01The challenge
An enterprise was coding thousands of vendor invoices by hand: assigning GL accounts, matching POs and routing approvals. The work was slow, inconsistent from one analyst to the next, and hard to audit. It also delayed the month end close.
02Our approach
We did not bolt AI onto the old workflow; we redesigned it. Instead of every invoice passing through a person, a model makes each routine decision and people handle only the exceptions.
- Decision models, not chatbots. We fine tuned small, open weight Google Gemma models with LoRA adapters to make bounded decisions: which GL account, which PO match, which approver. Each model chooses only from valid options, so it never invents an account that does not exist.
- Calibrated confidence. Every decision comes with a reliable confidence score. High confidence invoices post straight to the ERP. Complex or uncertain ones escalate to Jev for deeper reasoning, and anything still unresolved goes to an analyst, along with the model's reasoning.
- Governed by design. Every decision, score and override is logged for audit. Analyst corrections feed a retraining loop, and each new model must pass a benchmark before release.
- Runs in the client's cloud. The models are small enough to run on a single GPU inside the client's own Google Cloud (GCP) environment, so financial data never leaves their tenant.
03The outcome
Most invoices now flow straight through without a person touching them. Analysts spend their time on real exceptions instead of routine coding, and leadership gets a complete audit trail of every decision.
04Why it matters
Most enterprise decisions do not need a large, expensive general purpose model. They need a fast, accurate, auditable choice. We help teams pick the right model for each step, whether that is a small decision model, an agent or a person, and then govern it in production.