Invoices, bank statements, and SIE files read into the books

Swedish invoices, bank statements, and accounting exports in one workspace

Services: Document AI ; AI Fintech ; AI Workflow Automation
Delivery: Bookkeeping application, financial-document processing, and accounting-file parsing.

• Separate AI extraction routines for invoices and bank statements.
• PDF and image intake, bulk processing, and sorting tasks.
• A dedicated parser for Swedish SIE accounting exports.
• Client records, accounts, documents, tasks, reporting, and exports.

The input determines the processing method. A scanned invoice needs text extraction. A bank statement needs transaction rows. A Swedish SIE file already contains structured accounting data and needs a parser that understands its format.

Treating all three as a generic document summary would lose the fields the bookkeeper needs.

BookSmart accepts financial documents into client and account workflows. Its processing code includes PDF rendering, image preparation, OCR utilities, and OpenAI extraction.

The invoice prompt asks for a defined financial structure. The bank-statement routine has a separate output shape. Formatting and JSON-handling code prepare those outputs for application records.

Structured SIE files follow a dedicated parser with encoding utilities. Bulk-task and sorting services handle batches, while the frontend provides document tables, account views, notes, dashboards, and export tools.

Separate invoices from statements. The records have different fields and downstream uses. The backend has distinct extraction functions instead of one universal financial prompt.

Parse structured files directly. SIE files should retain their accounting structure rather than pass through OCR.

Keep numbers and dates intact during preparation. The text-cleaning prompt explicitly preserves invoice-related terms, numbers, and dates.

Track the processing task. Document, bulk-task, and sorting records let the application represent work that takes longer than an upload request.

Keep the file beside its record. Bookkeepers need to inspect the source when reviewing an extracted amount.

The delivered product connects document intake to client, account, and task records, with export paths for accounting work.

The available evidence establishes the processing paths and product scope. It does not include a labeled accuracy benchmark or before-and-after bookkeeping time measurement. Those should be measured separately for invoices, statements, and SIE imports rather than combined into one accuracy claim.

React 18, Vite, Redux Toolkit, Ant Design, Tailwind CSS; Node.js, Express, MongoDB; OpenAI GPT-4o extraction routines; PDF utilities, Sharp, Tesseract tooling and a Textract companion project; SIE parsing; S3, Stripe, jsPDF, ExcelJS.

Hypothetical planning example, not a measured project result or a performance forecast. All volumes and timings below are assumed.

Enter invoice fields or review extracted fields against the source document.

• Workload: 300 invoice documents per month.
• Manual handling: 6 minutes per item.
• Assisted handling, including human review: 2 minutes per item.
• Additional exception handling: 120 minutes in the same period.
• Modeled difference: 300 x (6 - 2) - 120 = 1080 minutes (18 hours) per month.

The example reserves another 120 minutes per month for unreadable or incomplete invoices. Bank statements and SIE imports are outside this calculation. This is potential staff capacity, not cash savings; setup, training, and software costs are excluded.

How to validate: Use a labeled invoice sample to track review minutes, correction minutes, failed extraction, and field-level accuracy. Count all failed documents in the sample, not just successful extractions.