8
AI
AI agent vs traditional workflow for invoice processing
Intelligence ↔ Predictability
Context
Automate processing of ~40K supplier invoices per month.
We receive invoices in every format imaginable. A vendor demoed an autonomous agent that reads the invoice, matches the PO, flags discrepancies and posts to the ERP.
Our finance controller wants every step auditable. My instinct: deterministic workflow (state machine) with the LLM used only for extraction, plus validation rules and human review below a confidence threshold.
Where are teams drawing the line between "agent" and "workflow with an LLM step"?
Constraints
- Accuracy
- > 99% on totals
- Audit
- Every step explainable
LLM agentvsDeterministic workflow + LLM extraction
Community verdict
LLM agent 25%Deterministic workflow + LLM extraction 75%
8 engineers · 3 opinions
With these constraints, what would you choose?
One choice per engineer. You can change it any time.
LLM agent
Deterministic workflow + LLM extraction
Predictability →
Trade-offs
Dimensions
LLM agentDeterministic workflow + LLM extractionout of 5
- Adaptability
- 5LLM agent scores 5 of 53Deterministic workflow + LLM extraction scores 3 of 5
- Predictability
- 2LLM agent scores 2 of 55Deterministic workflow + LLM extraction scores 5 of 5
- Auditability
Community discussion
3 comments
Have you made this decision in production? Share your reasoning.
Sign in to commentUse the LLM where the input is unstructured (extraction) and code where the rules are known (3-way matching, tolerances, posting). You get the best of both, and an auditor can read the state machine.
Also much easier to threat-model. An agent with ERP write access plus prompt injection hidden in a PDF is a scary combination.
Agents earn their keep in the long tail: credit notes, split deliveries, currency mismatches. Maybe route only the exceptions to an agent, with a human approving its proposal.