AI Arbitrage: 7 Tiny Agent Businesses You Can Start Today

Greg Isenberg gets the same comment on every video.

"Stop giving us billion dollar startup ideas. Give me something tiny I can start with AI that actually cash flows."

So he made an episode that's nothing but that. Seven tiny AI arbitrage business ideas, a framework for generating your own, and a live build where an agent scraped 222 job postings and wrote 14 cold emails before he finished explaining what it was doing.

I watched the whole thing so you don't have to. Here's the playbook.

What an AI arbitrage business actually is

Strip the buzzwords and it's this: somewhere on the internet, an asset is mispriced. A domain, a used pizza oven, an abandoned app. Someone would pay more for it than it's listed for.

The problem was never the spread. The problem was that finding these deals meant a human refreshing Craigslist at 7am every day.

That job is now an agent's job.

You point an AI agent at a messy feed (auctions, listings, job boards), give it a criteria list, and it drops a ranked list of deals into your Slack every morning. You take the spread.

Greg's framing: this is what people mean when they say agents are the new SaaS. You're not selling seats. You're selling an outcome, monitored around the clock by something that doesn't sleep.

One disclosure before the ideas: the episode is sponsored by GenSpark, and Greg builds everything in it with GenSpark Claw (their cloud take on OpenClaw, living in his Slack). The tool matters less than the pattern. Any agent that can browse, scrape, and message you works.

The 7 AI arbitrage business ideas

1. The dead domain flipper

Every day, good .com domains expire because someone forgot to renew them.

Greg's agent monitors the drops (GoDaddy Auctions, DropCatch) against his criteria: niche keywords, DR 20+, clean backlink profile, no adult or gambling history. Every morning it posts a ranked list of 10 domains worth bidding on to his Slack.

He's run this business before, manually. His old team bought dropped domains for $8, designed a logo for each one, and sold the pair for $3,000 to $5,000.

Same margins. Except now the scouting is automated, and changing your max bid means typing "make it $2,500" into Slack.

2. The local liquidation broker

Restaurants close constantly. When they do, someone has to sell the equipment, and that someone is usually a stressed owner posting a commercial smoke hood on Craigslist for $1,500 with a terrible title.

Greg's agent monitors BizBuySell, AuctionZip, BidSpotter, Craigslist, and his local bankruptcy court filings. It comps 40+ equipment types against live eBay sold prices and posts deal cards showing the spread.

First run: 327 listings scanned, 10 deals flagged, including that hood at roughly a 300% spread.

The model is brokering. Connect the estate to the new restaurant, take 15 to 30%, hold zero inventory.

3. Hiring signals as sales leads

This is the one Greg built live on camera, and it took about 5 minutes.

The logic: when a company posts a job for a Head of Growth, budget just got approved. When they're hiring three SDRs, they're about to buy outbound tools. A job posting is a company announcing it's spending money.

His agent scraped 222 postings from Hacker News Who's Hiring, RemoteOK, and Greenhouse. It surfaced 14 companies hiring marketing leaders, found the decision-makers, and drafted a personalized cold email for each one referencing the exact post.

The first batch had an HTML bug in the links. Greg fixed it by telling the agent, in a Slack message, that nobody converts from an email that looks broken. It diagnosed the problem (HTML entities bleeding into the addresses) and patched itself.

That's the whole debugging process now.

4. The "should I even call" memo

Buying a small business is mostly wasted diligence. You spend hours on a listing before finding the dealbreaker.

So: point the agent at BizBuySell or TrustMRR, have it pull the financials, cross-check reviews and web mentions, and produce a go or no-go memo in 6 minutes.

Two ways to make money here. Buy the good businesses yourself, or sell the memo as a service to people who want to.

5. Dead Product Hunt launches with living traffic

Thousands of products launched on Product Hunt 2 to 4 years ago are dead. Founder moved on, site abandoned. But some of them still rank, and the SEO traffic is still arriving at a door nobody answers.

The agent finds those. Then you email the founder and offer $5K to $25K. A lot of them will say yes just to stop paying the AWS bill.

You're buying traffic that took years to build, at liquidation pricing.

6. Fallen App Store apps

Same idea, mobile edition. The agent watches App Store rankings for apps that were top 100 a few years ago, have slid past 500, but still carry 10,000 to 100,000 reviews.

Those reviews are the moat. Ranking decay is usually a monetization problem, not a product problem.

Pull the developer contact, offer $10K to $50K, relaunch with better monetization. No cash? Bring the deal to an investor and keep a piece for sourcing it.

7. Competitive intel as a product

The simplest one. The agent monitors your top five competitors overnight: pricing changes, new pages, founder tweets, job postings, changelog updates. You wake up to a one-page brief of what moved in your market while you slept.

Use it for your own company, or sell it as a monthly subscription to founders who are too busy to stalk their competitors properly.

The framework: feed, asset, trigger, buyer, payday

The seven ideas are one idea wearing seven outfits. Greg's framework:

→ A messy feed. Job boards, auction sites, expired domain drops, app rankings, court filings. Somewhere data churns daily and nobody reads all of it.

→ A mispriced asset. A domain, a pizza oven, a dead SaaS, stale traffic, distressed inventory.

→ A trigger event. A drop, a closure, a job posting, a rank decline. The moment the asset becomes available or the buyer becomes motivated.

→ An obvious buyer. An operator, an agency, a founder, ideally someone with money already allocated.

→ A liquidity point. How you get paid: flip it, broker it, retainer it, or relaunch it.

If you can name all five for your idea, an agent can probably run it.

How to pressure-test your own idea

Greg's screening questions, which I'd steal verbatim:

→ Is there urgency? (A liquidating restaurant can't wait.)

→ Is there a real spread? (300% on used equipment, yes. 5% on retail goods, no.)

→ Can an agent actually monitor the feed? (Public listings, yes. Gated data, harder.)

→ Who pays first? (If you can't name them, you have a hobby.)

And where to hunt: places with constant change (marketplaces, listings, filings, job postings) and things people ignore (stale traffic, abandoned software, underpriced attention).

Court filings are the most underrated feed on that list. Almost nobody reads them, and they're public.

My take

I like these businesses more than most "start an AI agency" advice, for one reason: they don't depend on convincing anyone to believe in AI. The buyer of a $1,500 restaurant hood does not care that an agent found the listing.

That's the same reason boring businesses keep quietly printing money while everyone fights over the shiny categories.

Two caveats.

One: the agent finds the deal, but a human still closes it. Greg glosses over the part where you call Chris about his smoke hood, negotiate, and arrange a truck. The monitoring is automated. The brokering is a job.

Two: spreads attract competition. Twelve months from now, expired domains will have twenty agents bidding against each other. The durable version of every idea on this list is the one where you build a relationship layer the agent can't replicate (the restaurant owners who call you first, the founders who trust your memos).

Start with one feed. Get one deal. Then decide if it's a business.

FAQ

How much does it cost to start an AI arbitrage business?

Almost nothing on the software side. An agent tool subscription runs $20 to $30 a month, and most of the feeds Greg uses (Craigslist, GoDaddy Auctions, Hacker News, court filings) are free to monitor. The real capital need depends on the model: brokering and lead-gen ideas need $0 in inventory, domain flipping needs a few hundred dollars of float, and buying apps or dead SaaS products needs $5K to $50K (or an investor who has it).

Do I need to be technical to build these agents?

No. Greg builds the hiring-signals system live in the episode by describing it in plain English, answering a few clarifying questions from the agent, and pasting one command into a terminal. When something broke (malformed links in the cold emails), he fixed it by telling the agent what was wrong in a Slack message. If you can write a clear paragraph describing what you want monitored and what a good deal looks like, you can build these.

Which of the seven ideas should I start with?

The one where you already know the buyer. Arbitrage dies at the liquidity point, not at the finding-deals stage. If you know restaurant owners, do liquidation brokering. If you sell marketing services, do hiring signals. If you've ever bought or sold a site, do domains or dead Product Hunt projects. The agent solves the monitoring problem. Your existing network solves the harder one.

Is AI arbitrage a real business or a side hustle?

It starts as a side hustle and earns its way up. One good domain flip a month is beer money. But competitive intel sold as a subscription, or liquidation brokering run in three cities, compounds into real revenue. Greg's framing is that these are cash-flowing businesses you run with an AI employee, in the $1,000 to $3,000 a day range at the ambitious end, not venture-scale companies. That's the point.