AI at work: what AI agents actually do
Stories from my job: analytics, customer support, design and development.
Each one shows what I asked, what the AI agents did, and the numbers that came out.
16 stories, newest first
Keilu's Warehouse
Ideas, documents and decisions you worked on with AI get buried inside chats. For me, everything goes into one warehouse outside the AI tool, Linear in my case, and I just tell my shaliach "save it to the warehouse for me". That way it's clear where everything is, and tomorrow you can pick up right where you left off, even with a different tool.
Software as (small) gifts
A keyboard-language mistake became a gift idea: TypoSwap, a ready-to-paste correction and an installer you unwrap.
Breaking the (context) window
I asked to turn three office rooms into one shared space. With a fleet of shluchim each, I want us to talk more easily to the person beside us.
Wake up, Sky.
My son calls ChatGPT voice mode "Sky." I spent two and a half hours talking to it at work, with more than four PRs an hour. Now I want a sense of humor too.
Wait, what exactly is Grok Bot? A beginner's guide
A practical, personal guide to Grok Bot: what it is, how to start, what it costs, and what you can do from your phone, with a first task and safety tips.
How to go through social media and pull out pearls.
Sosia goes into my Facebook reading groups, reads, and pulls out organized quotes and mentions. Within an hour there were 28 quotes for the campaign and 5 real mentions of e-vrit.
1.2k dead clicks on the terms checkbox, and a bug that wasn't.
Dima reproduced it, Nitza designed it, and in about one workday we went from "bug?" to a measured fix.
A whole month of tickets, in seven minutes.
A monthly summary of 9,001 support tickets against our help center: 16 draft FAQs and two real production bugs.
From idea to engineering in two days: a spec-driven design process for the ticket bundles page.
Nitza, my AI designer, researched and built a repeatable design process, and its first run took our ticket bundles page all the way to a developer handoff.
Bar2Go pickup points at checkout, with one small change.
Kochava, my AI developer, first researched the checkout's shipping step, and only then added a link to the pickup points map.
A Linear brief before the weekly sync.
A routine that writes a Hebrew brief from Linear every Monday at 10:00, and caught 7 stuck tasks on its very first run.
What Clerk will cost, without guessing from login events.
Pricing Clerk by monthly retained users, including our reader app, showed our first estimate was an order of magnitude too low.
A weekly Clarity digest that stays quiet when there's nothing to say.
Every Monday my Clarity AI agent reports rage clicks, dead clicks and problem session recordings, but only when something actually matters.
Four questions from Chartbeat, and answers with real numbers.
One sentence to Basher during a meeting, verified traffic data pulled from GA4, and two days later I was on Zoom with the vendor.
One check-in sparked an FAQ for the help center.
Cross-checking our help center against real Glassix support tickets produced 8 draft Q&As in Hebrew, with nothing published without my approval.
From my phone, at the beach: a GA4 mystery dressed up as a campaign.
A fake traffic source in GA4 was logging 20,000 sessions a week. My AI agents tracked it down and fixed it while I stayed on the beach.