifier from your catalog?
- If you configured Pennylane/ERP routing (Phase 3), verify the correct supplier email appears for each matched entity
- If you configured Gmail labels (Phase 4), verify the correct label IDs appear for each entity and document type
If matching accuracy is low: go back to Phase 2. The fix is almost always in the entity catalog - missing registration numbers, incomplete email domains, or name variants the agent doesn't recognize.
Step 7 - Go live
- Connect your email source via automation (see From Inbox to Action below) - or establish a manual upload routine for low-volume scenarios
- Activate downstream routing in your automation tool:
- Pennylane/ERP forwarding using the addresses configured in Phase 3
- Gmail label application using the IDs configured in Phase 4
- Team notifications for SUSPICIOUS and "To Review" items
- Assign team members to the Review queue
- Set a review cadence - twice daily is a good starting point for most teams
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Classification accuracy is low | Poor PDF quality: scanned docs with low resolution or image-only content | Ask suppliers for native PDFs. For scans, OCR quality directly affects results. |
| "Not identified" too often on subsidiaries | Weak entity catalog: missing registration numbers or shared email domains across entities | Add stronger identifiers per entity (registration numbers, unique email domains). See Phase 2. |
| Legitimate emails flagged SUSPICIOUS | Trusted supplier's domain reported in threat databases (past incident, now resolved) | Add context to Stage 4 instructions for known-good senders. Don't lower the global threshold. |
| Response drafts wrong tone | Agent mirrors sender's style by default | Set explicit tone rules in Stage 5 instructions (formality level, greeting conventions, sign-off). |
| Slow processing on large batches | Agent processes sequentially; 100+ emails at once creates a queue | Split incoming feed into smaller batches (e.g., every 15 minutes instead of once daily). |
Updated about 2 hours ago