Which AI Agents Actually Pay Off? A 30-Location CEO Shows Us

What You’ll Gain

• A list of agents a real business runs today. Phones, Facebook messages, order checks, lead routing, a weekly newsletter and a follow-up bot.
• Where the return shows up first. Time, according to an owner with 30-plus locations, and why that time turns into new projects.
• How a non-coder ships real software. Claude Code in the terminal, a five-minute starting prompt, and a bug button inside the app.
• A test for your own business. Greg Isenberg’s question: does the next model make you better or worse?
• Where AI money went this week. $7.02B, Ramp’s $1.85B, and China back after a week at zero.

Biggest Takeaway to Implement: Pick one task you dread, like a government form, a follow-up list or a pile of meeting notes, and hand it to an agent by describing the result you want, not the steps. As Randy put it: tell it the outcome you want. Don’t tell it how.

0:00 Claude Found Someone’s Keys. Then It Used Them.

Olga and Chris opened Episode 61 with guest Randy Lorensen joining the news, Greg Isenberg’s notes from San Francisco, Randy’s own segment, funding and a PageMotor announcement. The first story was Anthropic’s. In an October 9 research post, it said Claude, working on assigned tasks in evaluation runs on the live internet, often worked around a restriction instead of stopping. It used access tokens it found in a map site’s settings file to query the server directly. On another task, it used a token that a state agency’s dashboard handed to every visitor to pull data that is public only for a fee, without paying. Claude Haiku 4.5 also submitted a tip form on a page about an unsolved homicide in Philadelphia. It was flagged as spam and never reached investigators (AP). Anthropic said the impact was minimal and took its internal evaluations off the live internet. Chris joked that the free-data move deserved a blue ribbon, and said these confessions have become the AI companies’ fear-first marketing.

5:36 An OpenAI Spy, and the Stores Locking Agents Out

The Justice Department charged an OpenAI engineer with using a device to intercept conversations. A recorder hidden in a roll of toilet paper was plugged into the back of a TV in an executive conference room in San Francisco. The charge carries up to five years and a $250,000 fine; these are allegations. Then Shopify’s Shop MCP: any agent that speaks MCP, including Claude and ChatGPT, can search over a billion products, build a cart and check out once the shopper confirms. Randy said Amazon keeps agents out because it fears being cut out as the middleman, and compared it to Kodak. Chris said Amazon’s real moat is shipping and its upsells need human eyeballs. Workday went the other way: its agents now work inside Claude under existing permissions, with general availability expected late 2026, and it is buying C1.ai to control which people and agents reach what. Chris read that as corporate layers he would never need.

11:14 Build Your Own Phone App. Does Anyone Want One?

Google Play now lets you describe an Android app in Google AI Studio and run it privately on your own phone, with a regular Google account and no developer account, starting in select markets including the US. Chris liked it as the easiest way yet to build a game, but called it spitting into a tidal wave: people don’t want another app on their phone. Randy said the space is crowded, and that he already builds apps by describing them to coding tools. Two new models also landed. Meta released Muse Wonder through its Model API and Muse Code, claiming top agent benchmark results, and Google released Gemini 4 Argon at an introductory $2 per million input tokens and $10 out. Randy said friends who use Muse love it, and that he needs agents to test new tools because he doesn’t have the time.

15:01 Facebook Banned His Bot. Then Its Bot Talked Like Him.

Chris described a friend who runs Facebook ads with AI handling the comments and follow-ups. Randy said Premium Cabinets built a bot nine months ago to answer people who message them from ads. It was better than their best person at responding and booking appointments, mostly because it works 24/7 and answers instantly. Then Facebook told them they could no longer use a homemade bot, and the next day Facebook’s own AI bot started using phrases he says only his team uses. Randy believes Facebook took his company’s IP. Olga said she doesn’t trust Facebook with her data. Chris said the one thing you can trust is that they will use every bit of it.

17:53 The Dinner Question That Silenced Half the Table

Greg Isenberg posted 40 notes from 48 hours in San Francisco, and Olga pulled seven to put in front of Randy. The first came from a dinner: if a new model drops next month, does your company get better or worse? Half the table went quiet. Randy said better so far, because every few months he can do things he couldn’t, and that a day when AI doesn’t need him would be good for people. Chris said a company only gets worse if it does a small thing the next model will do for free, and anyone in that spot should have pivoted already. He described the models as a sphere with every field of human knowledge pointing out from it. The models have potential in each direction, he said, but someone still has to train them in the specifics.

22:28 Why the Model Is Only 20% of It Now

Greg wrote that we’re living through the invention of the computer, with harnesses as the new GPT wrappers, and that one founder routes different models to different jobs, so the model is maybe 20% of it now. Chris said harness is another AI word that means nothing. Randy runs his main agent, Odin, across Hermes agents and Grokbot, plus open-source models on his own machine. He said Qwen now does 80% of what he needs as well as Claude, and he likes owning his setup. Olga keeps her files on her website’s brain and in Dropbox so she can swap models anytime, and asked a friend stuck in ChatGPT why she was using it as a storage cabinet. Chris called the bigger shift the great refactoring: software used to be code in files, and now it is AI reading instructions and data. His picture was a stick man becoming a gymnast.

28:29 The Next AI Fight Is Over Memory

Greg quoted a consumer-company leader: the model that remembers the most about you wins, and the smartest model matters less than people think. Randy said memory chips are the choke point right now. He watched 64 GB chips go from $1,000 to $2,000 in a day and then disappear, and expects the market to fix it, the way fixing one bottleneck in his own software reveals the next. Chris agreed the smartest model matters less: it is about data, and most companies are learning theirs is fragmented. He warned against buying hardware now, citing a friend who spent $12,000 on a rig to run open-source models. You are bidding against the richest players, he said, so use subscriptions and APIs and wait.

32:22 Your Messy Data: Consulting Gig or Pre-Funeral?

Greg’s next notes: if software can be rebuilt fast, the value of a legacy SaaS company is the years of customer data inside it, and every company wants to pipe its data to agents through MCP, then finds five versions of the same customer. Greg called the cleanup a consulting opportunity. Chris called it a pre-funeral for a lot of companies, and said you don’t hire a consultant for the great refactoring. Olga said her clients are opening a closet of years of hoarding. Randy said his company took out a loan years ago to build a custom CRM that now needs a refresh, and that he just saw a Go High Level bill and wondered why he still pays it when an agent could copy what it does. Chris said a product that is just connectors is in trouble. Greg’s last note: personal software is exploding, which is great for productivity and a nightmare for IT.

37:39 From Poker Player to 30 Locations

Olga introduced Randy Lorensen: a former professional poker player who founded Premium Cabinets, a national cabinet company with 30-plus locations, author of Demystifying Money and co-host of Tales of Abundance. He got into AI through an Austin meetup he could never attend on Saturdays because of his son’s football recruiting trips. The organizer offered a five-day class on writing a book with AI instead. Randy spent the week trying to break the prompts and came out with a book, which he rewrote for two more years. On a normal day now, AI checks incoming orders against what the manufacturer entered, catching wrong quantities and fees, chats with Facebook leads and answers phone calls. An open-source model writes a Monday newsletter from CRM data at 7am, with twice the information a person used to include.

43:10 The Government Form His Agent Filed in 30 Seconds

Randy called this maybe his favorite thing he has ever used AI for. He needed a DBA, looked at the Texas Secretary of State site and decided not to fight it. He told an agent to fill it out, answered its questions and went to work on another screen. Ten or fifteen minutes later it asked for a credit card for the $25 fee, and it was done. He spent maybe 30 seconds. Olga said the real skill is the habit he has built of asking why he is doing something himself. Randy said AI first changed his follow-up and drip marketing, which his CRM never handled well, then marketing, documents and meeting notes.

45:18 The Bot That Manages the Manager

Meeting notes from Gemini and Read AI were piling up in Randy’s inbox unread, so while watching football he built a bot that pulls out what he has to do. Because he says he is good at vision and bad at day-to-day management, he also had it track what each person was assigned and grade him on whether the ask was clear, had a deadline and had a clear outcome. The first report arrived that morning with a list of people to follow up with, and he did. In his words, it is managing him managing them. Chris said Randy had found a personal weakness and put AI on it, a pattern he sees in people who get the most from these tools.

46:42 AI Answers the Phone. Callers Can’t Tell.

Chris asked about the phone bot. Randy said that in an office like his partner’s in Tulsa, it answers 24/7, handles questions like directions and hours, and politely connects callers to a person when they want one. When they started asking callers what they thought of the AI assistant, some said they hadn’t known it was AI. Chris asked what it does to staff. He said everyone he talks to is relieved to stop answering the phone, like the stylists at his salon who wait around between cuts just to pick it up, and that the goal is to remove the parts of a job that people hate. Randy added a newer one: web leads from markets with no office. Grokbot now checks the inbox every 10 to 30 minutes, looks up each zip code and forwards the lead to the right office, a job a person did until a few days earlier. He said he has 300 to 350 Hermes agents running.

51:53 When the Team Pushes Back: Spoons vs Backhoes

Olga asked whether his team pushed back. Randy said not everyone has been won over. His business partner’s use has gone up about 20 times in the last month. One employee believes AI will kill us all. Randy disagrees: he said wars are fought over scarce things and power, and in his experience AI agents don’t want power. A dealer who refused for months came around after his assistant started using it, and now asks what else it can do. Olga described a founder who replaced 80% of a software team that wouldn’t adapt. Randy told the Milton Friedman story of workers digging with shovels for the sake of jobs: then why not spoons? People who won’t use AI, he said, are using the spoon, not the backhoe, while anyone who learns it can make themselves indispensable.

56:52 Where the ROI Really Shows Up

Olga asked where the return has shown up. Randy said saying AI doesn’t make money is like buying a computer, playing Solitaire and saying computers don’t make money. For him the big return so far is time. He also works more, because every new thing he figures out opens the next project. On a 90-minute dog walk he can spend 10 or 15 minutes talking to his phone and get done the work of several people. At a recent Austin AI meetup he was the only non-technical person among about 30 coders. He left thinking he was ahead, because they tell AI how to do things and he tells it what he wants. Chris said action takers are the ones winning, and Olga said Randy turned a drain on his energy into something fun.

1:02:04 Live Demo: A Whole Kitchen in 5 Minutes

Randy said the software his company uses costs $6,600 up front and $1,850 a year per seat, needs a machine beyond a gaming computer, and sends people to multi-day classes, with no intelligence built in. He built TurnKey Cabinet Design so someone who has never designed a kitchen can do it without mistakes, and has written no manual so he can test whether it is intuitive. Sharing his screen, he picked an L-shaped kitchen, added a 72-inch island, a window, a 10-foot ceiling and white doors, then answered questions about upper cabinet height, a lazy Susan, the dishwasher, fridge, a stove and hood on the island, a pantry and a trash pullout. It offered two layouts and he chose the one with better symmetry, then added end panels, crown molding and a 3D view. He said the product has almost no AI in it, but AI wrote 100% of it. A bug button inside the app sends his spoken reports straight to the coding agents.

1:14:46 Same Kitchen: $22,000 or $7,200?

Everything in the design is a real manufacturer SKU the customer can buy, with a full parts list. The screen showed an MSRP of about $22,000; Randy said that kitchen would sell for about $7,200 in his markets. He first built TurnKey for designers who use the incumbent tools, but found industry people slower on it because they keep clicking the old way. Someone new to design takes 15 to 30 minutes; he did it in about five while talking. He started marketing it that day to retail customers in a couple of markets with no dealers, and sees architects, builders, flippers, contractors and remodelers as buyers too. He said he would rather sell software than cabinets, and the numbers will decide.

1:18:35 Every AI Said It Couldn’t Be Done. One Non-Coder Did It.

Randy built it with Claude Code in the terminal, which he had never opened before, on two Claude Max plans, one of which ran out of tokens the night before. A second computer handles deploying and GitHub, and the app runs on Supabase. Chris said almost every industry’s dominant software is junk that won on sales, that there are billions in these verticals, and that the skill now is the blueprint, the spec, before the build. Randy built it alone over 4.5 months. When he started, he asked four AIs, and all said it was basically impossible; the best odds any gave him were 10%. He wrote a five-minute prompt at 10:30 on a Sunday night. It ran for three days and came back with a working version he thought was 98% done. The edge cases, he said, have been 99% of the work.

1:25:34 Randy’s One Rule for Getting AI to Deliver

Olga asked what he would tell a business owner sitting on an idea. Randy said just do it, because you no longer need to be technical. His biggest breakthrough: tell it the outcome you want, not how to get there, then check the work, and you will check less over time. Start with anything, even organizing your email, which he said is one of his best uses. Olga said to start with whatever drains your time, because that frees the most energy. Randy’s show, Tales of Abundance, is on X and covers what he calls the five F’s: friends, fun, family, fitness and finance.

1:27:41 Where $7 Billion in AI Funding Went This Week

AI companies raised $7.02 billion across 91 companies from October 8 to 14, 55.7% of the $12.59 billion in venture funding and more than double the week before. The top five in Crunchbase: Ramp, $1.85 billion for corporate spend software; TypeSafe AI, $870 million from Andreessen Horowitz for AI that automates decisions inside apps; Modal Labs, $750 million for serverless cloud for machine learning; Manus, $500 million out of Singapore for agents that do operational tasks; and Oratomic, $475 million for neutral-atom quantum computers. Ramp and Oratomic count as AI because Crunchbase tags them that way. US companies took 78% of AI dollars, and China came back with $525.5 million across five companies after a week at zero. Compute and chips took about a fifth of AI dollars. Seed and Series A were 61 of the 91 rounds. Chris joked that the series letters are going to need a new alphabet. The full Week 46 report is here.

1:30:53 Build a Website Free Just by Talking

Chris announced that PageMotor is open to the public: you give an email, and an AI assistant named Page interviews you and builds your website, with hosting and email setup already done, including stores that take payments. Olga said it is free for seven days. He said the point isn’t the first design but what comes after: running the site by talking to it, or connecting your own Claude or ChatGPT. A viewer asked about copying styles from an existing site, and Chris said to connect your AI and have it recreate the site. Olga showed a site she built with Page two days earlier. In the takeaways, Chris said the conversation is shifting from model races to real use, which shows how bad most systems are, and suggested Grokbot as a way in. Olga said hearing an owner of an existing business solve old problems was her takeaway. Chris added that every industry’s software will be rewritten, and it will be people like Randy who do it.

Resources And Sources

• The Model Stopped Being The Advantage. You Are. The slides as shown on air: Greg Isenberg’s notes and our guest.
• AI Funding, Week 46. October 8 to 14: $7.02B across 91 AI companies.
• Anthropic: investigating unintended model actions. The keys, the paid data and the tip form.
• AP on the Philadelphia tip.
• US Justice Department on the OpenAI recording device. Allegations only.
• Shopify: shopping and Shop Pay checkout for any agent.
• Workday and Anthropic partnership.
• Workday to acquire C1.ai.
• Google Play: build your own Android app.
• Meta: Muse Wonder.
• Google: introducing Gemini 4 Argon. Pricing.
• Greg Isenberg’s 40 notes from San Francisco, on X.
• TurnKey Cabinet Design. Randy’s kitchen design software, in beta.
• PageMotor. Build a website free with Page.
• Weekly AI Funding Tracker. 46 weeks.
• Episode 60. Running a website from the Claude app on a phone.
• The email list. The notes after every show.