I recently completed a customer project that synchronizes expense records from a CRM into QuickBooks Online using Make. The workflow polls the CRM for new expenses, extracts receipt information using AI, creates the corresponding purchase in QuickBooks, and updates the source record once processing is complete.
The most challenging part of the project wasn’t extracting the receipt data, it was determining whether the vendor already existed in QuickBooks.
For example, a receipt might contain:
ABC INDUSTRIAL SUPPLY - EASTERN REGION
While the existing QuickBooks vendor might be stored as:
- ABC Industrial Supply
- ABC Industrial Supply LLC
- ABC Industrial Supply - East
- ABC Industrial
From a human perspective, these are obviously the same business. For an automation, however, that matching process becomes surprisingly fragile.
The traditional approach usually involves building numerous rules to normalize the text, removing punctuation, ignoring capitalization, stripping “LLC” or “Inc.”, handling line breaks, performing partial matches, and maintaining an ever-growing list of exceptions. Every new vendor naming convention introduces another edge case to account for.
For this project, I decided to experiment with Make AI Agents.
Instead of relying entirely on string manipulation, I first retrieved the customer’s vendor list from QuickBooks and passed both the extracted vendor name and the existing vendor records to an AI Agent. The agent’s task was straightforward: determine whether an existing vendor was the closest match or indicate that a new vendor should be created.
The resulting workflow became much simpler.
Receipt
↓
AI Content Extractor
↓
Search QuickBooks Vendors
↓
Make AI Agent
↓
Vendor Found?
↙ ↘
Yes No
Create Expense Create Vendor → Create Expense
What stood out to me wasn’t the AI’s ability to read the receipt, modern OCR solutions already do an excellent job with that. The real value was allowing the workflow to reason through an ambiguous decision that would otherwise require dozens of brittle matching rules.
This was my first production implementation using Make AI Agents, and it changed how I think about workflow design. Rather than trying to anticipate every possible vendor naming variation with filters and text functions, I was able to delegate that reasoning to AI while keeping the rest of the automation deterministic and easy to maintain.
I think this is where AI Agents provide the most value, not by replacing traditional automation, but by solving the small, ambiguous decisions that are difficult to express through conventional business rules.
