Giani Brussich

I let an AI answer my customers. Here are the six times it broke.

November 2, 2026

I sell a personalized children's book. Every copy gets printed on demand with the kid's name inside. Every few days a real parent emails me. "Can you fix a typo?" "The godmother's name is wrong." "Where's my book?"

Those emails used to sit until I got home from my day job. That bugged the hell out of me. A parent buying a baptism gift shouldn't wait on my commute.

So I built an agent to answer them. It works now, but it broke six different ways on real orders first, and I learned more from those than from anything that went right.

What it does

Every email to the store gets read. The agent figures out what the parent wants, pulls up their actual order, and writes a reply using only facts it can prove. If money or a promise is involved, it waits for me.

How it's built

Think of it as three people at a front desk, cheapest one first.

The bouncer (free). Every 10 minutes a script reads the inbox. It tosses the junk: promos, app invites, duplicates. It can look but it can't touch.

The sorter (free). A small AI model running on my own laptop sorts what's left into "needs Giani," "routine," or "ignore." Customer emails never leave my machine at this step.

The writer (pennies). Only the emails that need a real reply go to Claude. It gets one shot, no tools, and about a page of context.

Before the writer says a word, a script builds a case file. What they paid, where the print job stands, tracking, and every email we already sent them. The AI writes from that file, not from its imagination.

Then plain old code checks the draft. It blocks vendor names, blame, promises, and any date or dollar amount that isn't in the case file. If nobody touches a draft for 2 hours and every check passes, it sends itself. Anything about a refund, a free reprint, or a call only I can make waits for me.

The scoreboard so far

These are small numbers, but every one is a real parent, and each failure below cost me money or trust on an actual order.

What broke

1. One order printed twice. The payment system pinged me twice, and both pings started a print job. That's about $21 gone once the presses roll, and charging a parent twice would've been a lot worse. Now the first ping writes down "I got this one" before doing anything slow. The second ping sees the note and stops. A daily check also cancels any print job beyond what someone paid for. Now any change that touches money gets tested by sending the same payment twice on purpose.

2. A name change missed the printer. A parent replied with the right godparent name. The email showed up. The book printed anyway, and the reprint cost $20. Now when a parent asks for a change, the agent pulls the print job while it's still free and holds the order. Parents are told they have 24 hours to make changes, but the printer waits 30, so the parent who replies at hour 23 still makes it.

3. Reminder emails never went out, and every check said green. For 20 hours the job ran, finished clean, and sent nothing. Three problems were stacked on top of each other. The email service's firewall didn't like how my script introduced itself. The key I used could only send, so my health check couldn't see anything. And the store's domain had never been set up with the email service at all. I never sent one real test email, so nobody noticed until a real parent's cart sat there waiting. Now nothing that emails customers counts as live until I've sent one real email and watched it land.

4. My own test orders looked like real customers. I test on the live checkout on purpose. Two scripts each had their own copy of the "is this a test?" rule, and they quietly drifted apart. So my test carts got reminder emails like real ones. Now there's one copy of the rule, used by both, and a check that fails loudly if a test order ever gets tracked.

5. The writer tried to post in a chat app. I gave the writer one turn. It had extra tools loaded, and it spent its one turn trying to reply in a chat app instead of writing the draft. I laugh about it now. That day I was pissed. Now the writer runs with zero tools, and that's written into the command itself, so there's nothing for it to press.

6. Replies went out from my personal email. The email tool I use day to day is signed in as me, not the store. Three customer replies went out from the wrong address in one day. Now all store mail goes through one script that can only send as the store, and a gate blocks my personal account any time the message mentions the store.

What I took away

The drafts were good the whole time. Every break happened in the plumbing around the AI: a payment that pinged twice, a change that beat the printer by a few hours, a send path I never tested, a rule that lived in two places, a writer with tools it didn't need, and mail from the wrong account.

That plumbing is most of the job, and it never shows up in a demo.

What's next

I'm pumped about where this is going. Trust the log, not the summary.