What is the difference between rule-based automation and AI agents?

Many small accounting firms rely on practice management software like Karbon or Canopy. These tools are excellent for creating task templates and setting date-based reminders. They can trigger a task when a project status changes. However, their automation is rule-based and breaks down when faced with unstructured client data, which is most of the firm's daily work.

Consider the most common workflow: a client emails a PDF of a brokerage statement or a K-1. A rule-based system cannot open that PDF, identify the document type, find the capital gains figure, and enter it into the tax software. This forces a human staff member to perform a 15-minute manual process: download the PDF, identify it, open the tax prep software, find the right client and form, manually type in the numbers, then update the task in the practice management system. This process is slow, expensive, and prone to data entry errors that can lead to incorrect filings.

Firms might try to patch this with document parsers, but these are often just another form of rule-based system. They are trained on a specific template. When a brokerage firm changes its statement layout, the parser fails. It cannot adapt. The core problem is that these tools lack cognitive ability. They cannot see, read, and understand a document the way a human can. They are designed for structured data and predictable sequences, but accounting workflows are messy and unpredictable.

The first step is a process audit. Syntora would map your end-to-end document intake workflow, from the moment a client email arrives to the final data entry in your tax software. We would analyze 50-100 real document examples to understand the full range of formats and data points you handle. This audit produces a clear data schema and a precise plan for the agent system.

Syntora builds multi-agent platforms using a custom orchestrator called Oden. When an email arrives, a Triage Agent uses Gemini Flash function-calling to classify the attached document. Oden then routes the task to a specialized Extraction Agent that uses the Claude API's tool_use capabilities to read the PDF and extract key data points into a structured format. This multi-agent design, built with Python and FastAPI, allows each agent to be an expert at its one job. We use Supabase for persistence, creating an immutable audit log of every action the agents take.

The delivered system is deployed on a DigitalOcean App Platform instance that you control. It connects to your email via a webhook. When a document is processed, the system presents the extracted data in a simple review interface for one-click approval by your staff. This human-in-the-loop step ensures 100% accuracy before data is committed. The full source code and a maintenance runbook are provided, giving you complete ownership.