Laiye Unveils ADP Self-Optimization Agent: AI Learns From Every Correction

International trade feels this pain.:

An Italian supplier's invoice shows "€ 1,250.05" The OCR reads "1250" — decimal wrong, amount wrong. Japanese receipts follow unfamiliar layouts; the model confuses tax lines with totals. Southeast Asian invoices mix per-line tax rates with lump-sum taxation, and your system can't allocate discounts automatically.

Reviewers fix the same errors, invoice after invoice. And the system never learns — until now. Invoice/Receipt Agent: no training, no configuration. It starts evolving on its own.

Laiye ADP (Agentic Document Processing) launches the Self-optimization. Free for a limited time.

Why invoice OCR always falls just short

Most teams using invoice OCR have experienced the same cycle. The system's aggregate accuracy looks acceptable. But reviewers still spend hours daily correcting the same edge cases — format mismatches, misplaced fields, folded documents, multilingual layouts.

The root cause is clear: a general-purpose model has seen millions of invoices, but it hasn't seen yours. It doesn't know your suppliers' formatting habits, the countries your employees travel to, or that one Japanese vendor always tucks their tax ID in a footnote in unusually small type.

Teams used to have two options:

• Use a generic model – works immediately, zero setup. You accept the reality it's almost good enough.
• Build a custom model – weeks of data labeling and training, plus costing several person‑months.

Neither was satisfying. That’s why we built a third path: start with a generic model, and let it evolve into one that knows your documents.

How Self‑Optimization Agent works

The Self‑Optimization Agent is available now inside Laiye ADP’s out‑of‑the‑box: Invoice/Receipt.

‍

Your reviewers do exactly what they already do: validate and correct extraction results. No new interface. No extra steps.

What changes is what happens behind the scenes.

Every correction is recorded silently. Once the system sees enough examples, it identifies repeatable error patterns and automatically generates an optimization.

But the optimization is not applied automatically. First, it goes through blind validation using real documents:

• The target field must improve.
• No other field can get worse.

The system only presents a recommendation card after validation passes. You decide whether to apply it.

Every optimization is explainable and reversible. We prioritize stability over speed — because in production, that’s what matters.

Privacy is built in: Learning stays inside your instance. No data crosses to other tenants. One platform, a hundred customers, a hundred different configurations — each shaped by its own documents.

Real example: from endless corrections to one‑time learning

A multinational consulting firm runs a shared service center for European travel expenses. Employees submit receipts from Italian restaurants, German hotels, French supermarkets.

Italian receipts were a persistent problem. European formatting uses commas as decimal separators (€45,80). The generic model kept reading “4580”. Reviewers fixed it manually. Repeatedly.

They turned on Self-Optimization. No change to their daily work. Over time, the system recognized the pattern, generated an optimization, passed blind validation, and pushed a recommendation to the finance manager. One click to apply.

One week after applying the optimization:

The system started remembering its mistakes and learning not to repeat them.

Accounts Payable: learn once, apply across hundreds of suppliers

Accounts Payable (AP) invoice processing differs from expense reimbursement. Suppliers are stable, but volumes are far higher. Each supplier has its own formatting signature: PO numbers with inconsistent prefixes, tax IDs hidden in footnotes, Japanese vendors using qualified invoice layouts that diverge from templates.

No single fix is hard. The pain is cumulative — reviewers correct the same supplier's same errors, month after month. With Self-Optimization Agent, the system gradually internalizes each supplier's format. Once recognition converges for a supplier, subsequent invoices of the same type require minimal or no manual intervention.

Learn once. Apply forever.

For enterprises processing invoices from hundreds of vendors monthly, this compounding effect becomes substantial.

Why we keep Human in the Loop

We considered auto-applying optimizations immediately after validation. We decided against it.

Invoice recognition feeds directly into payment and reconciliation. A single misapplied optimization could trigger cascading issues — incorrect payments, reconciliation mismatches, and downstream financial adjustments.

Our design principle: AI learns and analyzes. Humans make the final call. The Agent gets smarter about your business every day. You retain final approval.

Try it free

Get Started — Free for a Limited Time

Self-Optimization Agent is available now at no cost during the introductory period. No additional configuration is needed. No impact on existing workflows.

How to enable?

• Enter the Out‑of‑the‑Box – Invoice/Receipt agent
• Review and correct extraction results as you normally do.
• After you have corrected 5 documents, Self-Optimization becomes available.

This is your choice

• Not ready? The feature is off by default. You won't notice any change.
• Want to try? Turn it on. Learning starts immediately. You remain in control.
• Changed your mind? All correction history and past optimizations are preserved. You can re-enable it anytime.

Frequently Asked Questions (FAQ)

Q1: Does Self-Optimization Agent require retraining my team?

No. Reviewers continue their existing correction workflow. Learning happens silently in the background.

Q2: How long before I see improvement?

It depends on correction volume. The system needs enough samples to identify reliable patterns (The feature becomes available after just 5 corrected documents.)

Q3: Can an optimization be rolled back?

Yes. Every optimization is reversible. The full correction history is preserved.

Q4: Does the system share data across different companies?

No. All learning is instant-isolated. Your corrections train only your instance.