Three decisions we made that most AI tools don't

Most AI CS tools let the LLM decide what to do and then suggest an action. giiHelpdeskAgent does the opposite: your rules run first, and the LLM only drafts the reply - it never decides the action.

Here's why this matters: an LLM can be confident and wrong. It has no memory of your last 50 refund decisions. It doesn't know that you never approve refunds on orders that have already shipped. It doesn't know that a certain SKU has a known issue and should always be escalated to your supplier team.

Your rules do know all of this. So we encode them in code, where they can't be misremembered, overridden by a tired agent, or hallucinated away.

This is an architectural decision, not just a policy. We designed the system so that storing email content is technically impossible - not just prohibited.

Every time a Gmail thread is analyzed, the content is processed in memory, used to generate the analysis and draft, and then discarded. Nothing is written to our database. Our audit logs record hashed identifiers (thread ID hash, order ID) - never the email text itself.

Why did we build it this way when storing emails would give us more training data and better features? Because your customers' emails are not yours to give us. And because any system that stores sensitive communications is a system waiting to have a data breach.

giiHelpdeskAgent never cancels an order, processes a refund, or sends an email without a human explicitly clicking "Confirm." Not on simple cases. Not on high-confidence matches. Always a human in the loop.

We built several prototypes of "fully automatic" flows. They worked most of the time. "Most of the time" is not good enough when the downside is a cancelled order that shouldn't have been cancelled, or a refund that's now been issued and can't be reversed.

The speed we deliver - 90 seconds from email open to action complete - comes from eliminating the manual research and decision steps, not from removing the human. Your agent still makes the call. They just make it with full information, in 6 seconds instead of 8 minutes.