Which AI Projects Fit a One-Person Company — and Which Ones Quietly Eat You Alive

The easiest test to reach for is "I could build this myself." That was reasonable in 2023. It isn't now — building the thing is no longer the hard part.

What gets people isn't the cost of starting. It's the cost of keeping it alive. So ask a different question: does this keep running without eating more of my hours?

01 · First, the filters

1. Delivery has to approach zero. The first delivery takes ten hours; the hundredth should take ten minutes. The gap is whether you wrote it down as a process. Anything that needs fresh thinking every single time keeps you capped at one person, no matter how good you are.

2. The demand has to be homogeneous. A hundred customers asking a hundred versions of the same question. The more bespoke it gets, the closer you are to freelancing.

3. Marginal cost has to be predictable. Customer 101 should cost a few API calls, not another all-nighter.

4. Nothing that requires credentials to get in the door. If the deal needs a licence, a wall of enterprise client logos, or "our team of fifty," a solo founder has already lost at the starting line.

5. Traffic can't be fully rented. A business that goes to zero the moment a platform changes its rules is a house built on someone else's land.

Four out of five and it's worth testing. Two or three, and you'll end up busy without accumulating anything.

02 · Seven project types that actually fit one person

1. Single-purpose tools (micro-SaaS)

One page, one problem: batch-fixing invoice details, subtitle translation, rewriting a résumé against a specific job posting, bulk-renaming and watermarking images. Narrow scope, low maintenance, predictable subscription revenue at $5–$20 a month. A few thousand users is entirely manageable on your own.

The difficulty is restraint. Expand the scope and the single-point advantage evaporates. Adding features is far easier than holding back.

2. Packaging domain knowledge as an agent

Take a workflow you know cold — customs clearance, public tenders, clinical trial recruitment, school administration, cross-border compliance — and turn it into an agent or a skill pack. Model capability is the same for everyone. Your process isn't. That gap is the arbitrage.

Charge by seat, by project, or as training plus ongoing support.

The ceiling sits in the same place: clients are buying you, and growth hits your calendar immediately. The way out is to document the delivery and sell the process instead of your time.

3. Productized services

Turn a service into a fixed-price package: 50 short-form video cuts for $299, 100 localized product listings for $499, 20 industry articles for $899. The demand is well defined, delivery batches cleanly, the price is on the page, and there is no sales team to pay.

The catch: this is still time-for-money, just wearing a nicer shirt. Unless every order gets folded into a template, month twelve feels exactly like month one.

4. Directories, comparison sites, rankings

This site is one: scattered tools, rankings, and prices, organized into something searchable. It survives on one operator for two unglamorous reasons — the content has a fixed update cadence (daily, weekly, monthly rankings), and the collection, cleaning, and rendering are all scriptable.

It also needs no support desk and no per-order delivery. Long-tail search compounds.

Revenue comes from display ads, sponsorship slots, affiliate commission, and paid listings.

The cost sits up front. Content is a heavy asset with roughly a twelve-month stretch before it earns anything, and once you stop updating, traffic decays faster than it grew.

5. Digital products: templates, datasets, benchmarks

Files you can deliver and copy: a Notion system, a prompt library, an industry dataset, an eval set for one specific task. Build once, sell forever, marginal cost zero.

But the ticket is low, volume decides, and you need distribution you already own. Shipping it and having nobody find out is the most common way this dies.

6. Small AI integration work

Wire up a RAG support bot for a local business, automate a spreadsheet report, clean up a sales team's lead list. Deals land in the low thousands of dollars. The technical depth is modest; people who understand the business and can actually ship are rare.

Charge a project fee plus a monthly retainer. The retainer is where the cash flow actually lives.

The annoyance is that every client's systems are different, so it never replicates. Which is exactly why the retainer is non-negotiable. Without it you will spend your life hunting the next client.

7. Content IP, produced with AI

Newsletters, tutorials, short-form comics, podcasts. AI collapsed the production cost. It did nothing to the cost of getting read.

Still worth doing: a distinct point of view accumulates readers, and readers are the most platform-resistant asset you can own.

It's just the slowest path here — eighteen months before it means anything. Treat it as a distribution asset rather than a revenue line and it usually works out better.

03 · Six one-person companies worth studying

None of these six are really "AI companies." Each one took a single slice of the delivery chain, pushed its cost toward zero, and turned the savings into either price or margin. Read them for the cost each one attacked, not for the product shape.

1. Pieter Levels — PhotoAI, InteriorAI

AI headshots and interior design renders, run by one person on a deliberately thin stack, with public revenue in the seven figures.

The move: treat the shared base as the asset, not the products. Several products run on one login, one payment integration, one image pipeline, so the second product costs almost nothing to launch. Most solo builders are spent after one product because every new idea means rebuilding the base. He built the base thick first, then hung products off it.

2. Danny Postma — HeadshotPro

AI professional headshots at $29–$59, delivered in thirty to sixty minutes, fully automated, publicly in the seven figures.

The move: compress the demand into one input-to-output sentence. "Turn a phone selfie into a photo you can put on LinkedIn" solves three problems at once — the ad explains itself, the user knows what to upload, and anyone can judge the result at a glance. Conversion is high because the demand has no ambiguity in it, not because the creative is clever.

3. Tony Dinh — TypingMind

A one-person AI chat front end, sold on interface, bring-your-own API key, and a one-time licence.

The move: delete a metric with your pricing. A subscription means re-earning every customer every month; a one-time licence makes churn disappear from the dashboard. One person can't run a retention operation, so he removed the need for one. That is product design doing double duty as cost design.

4. Jon Yongfook — Bannerbear

A one-person API for automated image and video generation, five figures a month, selling to developers rather than end users.

The trade: outsource support to the documentation. Developers read the docs before they pay, so writing them properly once is like hiring a pre-sales engineer who never sleeps. End users don't read docs; developers do. Choosing your customer type is really choosing whether you do support at all.

5. Damon Chen — PDF.ai

One person behind "chat with your PDF," publicly around $30k a month.

He didn't create the demand. He caught it. "chat with pdf" was already something people typed into a search box. He didn't invent a scenario; he intercepted an existing intent and turned it into a product. Most solo products die trying to educate a market. He never got to that stage.

6. Bhanu Teja — SiteGPT

One person behind "put an AI support agent on your website," around $20k a month, selling to SMBs.

The move: pair a high ticket with a fixed pipeline. Businesses pay more, but they also ask for more. His answer was to collapse delivery into three steps — upload the docs, generate the script, paste one line of code — which lets a high ticket coexist with low delivery effort.

What gets cut differs every time: rebuild cost, explanation cost, retention cost, pre-sales cost, market education, custom delivery. The one thing none of them cut is judgment about what to build and what not to. The model can't take that over, and that is exactly where the moat still sits.

One caveat before copying any of it: none of these economics transfers cleanly. Payment rails decide whether you can run subscriptions at all, the language of your market decides your acquisition cost, and compliance decides whether you can sell to businesses. The demand definition and the automated delivery travel; the distribution advantage does not.

04 · Four that look great and are wrong for one person

• Training your own foundation or vertical model. Capital- and compute-intensive, and your hours are the least valuable input in that equation.
• Anything that needs a licence — medical diagnosis, investment advice, legal opinions. Compliance eats the entire margin.
• Heavy custom outsourcing. Every project starts from zero and leaves nothing behind.
• Thin wrappers riding one platform's traffic. One rule change and the business is gone.

05 · The cost structure, or how to hold 80%+ gross margin alone

A one-person company has three costs: model API, hosting and domain, payment processing. No payroll means your margin ceiling is set directly by the token price you pick.

Which makes model selection a financial decision, not a technical one. On the HaiAi123 intelligence index (as of November 7, 2026), the price spread across the board is roughly 40×, while the score spread is under 2×:

Run the numbers: a task taking 5,000 input and 1,000 output tokens costs about $0.001 on the volume tier and about $0.04 on the frontier tier. At ten thousand calls a day, that is nearly $400 a day.

So the answer isn't "use the best model." It's tiered routing: batching, formatting, classification, and extraction go cheap; judgment calls, customer-facing copy, and hard reasoning go expensive. That routing layer is the real moat for a lot of solo businesses — it sets your margin, and it decides whether you have any room to compete on price at all.

06 · A 90-day validation path

Weeks 1–2 — deliver 20 orders by hand. No product. One post or one email: "I'll do X for you, first twenty free or half price." Watch who replies.

Weeks 3–4 — lock the repetition down. Prompts and scripts. This is the fork between a business and a job.

Weeks 5–8 — ship the minimum. One landing page, one payment link, one automated delivery flow. No accounts, no dashboard.

Weeks 9–12 — find 10 paying users. If you can't, change the use case, not the product form.

The only hard test: before the product existed, did anyone pay? If yes, keep going. If no, the problem is on the demand side, not the technical side.

07 · Building is solved. Distribution isn't.

Anyone can ship now. Being seen is the part nobody warns you about, and it is where a solo operator is worst off: no ad budget, no content team, no sales reps.

Three paths one person can actually walk:

• Long-tail search. Write a hundred specific questions instead of one definitive essay. This is how directories and tool sites live, and how PDF.ai and Bannerbear got their start.
• Vertical communities. Answer consistently in the three places your users actually are, instead of advertising in thirty.
• Direct outreach. Twenty specific people, one specific email, every week. Low conversion, and the only channel that depends on no platform at all.

Closing

Choosing what to build for a solo business is the inverse of venture capital: don't back the biggest thing, back the smallest thing that sustains itself. One sentence to test against — one person can start it, and once started, it runs without being pushed.

Written November 2026. Ranking data from the HaiAi123 intelligence index, built on Artificial Analysis evaluations (Intelligence Index v5, as of November 7, 2026). Prices are per million tokens and shift with new model releases; for reference only, not a basis for procurement or investment decisions. Revenue figures above are the founders' own public disclosures, reported under differing methods, and should be read as orders of magnitude only.