LLMs are powering the next leap in parsing

Better parsing means sharper detection, and as LLM technology advances, so do we. Over the past few months, we’ve rolled out upgrades that significantly improve parsing accuracy across a wide range of document types.

Broader parsing coverage across documents

In our last update, we shared how LLMs boosted bank statement parsing. Now, we’ve expanded that progress to invoices, utility bills, tax forms and IDs, while also improving parsing for business filings, financial statements, cheques, investment statements, leases, credit card statements, benefits statements, social security cards, and more. These improvements strengthen the extraction of key fields like names, dates, and addresses — the very details fraudsters are most likely to manipulate.

Invoice parsing

Invoices now benefit from significant leaps in accuracy:

• Name coverage: 82% → 92%
• Address coverage: 67% → 86%
• Date coverage: 56% → 90%

Over 98% of invoices now return parsed details, with new fields like seller name (96%), total invoice amount (92%), and seller address (82%) captured at scale.

Utility bill parsing

Utility bill parsing is now live for all customers, with performance gains across key fields:

• Name coverage: 88% → 97%
• Address coverage: 75% → 97%
• Date coverage: 77% → 93%

We’re also capturing richer context like service provider name (92%), account number (92%), provider URL (83%), and amount due (73%).

Tax form parsing

Tax form parsing has expanded to new subtypes and achieved step-change improvements:

• W2 EIN values: 44% → 99%
• 1040 total income: 43% → 96%
• Schedule C total other expenses: 22% → 96%
• 1120 corporation name: 27% → 96%
• 1120-S corporation address: 4% → 97%
• 1065 ordinary business income: 26% → 99%

ID parsing

ID parsing now delivers stronger accuracy across the fields most critical for identity verification:

• Name coverage: 73% → 89%
• Address coverage: 50% → 68%
• Issue date coverage: 50% → 86%

Alongside the issue date, we have also improved our parsing of date of birth (84%) and expiry date (73%). Beyond these core fields, we now extract the document’s country code (93%) and number (92%).

What this means for fraud detection

Parsing improvements might not grab headlines like deepfake scams or synthetic fraud, but they’re the bedrock of effective fraud detection. When institutions can trust the accuracy of extracted names, addresses, and dates across invoices, utility bills, identity documents and tax forms, they can catch fraud signals earlier and with more confidence. Additionally, these improvements in name and address parsing assures that we have a more reliable foundation for validating identities.

With these updates, Inscribe customers benefit from stronger detection, broader coverage, and faster insights, ensuring fraudsters have fewer places to hide.