Connecting Legacy Accounting to Modern Commerce

Challenge

The business ran on QuickBooks Desktop. Thousands of parts existed only inside an on-premises company file: no categories, no images, no web presence. Customers search for parts the way they talk, with misspellings and partial part numbers, and every web order had to land back in QuickBooks as a sales order without anyone typing it in twice.

Approach

• Use the integration QuickBooks Desktop actually supports. The QuickBooks Web Connector polls a SOAP service we built and exchanges qbXML. Large item queries are paged with iterators so a sync never tries to swallow the whole catalog at once.
• Let AI help, with guardrails. A locally hosted language model proposes a category, subcategory, vehicle type and tags for each item, with a confidence score. Low-confidence results are flagged for a person to review instead of being trusted blindly.
• Protect the sync from the model. Concurrency limits, adaptive timeouts, a circuit breaker and exponential back-off mean a slow or unavailable model degrades categorization, not order flow.
• Search for intent, not exact strings. Fuzzy matching handles typos and part numbers; vector search over local embeddings handles meaning.

Implementation

• A Web Connector SOAP service with session handling and qbXML models for items, inventory, customers and sales orders.
• Two-way order flow: web orders are written to QuickBooks as sales orders, matched by reference number and marked as web-originated.
• A queued, batched categorization pipeline with confidence scoring and review flags.
• Hybrid search (fuzzy + semantic) and a retrieval-augmented assistant that knows stock levels and hands off to customer service when it can't help.
• An image pipeline with validation, de-duplication by content hash and a blocklist.

Outcome

The accounting catalog became a categorized storefront that tolerates the way customers actually type ("transmision", "alernator"). Orders reach QuickBooks Desktop on the Web Connector's normal sync cycle, with no re-keying.

Technical perspective

The Web Connector is a polling protocol driven from the customer's own machine, so sessions, iterators and pending requests must survive between polls. Putting a language model anywhere near that loop meant designing for the model being slow or absent, so the business never waits on AI to take an order.