Build a Custom AI Tax Preparation System You Own

Many accounting firms try to solve the document intake problem with generic tools. They use a document management system with a built-in OCR feature or an off-the-shelf data extraction tool. These tools fail because they lack the specific context of tax preparation. They can read a number, but they do not know if it is from Box 1 of a W-2 or Box 1a of a 1099-DIV. This generic approach creates more review work than it saves.

For example, a 10-person firm uses a popular document management portal where clients upload their tax documents. The portal's OCR feature extracts the text from a PDF of a consolidated 1099 from a brokerage. It dumps all the text into a flat file, mixing dividends, capital gains, and miscellaneous income. A staff accountant then has to manually find the right numbers in the text block and copy them to an Excel workpaper, which is the same work they were doing before.

The fundamental issue is that these systems are not architected for accounting workflows. They cannot handle form variations between tax years or different financial institutions. They provide no validation to check if capital gains distributions match the summary totals. The result is a system that feels automated but still requires 100% manual review and correction, completely defeating the purpose.

Syntora would begin an engagement by analyzing a sample of 50-100 of your firm's actual tax documents to identify common forms and their variations. This discovery phase is critical for determining the specific extraction logic required for each document type. The technical approach would leverage AWS Textract for initial Optical Character Recognition (OCR), providing high-fidelity text and positional data from scanned PDFs, even those that are skewed or low quality.

A core component of such a system involves engineering specialized prompts for the Claude API. For each identified document type, a custom prompt would instruct the model to accurately find and extract specific financial fields, such as 'Wages, tips, other compensation' from a W-2 or 'Total ordinary dividends' from a 1099-DIV. The extracted data would then be validated against expected data types and formats using Pydantic models to ensure accuracy and consistency.

The extraction logic would be wrapped in a FastAPI web service, suitable for deployment on serverless platforms like AWS Lambda, which operates with a pay-per-execution model, incurring no cost when idle. When a staff member uploads a client's PDF packet to a designated S3 bucket, a Lambda function would trigger the service. This service would process the entire packet, document by document, and post the extracted, structured data as a draft journal entry to an accounting platform via its API, such as QuickBooks Online.

For result review, a simple dashboard would be developed, potentially using Supabase. This dashboard would log each processed document, providing links to the original PDF, the structured JSON output, and its status in QuickBooks. The system would be designed to flag documents for manual review in an exception queue if the Claude API returns a confidence score below a specified threshold, such as 98% for any critical field. This design pattern enables efficient quality control, allowing a single individual to manage reviews for a high volume of documents by focusing on flagged exceptions.