Background
Dataforge built AI Enterprise as a Job Control System (JCS)—tracked jobs with review, integrations, and a full run history—not a free-form chatbot.
Before putting the product in front of beta customers, we put it on our own accounts payable work: vendor invoices and general expenses that previously needed manual reading, vendor checks, currency handling, and voucher entry in QuickBooks Online.
Business Challenge
Invoice and expense documents are not uniform. Some arrive missing required fields. Vendors may be new. Currency may be USD or CAD. Classification for expenses needs to stay consistent over time.
A template-only OCR path breaks down on that mix. An open-ended AI agent that calls tools back and forth is also a poor fit: cost rises, behaviour is hard to audit, and accounting still needs clear checkpoints when policy requires a person.
We needed a controlled workflow that could post clean voucher data into QuickBooks Online from email, pause when a vendor was unknown, and keep standing guidance without rewriting code for every vendor quirk.
Solution
AI Enterprise ingests invoices and expenses from Office 365 / Exchange email. Each document runs as a tracked job against QuickBooks Online.
The workflow checks the vendor against our approved-vendor database. If the vendor is missing, an interaction alerts accounting. After confirmation, staff can ask the system to create the vendor file; it fills public vendor details it can find and records a short note on what the vendor does for later expense guidance.
Currency is detected (CAD or USD) with conversion when needed. Jobs confirm checks before posting; accounting reviews afterward. When the document or policy requires it, the job pauses for operator input first.
Operator notes act as standing policy. Accounting can change a note on a vendor; later jobs follow it going forward without relying on someone to remember. Credentials for connected systems are encrypted at rest and are not retrievable by portal users.
Implementation
IT owns most workflow tuning. When accounting sees something off, we run diagnostics aimed at general behaviour—not one-off data fixes—and adjust the path. Accounting has begun writing operator notes directly; that has worked well for vendor-specific guidance.
Agents inside the workflow are highly limited. Many steps do not hold open tool access: they process data, create a draft, and run that draft through a validator. That avoids back-and-forth tool loops, keeps behaviour constrained, and holds cost to about 5 to 10 cents CAD per invoice in our current mix.
Volume is about 100 documents a month across vendor invoices and general expenses. Document quality varies widely; the same job model handles that mix better than a fixed template system.
Project Considerations
Posting happens after confirmation checks, with human review afterward. When the document or policy requires it, the job pauses for operator input before continuing.
Time saved is modest at this volume—about three hours a week. The clearer gains are voucher quality, more consistent expense classification, and not spending those hours on data entry.
Results above 90% automation apply since the latest workflow version and will vary with document mix.
Environment Comparison
| Previous Environment | Current Environment |
|---|---|
| Manual read of email invoices and expenses; entering data into QuickBooks Online | Tracked AI Enterprise jobs from Office 365 / Exchange into QuickBooks Online |
| Approved-vendor checks and new-vendor setup handled ad hoc by staff | Approved-vendor check in-flow; interaction alert; optional vendor create after confirmation |
| Currency and classification depended on whoever entered the voucher | CAD/USD detection with conversion; cleaner voucher data and more consistent expense classification |
| Vendor quirks lived in people's heads or scattered notes | Operator notes scoped to vendors; changes apply to later jobs as standing policy |
| Template OCR fragile on messy documents; open agents expensive and hard to audit | Controlled workflow with draft-then-validate steps; ~5–10¢ CAD per invoice |
Key Outcomes
- Over 90% automation — vendor invoice volume on the JCS path with little or no entering data by hand since the latest workflow version
- 5–10¢ CAD per invoice — draft-then-validate steps keep tool use limited and run cost low
- About 3 hours/week back — modest at ~100 docs/month; larger win is voucher quality and classification
- Standing operator notes — accounting changes a vendor note once; later jobs follow it without relying on memory
Long-Term Track Record
The path remains in daily production at Dataforge. Operator notes continue to accumulate as accounting encodes vendor and expense guidance once instead of re-teaching the same exceptions.
Cost per invoice has stayed in the 5 to 10 cent CAD range under the draft-then-validate design. We use the same paper trail—status, history, cost, and interactions—that customers see when they run jobs in the portal.