# How to Build a One-Person Business Using Claude: Dario's 3 Filters > Source: > Published: 2026-09-19 00:00:00+00:00 # How to Build a One-Person Business Using Claude: Dario's 3 Filters Anthropic's CEO says a solo founder could build a billion-dollar business with Claude. Here's the three-filter framework, explained with a real example. ## What did Dario Amodei say about a one-person billion-dollar business? Anthropic CEO Dario Amodei suggested, in response to an interview question, that the first one-person billion-dollar company could emerge this year, built on top of Claude. He didn’t hand out an official checklist, but the examples he used point to three conditions such a business would need: it deploys its own capital, it’s built on software, and its sales and support run without a human doing the repetitive work. Those three conditions work as filters for anyone trying to figure out what to actually build. ## TL;DR - **Dario’s comment wasn’t a formal framework** , it was an answer to an interview question, but the examples he gave (like a proprietary trading firm) can be turned into three practical filters for idea selection. - The first filter, **deploying your own capital** , covers businesses like trading firms, real estate flipping, or car dealerships, where you don’t need many customers but you do need money and risk tolerance. - The second filter is **software** , because tools like Claude Code now let one person build products that used to require a full engineering team. - The third filter is **automated sales and support** , which rules out anything needing heavy customization, compliance overhead, or a human on every deal. - A working example, called **Agent Report Card** , applies all three filters to land on a niche: QA and evaluation software for AI customer support agents. - The business runs on three components: **the core product** (automated agent testing),**a lead-qualification and support layer** (Claude handling routine questions, escalating money and data decisions), and**a lead-generation workflow** (Clay plus Claude drafting outreach). - Getting from working prototype to real revenue still means solving the hardest problem in any solo business: **finding and converting paying customers** , not just building the software. ## What are Dario’s three filters for a one-person business? The framework breaks down into three sequential filters, each one narrowing the field of viable ideas. **Filter one: capital deployment.** The business trades or invests its own money rather than selling a product to customers. Dario’s example was a proprietary trading firm, but the same shape applies to house flipping or used-car reselling. You buy something, improve or trade it, and sell it for more. The advantage is that you don’t need a large customer base or sales team. The catch is you need capital and a real appetite for financial risk, which rules this route out for most people starting from zero. **Filter two: software.** This is where most solo builders land, because tools like Claude Code have made it possible for a non-engineer to build functioning software products. The problem is that “software” is not automatically a good one-person business. Plenty of software products still require custom onboarding, a sales team, and ongoing support that eats up a founder’s time. **Filter three: automated sales and support.** The offer has to be repeatable, need minimal customization, and be easy enough for customers to start using without heavy consultation, regulation, or one-off support tickets. If the support questions can be answered by an AI agent instead of a human, the business clears this filter. Simple, well-scoped products (a file converter, an ad reviewer) tend to pass. Anything requiring a custom sales conversation for every deal tends to fail. Run an idea through all three filters and most ideas fall out. What survives is a narrow category: software products with a clear, explainable value proposition that customers can self-serve. ## How do you pick a specific idea inside that framework? One way to stress-test candidate ideas is to have Claude argue both sides of each one: who would actually pay, whether the result can be delivered by software alone, and whether a single person could realistically sell and support it without help. Three ideas were tested this way in the example covered here: a scheduling tool similar to Calendly, a tool that researches companies and drafts cold outreach, and a tool that stress-tests AI customer support agents before a business puts them in front of real users. The scheduling tool was easy to build but competes in a saturated market full of mature incumbents. The outreach tool was easy to explain but couldn’t prove its emails would actually convert. The third idea, testing AI agents, had the clearest and most provable result: connect an agent, run it through difficult scenarios, and hand back a report showing exactly where it failed. That idea became the working example: **Agent Report Card**, a piece of QA and evaluation software for AI agents. ## What does Agent Report Card actually do? The core insight is that AI agencies deploying customer support bots for clients need proof those bots won’t do something costly, like inventing a policy, issuing a wrong refund, or leaking private data. Without dedicated software, someone has to manually run through test conversations by hand, and every time the underlying prompt or model changes, that testing has to happen again. Agent Report Card connects to a client’s support agent and runs it through a fixed set of test scenarios (in the demonstrated version, 16 tests), covering both normal questions and adversarial ones designed to provoke a risky response, like a billing dispute or a private data request. It scores the agent, flags failures, and can pass failed conversations to Claude to diagnose the root cause and suggest a tighter instruction. Because AI agents are non-deterministic, fixing one failure doesn’t guarantee a perfect score on the next run, which is part of why a larger, repeatable test set matters more than a single pass/fail check. The output is a client-facing report: the score, which tests ran, what changed between runs, and what’s still unresolved. The agency’s private conversations and prompts stay inside its own workspace. The customer isn’t paying for a dashboard, they’re paying for documented proof that an agent was tested before going live. ## How does a one-person business handle sales and support without hiring anyone? The framework splits the operation into three layers beyond the core product: an intake and support agent, and a lead-generation workflow. For intake, a prospect fills out a trial form describing their situation (for example, an agency managing over a dozen agents that has already had one issue an incorrect refund). Claude reads the submission, evaluates fit, and recommends a next step, such as a small trial. The founder still makes the final call before anything moves forward. Claude doesn’t send contracts, charge a card, or make commitments on its own. For support, routine questions (like how to rerun failed tests) get pulled from a product guide and answered directly. Anything involving money or account-level data, like a refund or account deletion request, gets drafted but routed to the founder rather than executed automatically. That split is the operating principle: repetitive questions get automated, judgment calls stay human. For lead generation, a B2B data tool (Clay, in the demonstrated setup) identifies companies publicly deploying AI agents, and Claude checks public sources, explains why a company might be a fit, and drafts an outreach message. Drafts are reviewed and approved by the founder before sending, since finding a plausible lead and writing a decent message isn’t the same as landing a customer. ## Is this actually a path to $1 million or $1 billion in revenue? Nobody should take “billion-dollar business” literally as a near-term outcome for most solo builders. The math also matters: at $499 per month, a product like Agent Report Card would need over 150 active paying customers just to cross $1 million in annual recurring revenue, and getting there requires an outreach funnel that actually converts, which is a much harder problem than building the software itself. ### Built like a system. Not vibe-coded. Remy manages the project — every layer architected, not stitched together at the last second. The realistic first milestone isn’t $1 million, it’s getting a handful of paying customers using the product and generating feedback. Framing the offer clearly helps here: define the specific pain (agencies manually testing agents with no proof of quality), the specific person (an AI automation agency deploying support bots for clients), and the specific promise (a defined set of test scenarios and a client-ready report). Starting narrow, on customer support agents specifically rather than “AI agents” broadly, makes the promise sharper and easier to sell, even though general-purpose AI eval tools already exist with more funding and more customers. ## Frequently Asked Questions ### Did Dario Amodei give an exact three-step method for building a one-person business? No. He answered an interview question about what a one-person billion-dollar company might look like, using examples like a proprietary trading firm. Those examples can be generalized into three filters (own capital, software, automated sales and support), but there’s no official Anthropic framework by that name. ### What tools are needed to build a business like Agent Report Card? The demonstrated stack used Claude for reasoning and diagnosis, Claude Code to build the application, a database to store test history and reports, and Clay for B2B lead data and outreach personalization. ### Why focus on customer support agents instead of all AI agents? Narrowing the scope makes the offer specific and testable. Sales agents, finance agents, and other agent types each need different test scenarios. Targeting customer support agents lets the product get good at a defined set of high-risk situations, like refunds, billing disputes, and data requests, rather than becoming a generic tool competing against better-funded evaluation platforms. ### What’s the hardest part of running a business like this alone? Not the software. Building the product, the intake flow, and the outreach drafts is achievable with current AI tools. The harder problem is the same one every business faces: getting strangers to actually reply to outreach, trust the product, and pay for it. ### Can Claude run the entire business without human oversight? Not in the version described here. Claude handles research, drafting, qualification, and routine support answers, but decisions involving money, refunds, or customer data are routed to the founder rather than executed automatically.