How to build a truly profitable mini AI company
Shen Translation Bureau is a translation team under 36Kr, focusing on technology, business, workplace, life and other fields, with a focus on introducing new technologies, new perspectives and new trends from abroad.
Editor's Note: Stop obsessing over Silicon Valley's "financing and cash-burning" myths. Medvi, which generates $1.8 billion in annual revenue with just two brothers and AI, proves that using AI to revamp those tedious, slow traditional businesses is the most lucrative wealth-building shortcut for ordinary people. This article is a translated piece.
I believe we are on the path to birthing the first "one-person billion-dollar company". That day has not truly come yet, but two-person billion-dollar companies have already emerged, all built with the help of AI.
—— Dario Amodei, Chief Executive Officer of Anthropic
In the past, the concept of a "one-person billion-dollar company" sounded like a fairy tale. Now, the CEO of Anthropic claims that there is a 70% to 80% probability of this happening this year (2026).
This sounds crazy, but you will not think so once you look at what has already happened.
This article breaks down what this AI Agent revolution means, what lessons we can learn from the first "two-person billion-dollar company", and how you can replicate this playbook on a smaller scale.
You will learn:
Why micro AI companies can generate huge revenues nowadays
What new business model a two-person company with $1.8 billion in annual sales reveals
Why the biggest opportunities lie in ordinary businesses with "extraordinary profit margins"
How to find those tedious workflows that people are already paying for
How to sell and deliver results manually before automating them
How to acquire an existing online business and upgrade it with AI
The 4 leverage points where AI can truly generate leverage effects
Ready?
Let's get started.
A Two-Person Company That Earns $3 Million Per Day
Matthew Gallagher built a company in two months, which now reaches $1.8 billion in annual sales.
His company Medvi sells GLP-1 weight-loss drugs directly to consumers (DTC) through telehealth.
The code was written by AI, the website was built by AI, the ads were created by AI, and customer service is also handled by AI.
First month: 300 customers.
Second month: 1,000 new customers.
First full year: $401 million in sales.
Profit last year: $65 million.
Daily revenue exceeds $3 million.
Who is the only employee he hired? His younger brother.
That is the entire lineup of this company.
The automated operation model of this company has drawn strict scrutiny and class-action lawsuits. However, there are still many valuable experiences to extract from it.
Even If You Do Not Approve of This Business, The Experiences Worth "Learning Stealthily" From Medvi
1. Own the customers, rent the backend
The backend of most markets is already mature — suppliers, manufacturers, fulfillment and delivery, compliance reviews, and various platforms are all readily available. The bottleneck is almost always customer acquisition (that is, distribution channels).
If you can capture the customer acquisition link at a lower cost and faster speed than existing industry players, you can build directly on top of other people's infrastructure and take the profit margin for yourself.
2. Gold mines are hidden in slow, bloated markets
Medvi succeeded because Matthew found a market with clear demand but extremely high friction. Then he used AI to cut out manpower, coordination and waiting time.
Look for such markets: customers are already paying money, but the way companies operate is still stuck in 2016 — slow responses, manual onboarding, weak follow-up, high fulfillment costs, bloated teams, poor website experience, and lack of personalization.
On a smaller scale, this could mean buying a running online business and upgrading this "money-making machine" with AI.
The $10,000 Monthly Mini Version of the "Billion-Dollar AI Playbook"
You have two paths to choose from.
Path A: Build From Scratch
The easiest way to start a business with AI agents is: first use AI behind the scenes to sell a useful result in the form of manual service; after confirming that people are willing to pay for it, then automate the workflow.
Let's go through it step by step.
Step 1: Choose a vertical workflow that already has a budget
Look for workflows where people are already spending time, or paying to hire employees, freelancers, outsourcing agencies, assistants, buying software, or investing in manual sorting.
A good problem usually has the following 4 characteristics:
It occurs frequently enough to make people feel real pain
It consumes time, money, speed, accuracy or revenue
It is complicated enough that traditional software has not been able to solve it well so far
Easy to review: AI can generate the first draft, then humans check and correct errors
Ideally, you are looking for a tedious workflow that relies on manpower behind it:
Invoice and bill processing
Customer support ticket triage
Sales lead screening and evaluation
CRM (Customer Relationship Management) data cleaning
Administrative work for customer onboarding
Document information extraction
Proposal and quotation drafting
Initial resume screening for recruitment
Sales background research
Customer report production for marketing agencies
Your entry point must be very specific:
"I use 'AI + human review' to help [specific group of people] reduce [specific repetitive workflow] from [current cost/time] to [better result]."
For example:
"I help freight brokers convert bill emails from carriers into neat, payable records within 24 hours, without the need for dedicated staff to copy PDF content into spreadsheets one by one."
Tips: Reddit is a powerful tool for this kind of research.
You can search for the following keyword combinations (search in English):
"[Niche field] + I hate"
"[Niche field] + takes forever"
"[Niche field] + still doing this manually"
"[Niche field] + spreadsheet"
"[Niche field] + any tool for this"
"[Niche field] + how do you handle"
What you are looking for is: different people complaining about the same pain point repeatedly in different months, and there is no decent solution in the comment section.
Step 2: Verify the pain point with workflow evidence
Find 20 to 30 people who are currently troubled by this problem. Send them private messages, leave comments under their posts, ask questions in communities, or directly contact business founders.
The script can be like this:
"Hi, I saw your message saying that you have to manually organize sales call records every time after talking to customers. I would like to ask: how exactly do you handle this now?"
You are looking for evidence that proves this workflow pain point really exists:
They will explain their current approach to you
They will show you their spreadsheets, inboxes, CRM, call records, documents or specific processes
They admit that the current approach is slow, annoying, costly, or error-prone
They are already spending money on hiring people or buying tools to solve this problem
Step 3: Sell the Paid Pilot
Go back to those people who showed you their workflows, and propose a paid pilot plan to them:
If people are willing to pay for the manually delivered version, it means that the demand is real.
Important Note: Charge high enough to verify that the other party is truly determined.
Step 4: Deliver manually first, then automate repetitive links
Use AI behind the scenes first to complete the result delivery manually.
Then observe the entire workflow carefully:
Which step is repeated every time?
Which step takes the longest time?
Which step is the most error-prone?
Which step do customers care most about?
These are the parts you need to prioritize for automation.
Keep the human review layer before the quality is stable and reliable.
Step 5: Decide whether to keep it or drop it
After completing the first 5 to 10 paid pilots, review the data.
Ask yourself the following questions:
Did customers pay readily without you having to persuade them hard?
Did they come back for repurchases, or put forward more demands?
Did the delivered results save them time, reduce errors, or generate revenue?
Can you deliver the results with a healthy profit margin?
Can the workflow become more automated as you serve one more customer?
Do you have a clear path to reach more buyers of the same type?
If all the answers are yes, turn it into a "productized service" and start building the automation system.
Path B: Acquire and Upgrade
Don't want to build from scratch?
You can directly buy a business that already has existing demand.
Go to trading platforms such as Flippa, Acquire or Empire Flippers to take a look.
Search for projects with verified demand.
Focus on finding targets with the following characteristics:
Existing revenue
Returning customers
Email subscription list
Organic SEO search traffic
Simple and clear product or service model
Backward operations
Slow customer response
Poor content production system
Rudimentary customer onboarding process
Weak follow-up
Manual fulfillment
The ideal target is a very boring business where the original owner has become slack, tired, or hit a bottleneck.
Important Note: Do not pay for "potential".
"This site will definitely make money if someone manages it properly" — that is not the target you are looking for. What you need is a site that is already making money but has obvious flaws in its operation system. AI should not be the reason why this business can survive, but the reason why it can become more profitable.
Your Acquisition Due Diligence Checklist
1. Veto Items (Must-pass Questions)
As long as one answer is "No", turn around and leave.
Is it already making money now?
Do I clearly know where its traffic comes from?
Is the revenue data real and verifiable?
Can I quickly understand this niche industry?
Can I clearly state which specific part will get better after accessing AI?
2. Growth Potential Items (Bonus Questions)
The more "Yes" answers you get here, the better the opportunity.
Is the traffic stable or growing?
Is there an email list?
Are there repeat customers?
Is the original owner doing everything personally and overly dependent on individuals?
Is fulfillment and delivery done manually?
Are customers asking the same questions over and over again?
Is there obvious upsell space?
Can AI improve profit margins or conversion rates within 30 days?
Am I willing to invest 12 months in this business myself?
A Simple Rule
All answers to the "veto items" must be "Yes".
At the same time, there must be at least 5 "Yes" in the "growth potential items".
The "veto items" ensure that what you buy is a real operating business.
The "growth potential items" prove that this business has enough flaws for AI to fix.
Case: From a Half-dead Affiliate Marketing Blog to a Micro Media Business
Last year, I bought a small website about [content omitted].
The original state of this business was like this:
10,000 to 20,000 visitors per month
Earn $700 per month through affiliate links
No email list
Outdated articles
No own products
No real conversion funnel
The upgrades I did with AI were very simple:
Updated all articles that have search rankings
Added a test quiz titled "Which tool is best for your workflow"
Started collecting visitor emails
Created a paid remote work configuration guide
Launched a weekly email newsletter
Created comparison pages for various tools
Used AI to generate outreach emails to secure sponsorships from tool companies
Adapted the articles and distributed them to social media platforms such as X and LinkedIn
Now, this website is no longer just an affiliate marketing blog, but has become a business integrating small media, lead generation and own products.
#Why "Small, Simple and Profitable" Beats "Large and Chaotic"
Start with a streamlined company.
This means keeping only the following elements:
A clear and specific service plan
One landing page
One payment collection system
One customer acquisition channel
One AI agent workflow