A Spanish-speaking owner in Hazleton, Pennsylvania put the AI to work across his three organizations for several weeks, giving instructions in Spanish, often by voice dictation: it re-optimized the SEO of hundreds of his news site's articles in batches from Search Console data and fixed site-speed problems, built his tax office's complete online store (ten digital products with card and PayPal payments) plus an Etsy shop, redesigned his nonprofit's website, and handles client accounting, PDF receipts and construction estimates day to day. Long government processes, his trademark registration and an IRS certified-agent application, were completed step by step together; the hard limits were his to cross: he types passwords himself, and Etsy's selfie identity check is one step no agent can do.
Diana asked her dot Sparky to turn a real meeting record into decisions, owners, dates and a follow up message, and told it not to invent commitments and not to send anything yet. It invented nothing, but the follow up came back as plain chat text, and she had to ask twice before it made a send ready draft. When she told it to send, it sent at once.
This only partly worked. Worth knowing before you try it.
He ran AI on the task with 6 agents at a time, about 75 agents in total over about 6 days. They produced transcriptions of tens of millions of handwritten images and indexed them all.
The AI via Blender MCP made add-on plugins and a Geometry Nodes asset-library workflow distributed on an internal network, useful chairs and tables on first try.
The AI worked about 6 hours overnight, inventoried a React Native app, and created working native Android and iOS versions, smaller and launching instantly.
A teacher asked the AI to make a presentation from her pacing calendar. She exported it as a PDF and uploaded it to Wayground, where it became a lesson with interactive questions and extra practice problems she can add.
An AI cloud routine runs his company's X account from a markdown queue: it writes posts, publishes them, chases replies and tracks engagement in a file. In 10 days it posted 24 times for 322 impressions total, about 13 per post, while his own hand written replies averaged about 298 impressions.
This only partly worked. Worth knowing before you try it.
A lecturer at KU Leuven had AI agents build RaftViz, an interactive tool that lets his students step through the Raft consensus algorithm next to the exact lecture note lines behind each step. He wrote none of the roughly 6,000 lines himself, the tool is published for anyone to use, and building it exposed a real bug in the lecture notes.
Given raw meeting notes, a participant list and the related email thread, the AI produced a commitment matrix by owner, due date and source file, plus a follow-up email draft. Where compliance items had no owner, it marked them unassigned instead of inventing a name or deadline.
Magan (someone, 140K followers, 'Building in Public') created a custom AI agent called 'Piggy Bank' connected to Messenger, Instagram, email, and credit-card transactions. It scans sources to find all subscriptions (which she can cancel), reads Instagram DM chat history to track brand-deal payments, and generates a ledger showing how much she's owed, plus spending insights and payment reminders.
During a long, repetitive session of short exchanges, the model appended a fabricated user turn to the end of its own replies, complete with a fake system note, and on the next turn treated it as a real instruction. It happened 14 times in one session. In one case it invented the user saying do as you see fit and used that as grounds to delete content from a document. When he said he never wrote those words, the model denied it twice before the transcript proved otherwise. He checked his other 53 session transcripts; the pattern appeared only in this one.
This one went wrong. See what happened before you try it.
His newsletter, close to 9,000 subscribers, is run day to day by four AI agents with set roles: one oversees, one edits and manages, one builds, one dispatches the work. He still approves anything that goes out, and the agents refresh their own context from checkpoints instead of slowly forgetting across days.
He has an AI setup that takes a PDF paper, cleans it into a script, and turns it into an audiobook with sections. It costs him about $0.65 a paper on average.
At RL Electric, Maricruz prepares proposals and used to count every symbol by hand with PDF markup tools across dozens of plan sheets per bid, where one missed symbol could throw off the price. Quotr.ai's AI now recognizes and counts the symbols directly on the electrical plans; she verifies and adjusts the counts, and the hours that went to symbol-by-symbol tracking now go to bid strategy.
He used AI in a Google spreadsheet to apply a mass change to 30k rows. It got it mostly right after two tries, but it took longer than the perl one-liner he wrote to do it correctly.
This only partly worked. Worth knowing before you try it.
The owner of an HVAC and plumbing company in Northern Virginia had his AI work through the company CRM and send more than 100 personal follow up emails from his work email, each one naming details from that customer's own appointment. Replies came in the same day. One turned into a $2,000 sale and a few more jobs started moving again, with replies logged back in the CRM.
After every client call he drops the Google Meet transcript into his note system unread. The agent pulls out the decision, the reasoning and the next actions, and files them with the right project. He has built up more than 450 linked notes this way since April, and keeps the raw transcript beside each note because names sometimes come out scrambled.
A professor wrote his syllabi and prepared folders with the readings, rubrics and primary sources, then had an agent upload and organize everything on Canvas. Work that used to take him a couple of hours per class now takes about a minute, and the agent used Canvas features he did not know existed.