Why we put a human-review layer between Claude and QuickBooks

Claude can read a PO, find the customer in QuickBooks, match the lines against the catalog, check the prices, and create an Estimate. We let Claude do all of that. We do not let Claude press Send without a human in the loop.

QuickBooks is the source of truth for what gets shipped and what gets billed. If an Estimate has the wrong customer, the wrong customer gets billed when it converts to an Invoice. If an Estimate has a wrong SKU, the wrong product gets pulled from the warehouse. If an Estimate has a wrong price, the customer gets a price they didn't agree to. The cost of a bad QB record is shipping cost, accounting cleanup, and possibly a lost account.

Compare that to the cost of an unsent draft. Marcia opens it, sees that Claude mapped DM19012 to "Services" instead of the right SKU, picks the right SKU, hits Send. Twenty seconds of friction. No customer impact. No accounting cleanup.

The two costs are not comparable. One is annoying. The other is expensive.

So we built SideQuest so that Claude does all the work up to but not including the submit. The draft sits in QuickBooks, or in SideQuest's local draft store if Claude is uncertain enough that even drafting it in QB would be wrong. A rep reviews. The rep can override anything: the customer, the line items, the prices, the discount, the memo. The rep hits Send when they're satisfied.

This matters more than the design feels like it should. Distributors we talked to who'd tried other automation tools all had the same horror story: an automation pushed something live, the thing was wrong, the customer was confused, the team didn't realize for two days because they trusted the automation. After that, they stopped trusting the automation entirely and went back to typing.

The human-review layer is the trust mechanism. Marcia trusts SideQuest because Marcia is still the one pressing Send. SideQuest didn't take her job. It took the tedious 90% of her job. The 10% she keeps is the part that requires judgment.

A side benefit: when Marcia overrides Claude's mapping, that override is a training signal. SideQuest logs it and updates the per-customer cross-reference table. Next time the same SKU comes in from the same customer, Claude maps it the way Marcia mapped it. The system gets smarter with use, and the smarts come from Marcia rather than from a black-box model.

A system that pressed Send by itself would be faster. It would also be the thing Marcia turns off after the first wrong invoice. We optimized for the system Marcia wants to keep using. That meant building the layer that puts her in charge of the only step that matters.