Plus the one rule that keeps all three out of trouble.
How to turn your AI Assistant hobby project into a thriving business — Image by Gemini
In “Build Your Own Assistant, and Keep It,” I made the case for an AI that lives on your own hardware, answers to you, and doesn’t report back to anyone. Readers came back with a fair question almost immediately: Fine, I own it. Now what do I actually build with it?
This piece answers with three businesses. None of them needs venture capital. All of them can start as a side project on a single machine. Each one is far more practical this fall than it was a year ago, but each one will fail if it’s built carelessly. I’m not saying this to dissuade anyone, just go in with your eyes open because the business vision is the only barrier to entry.
First, the news that makes this moment matter.
The Week the Storefronts Picked Sides
On the night of September 20th, users of Meta’s new AI assistant, Muse, hit a wall when they tried to buy things on Amazon. A popup told them that continued access by an unauthorized AI agent violated Amazon’s Conditions of Use.
Muse had launched on September 8th. It lasted twelve days.
Amazon’s complaint was specific. Meta gave no advance notice, the agent browsed without identifying itself as software, and it appeared to capture and store customer credentials. Meta says Muse has no visibility into passwords or payment methods. Either way, the outcome was the same: blocked.
Amazon isn’t an outlier. Over the past year it has pushed back on shopping agents from Google, OpenAI, and Perplexity. eBay rewrote its user agreement earlier this year to ban AI shopping bots, and the motive is easy to see. eBay takes a cut of the final sale price, so an agent built to win auctions cheaply shrinks eBay’s revenue.
Now look across the street. Shopify turned on its Agentic Storefronts by default for eligible merchants this spring, making millions of stores searchable by assistants like ChatGPT, Microsoft Copilot, and Google’s AI Mode. No scraping required. The endpoints that software agents speak to are already live on every store.
Two retail giants, two opposite philosophies. That divide is the map for everything that follows.
Your Agent, Not a Consumer App
Before going further, a question worth settling: can you build these businesses on Muse, ChatGPT, or Copilot?
You can experiment with them. You shouldn’t build on them. Consumer assistants are designed for consumer convenience. They belong to someone else, they follow someone else’s partnerships, and when two giants go to war, as Amazon and Meta just did, every user of that assistant becomes collateral. Your workflow can vanish overnight because of a dispute you had no part in.
An enduring business needs developer-level control. That means one of two setups.
The first is a private model running on your own machine, served by a tool like Ollama, LM Studio, or vLLM. Nothing leaves the box unless you send it. This is the fully sovereign option, and it’s the one my earlier piece walks through.
The second is an MCP-connected client, such as Claude Desktop, Cursor, or a small Python harness you write yourself, that uses your own API keys and your own tool connections. MCP stands for Model Context Protocol, the emerging standard for how AI assistants talk to outside services. If the model behind that client is hosted in the cloud, your prompts travel to it, so choose based on how sensitive the work is. But the keys, the tools, and the rules are yours.
One honest caveat: owning the agent doesn’t exempt you from anyone’s terms of service. Amazon’s rules apply to your homemade agent just as much as to Meta’s. What ownership gives you is choice. You decide which platforms to work with, how your agent identifies itself, and when it stops and asks. That choice is what the rest of this article is about.
Read the Business Model, Not the Website
Here’s the simplest test for whether a platform will welcome your agent or ban it. Ask how the platform makes money.
Attention platforms earn money from your time. They sell ad impressions, sponsored placements, and minutes spent scrolling. An agent threatens all of that. It doesn’t scroll. It skips the sponsored rows, applies your constraints, and picks the one item that actually fits. Every agent-driven purchase is an ad impression that never happened. Attention platforms will block, throttle, and litigate.
Transaction platforms earn money when something settles. A hotel booking confirms. A lease gets signed. A niche product sells and the platform takes its fee. These platforms don’t care whether a finger or a piece of software pressed the final button. They care that the order cleared.
Your agent should live almost entirely in the second world.
When a service offers an official path, whether an MCP server or an authenticated application programming interface, use it. When no official path exists, your agent can fall back to reading pages inside your own logged-in browser, at a human pace. What it must never do is disguise itself, borrow someone else’s credentials, or sneak past a platform that has already said no. That isn’t clever engineering. It’s how you lose a fifteen-year-old account in one afternoon.
The Golden Rule: Read Freely, Propose in Drafts, Commit Only With a Human
All three businesses rest on one boundary. Your agent works across three levels of authority, and only the first two are its own.
Read. The agent inspects filings, prices, listings, inventory counts, and reviews. Reading is cheap, continuous, and safe.
Propose. The agent drafts the rebalance, the listing, the rental shortlist, or the itinerary. It does the arithmetic, shows its reasoning, and stages everything for one final action.
Commit. The agent stops and waits. Money never moves, contracts never get signed, and messages never go out without a human saying yes, with a click, a key, or a spoken confirmation.
Engineers will recognize the spirit of a two-phase commit, but in practice, it’s a human-in-the-loop air gap. It protects you from model mistakes. It protects you from instructions hidden in web pages your agent reads. And it keeps you inside nearly every platform’s rules.
Every business below depends on it.
Idea One: The Patient Investor
What It Is
Picture an investing partner with no appetite for drama. It doesn’t day trade, chase trending stocks, or read social media for sentiment. It reads filings, watches your balances, and about once a month taps you on the shoulder with a calm suggestion. Then it waits.
Language models are poor at predicting tomorrow’s market. Anyone selling an agent that claims otherwise is selling fiction. What they do well is read a seventy-page annual report without getting tired, notice when your portfolio has drifted from your plan, do the rebalancing math, and explain it in plain language.
How a Month Runs
Say your plan is conservative: seventy percent in a broad, low-cost index fund, twenty percent in short-term Treasuries, and ten percent in cash. You wrote it with a clear head, and the point is to stick to it when markets get loud.
Every morning, the agent checks your balances through a read-only connection to your brokerage. Most mornings, it logs nothing and goes quiet.
Then a rally pushes your stocks to seventy-four percent. The agent flags the four-point drift, calculates the trade that fixes it, and drafts a limit order, which is an order that executes only at a price you set or better. Then it speaks up:
“Stocks are four points above target. I’ve drafted an order to sell about fifteen hundred dollars of the index fund and move it into short-term Treasuries. Want to review it?”
You can question the math, ask about the tax hit, or say not this month. Nothing leaves your network until you confirm.
In the background, the agent pulls annual and quarterly reports from the SEC’s public filing system, EDGAR, and files them into a searchable library on your own drive. Ask it how a company’s debt changed from last year, and it answers with the actual passages. Your research and your reasoning stay on your hardware.
The Business
There are three ways to take this idea, and only some of them are simple.
Use it yourself. Not a business, but possibly the most valuable version. The biggest threat to most people’s returns is their own impulses, and an agent that holds you to your own plan is worth real money.
Sell the software. You build the tool. Customers install it on their own hardware, connect their own read-only brokerage keys, and pay you for a license or ongoing updates. You never touch their money and never tell them what to buy. This is the cleanest business, and it fits the sovereign idea perfectly.
Advise others for pay. This is the regulated route. In the United States, giving personalized investment advice for compensation generally requires registering as an investment adviser with your state or the SEC, under the Investment Advisers Act of 1940. Wrapping that advice in software doesn’t get you around it. Treat this line as absolute until a securities lawyer tells you otherwise. I’m not a lawyer or a financial advisor, and nothing here is investment advice.
Operational Risks
Execution errors. A misplaced decimal on a trade is a very different mistake from a typo in an email. Keep brokerage access read-only by default, route every order through confirmation, hard-code a maximum order size, and log every request.
Taxes. Rebalancing can realize capital gains. The agent should track how long you’ve held each position and flag possible wash sales before it drafts anything.
Your First Month
Write your plan in plain sentences. Connect one account, read-only. Have the agent send a weekly drift report and nothing more. Only once you trust its arithmetic should it start drafting orders.
Idea Two: The Curated Shop
What It Is
Conventional dropshipping earned its bad reputation. Think storefronts full of products nobody inspected, marked up from overseas suppliers, with margins that depend on shoppers not knowing better.
The curated shop runs on the opposite principle. You pick one category you know deeply, like homelab rack accessories, mechanical keyboard parts, or specialty espresso tools, and you sample, test, and document every item you sell. Storage, packing, and shipping go to a domestic fulfillment company, often called a third-party logistics partner, that you connect to by API. Your assistant runs the operations in between.
Your taste is the product. The software keeps the shelves straight.
The Division of Labor
The agent handles what scales. It keeps inventory counts in sync between your supplier and your store. It writes listings with precise specs. It checks that each order’s margin still covers fulfillment, sends the order to the warehouse, and passes tracking back to the customer. It sorts incoming email and drafts support replies for you to approve.
You handle what doesn’t. You test samples, vet suppliers, set prices, and settle anything unusual, especially refunds and disputes. An agent can route an order, but it can’t feel the stitching on a canvas bag or spot a cold solder joint on a circuit board.
The Distribution Shift: Selling to Other People’s Agents
Here’s what changed this year. More and more shoppers send their assistant to find a product instead of browsing for it themselves. That assistant skips the pretty storefront and reads structured product data directly.
On Shopify, that happens through its Storefront MCP servers and the Universal Commerce Protocol it built with Google. A shopper’s agent queries your catalog and reads exact materials, verified dimensions, honest lead times, and real stock levels. If the order goes through, your own assistant checks it, you approve anything out of the ordinary, and the warehouse ships it.
That means your product data is now your shop window. A small store with clean, accurate data can appear right next to major brands inside someone’s AI assistant. It’s one of the best openings independent makers have had in years.
One caution. The ad layer is already arriving here too. Shopify’s summer release added Promoted Placements, which let merchants pay for visibility in AI-driven shopping. The agent-driven web won’t stay ad-free on its own. Accurate data and a real reputation are your defense.
Operational Risks
Inventory drift. If your supplier sells out and its data feed lags, you’ll sell things you can’t ship. Reconcile stock daily, and have the agent pause any listing whose feed hasn’t updated in twenty-four hours.
Shipping promises. The Federal Trade Commission’s Mail, Internet, or Telephone Order Merchandise Rule requires you to ship within the time you promise, or within thirty days if you don’t state one. If you’re delayed, you must tell the customer and offer them the choice to wait or get a prompt refund.
Chargebacks. Defective products lead to disputes, and disputes lead to frozen merchant accounts. Inspect every first batch by hand, and never let the agent deny a refund on its own.
Your First Month
Pick one niche. Order samples of ten products and keep the three you’d happily give a friend. Open a store, connect one fulfillment partner, and let the agent draft listings for you to edit. Sell to people you know before strangers find you.
Idea Three: The Scout
What It Is
The scout is a search concierge for high-friction decisions with lots of constraints: medium-term rentals, complicated trips, lodging that has to work around a job.
These searches aren’t browsing problems. They’re optimization problems. Take a traveling nurse starting a thirteen-week hospital contract. She doesn’t want two hundred apartments. She wants a furnished place with a lease that fits her contract, dedicated parking, a commute under twenty minutes at night-shift hours, and the real move-in cost, including pet fees and deposits, stated up front.
Humans solve this by opening dozens of tabs and building a spreadsheet by hand. An agent treats the web as the database it secretly is.
How a Search Runs
Intake. The client describes what they need in plain conversation. The scout asks the questions they didn’t think of. Will you commute at rush hour? Do you need secure package delivery?
Gathering and normalizing. The scout collects listings from supported sources and adds up the true cost, including application fees, admin fees, pet rent, and deposits. It estimates the commute at the hours the client will actually drive, not at midday.
Fraud filtering. It flags warning signs, like the same photos appearing under a different address, or a landlord who wants money before a showing.
The shortlist. The client gets a short briefing with three to five strong options, each with its tradeoffs and full costs. When they pick one, the scout drafts the showing request. The client approves it before it goes out.
Travel works the same way. The scout weighs fares, layovers, refund rules, and baggage fees together and presents a few clear itineraries instead of forty options sorted by whoever paid to be on top. It stages the booking. The client confirms it.
That confirmation step matters more than it looks. Many rental and home-services sites charge landlords and contractors for every inquiry. An agent that fires off twenty tour requests is spending other people’s money on leads that go nowhere. The scout reads widely and asks sparingly.
The Business Model and the Incentive Trap
Charge the client. A flat fee per completed search, or a monthly retainer for people and companies that relocate often. Your incentives line up perfectly with the client’s.
Or take referral commissions, carefully. This is the trap. Once listings pay you, you have a reason to favor the ones that pay most, which is exactly how the ad-driven web lost people’s trust. If you accept referral fees, disclose them plainly, and build your ranking so it can’t see who pays.
Legal Lines to Respect
Fair housing. Federal and state fair housing laws prohibit steering people based on race, color, national origin, religion, sex, familial status, or disability. Your filters should describe the home, like price, size, pet rules, and commute time, and never the people who live nearby.
Real estate licensing. Some states, including Texas, treat paid apartment-locating as real estate brokerage, which requires a license.
Seller-of-travel rules. California, Florida, Washington, and Hawaii, among others, require registration if you sell or arrange travel for their residents.
Check your state and your clients’ states before you charge anyone. A short consultation with a local attorney earns its fee here.
Your First Month
Do the job by hand for five people you know who are moving or traveling. Write down every question you asked and every step you took. That list becomes your agent’s intake interview and workflow. Automate the searching first, the summaries second, and outbound messages last, always behind a confirmation.
Choosing Your Path
The patient investor needs almost no capital beyond a computer you already own, and the technical lift is light. Its bottleneck is regulation: you can’t advise others without a license. It’s best for builders who want discipline and clarity in their own portfolio, and possibly a software product later.
The curated shop needs roughly one to three thousand dollars for samples and setup, and moderate technical work. Its bottleneck is physical: supplier reliability and product quality. It’s best for people with real taste and expertise in a tangible niche.
The scout needs almost no capital and moderate technical work. Its bottleneck is trust, plus the licensing checks in your state. It’s best for problem-solvers who enjoy logistics and working closely with clients.
Whichever you choose, three habits carry you. Work with platforms whose business models welcome you. Let the agent read freely and draft everything, but commit nothing without a human. And keep your data, and your clients’ data, on hardware you control.
The Bigger Picture
The clash between the big platforms isn’t a technical squabble. It’s a contest over who sits between a person and their decisions.
Advertising platforms need you inside their feeds, where impressions can be sold. Transaction platforms are opening their catalogs to agents through protocols like MCP and the Universal Commerce Protocol, because agents bring them buyers.
A private assistant gives you a third position. It answers only to you, runs on your desk, and serves no ad network. Point it at the platforms that want to do business in the open, build every workflow around a human’s final yes, and keep your data at home. That’s how you build a business on durability instead of on the next loophole.