{"slug": "i-tried-to-build-a-company-where-ai-autonomously-finds-wins-and-delivers-work-s", "title": "I Tried to Build a Company Where AI Autonomously Finds, Wins, and Delivers Work. Here's Why It Failed.", "summary": "A developer spent six months building an AI-powered system to autonomously find, win, and deliver freelance work, but despite a fully functional pipeline and 300+ tests, it generated zero contracts and a net loss. The experiment revealed that AI-discovered opportunities lacked real human connections, and anonymity constraints plus platform risks further hindered success.", "body_md": "\"Just tell Claude Code what to do, and let AI handle everything — finding leads, qualifying them, doing the work, and delivering it.\"\n\nThat was the idea I spent about half a year trying to build, on my own. Spoiler: the system worked. All the tests passed. And I still ended up with **zero contracts, zero dollars earned, and a net loss**.\n\nIn this post I'll walk through the actual architecture I built, what happened across two experiment cycles, and why \"being able to build it\" turned out to be a completely different problem from \"being able to sell it.\" I'm writing this in the hope that it saves someone else from repeating the same mistake.\n\nI designed a 4-stage pipeline:\n\n```\nResearch Agent (find opportunities)\n      ↓\nQualification Agent (score them)\n      ↓\nDelivery Agent (execute & deliver)\n      ↓\nLedger (log everything, track KPIs)\n```\n\nThis agent searches for potential work based on a configurable domain, keywords, geography, and criteria. One important design decision: none of these were hardcoded. I wanted to be able to pivot the target domain without touching the code, since I fully expected the first domain to fail.\n\nEvery candidate opportunity gets scored on:\n\nI iterated on this scoring logic quite a bit. For example, I changed the risk-score weight from `-3` to `-5` partway through, so that an extremely risky candidate could single-handedly sink an otherwise strong score below the pass line.\n\nOnce a job was won, this agent used the Claude Code CLI as the execution engine to produce the deliverable. I deliberately avoided metered API billing and ran everything within an existing Claude Code subscription — I didn't want to rack up AI costs before earning a single dollar.\n\nEvery opportunity, from discovery to outcome, gets logged. To avoid the trap of \"this feels like it's working,\" I tracked exactly six KPIs:\n\n| KPI | Description | \n|---|---|\n| Opportunities discovered | Number of candidate jobs found | \n| Jobs completed | Number of jobs actually finished | \n| Revenue generated | Cash actually received (not invoices, not verbal agreements) | \n| AI cost | Cost of AI usage | \n| Human time | Hours of human involvement | \n| Revenue per human hour | Revenue ÷ human hours | \n\nDeliberately excluding vanity metrics like signups or pageviews was, in hindsight, the right call. But as you'll see below, having the right metrics doesn't help much when the actual number of jobs is zero.\n\nI also wrote 300+ tests. Engineering-wise, this system was solid.\n\nIn the first cycle, I let the Research Agent autonomously search for opportunities from scratch. Results:\n\nDigging into why all 10 failed, every single one fell into at least one of these buckets:\n\nIn short: the AI found opportunities that looked plausible on paper, but none of them connected to a real person who was actually struggling with the problem and willing to pay to solve it.\n\nAfter Cycle 1, I flipped the division of labor:\n\nThis was a big shift — from \"AI does discovery\" to \"discovery is human, AI only reinforces.\"\n\nWhile comparing platforms at this stage, I hit a wall I hadn't anticipated: **wanting to stay anonymous (no real name, no face photo) massively narrowed the field of usable platforms.**\n\n| Requirement | Outcome | \n|---|---|\n| Real name + photo mandatory | Excluded | \n| Pseudonym/business name OK, but requires ID + selfie verification | Skipped — too much friction | \n| Anonymous-friendly, lightweight verification | Kept as a candidate | \n\nI also concluded that having AI autonomously crawl sites to find jobs was in a legal/ToS gray zone. In fact, one platform I was evaluating later discontinued its \"auto-propose\" feature entirely, switching to a \"human must review before proposing\" model. I took that as external confirmation that fully autonomous bot applications carry real platform risk.\n\nWhat I ended up building instead: save search filters, receive email notifications for new listings, and have the AI read and analyze those notifications — never touching the site directly. It avoids ToS risk while still using AI for the analysis part.\n\nEven with this more careful setup, Cycle 2 also ended in zero contracts and zero revenue.\n\nTechnically, everything worked exactly as designed — the qualification logic, a \"Policy Gate\" that automatically filtered out sites with no-solicitation notices or bot bans, and the logging pipeline all functioned correctly.\n\nThe problem was never engineering. **Winning work is not a problem software can solve.**\n\nWhen a person or company decides to hire someone, the deciding factor isn't \"is this AI technically impressive?\" It's \"can I trust this person/company to actually deliver?\" Trust comes from track record, referrals, and prior relationships — not something an automated system can manufacture on your behalf.\n\nLooking back, I was essentially betting on:\n\nBuild it → put it out there → someone finds it\n\nAnd the \"someone finds it\" part was where the whole plan broke down. Interestingly, I'd hit the exact same wall before, with a content site that depended on search traffic, and with several mobile apps I'd shipped — same structural problem every time. I could build things. I had never actually designed a distribution channel.\n\nThe original vision — \"AI autonomously wins its own business\" — is on hold indefinitely.\n\nBut the engineering work itself — designing and implementing a full discover → qualify → execute → log pipeline — wasn't wasted. I'm now repositioning that same technical capability: instead of using it to run my own autonomous business, I'm offering it as implementation work for companies that want to build AI-driven automation into their own operations.\n\nConcretely, this includes:\n\nIf you're trying to figure out how to actually wire AI into a real workflow — beyond just prompting a chatbot — feel free to reach out.\n\nBuilding something that automatically earns you money is a genuinely exciting idea for any indie developer. And with today's AI tooling, building a technically functional system is very achievable solo.\n\nWhat I underestimated was the gap between \"a system that correctly does what it's designed to do\" and \"a system that convinces someone to actually pay for it.\" That gap isn't closed with more code — it's closed with trust, built the slow way.\n\nHope this is useful to anyone chasing a similar idea.", "url": "https://wpnews.pro/news/i-tried-to-build-a-company-where-ai-autonomously-finds-wins-and-delivers-work-s", "canonical_source": "https://dev.to/zerozeroyoutube/i-tried-to-build-a-company-where-ai-autonomously-finds-wins-and-delivers-work-heres-why-it-5f3c", "published_at": "2026-09-08 06:12:03+00:00", "updated_at": "2026-09-08 06:31:00.362977+00:00", "lang": "en", "topics": ["ai-agents", "ai-products", "developer-tools"], "entities": ["Claude Code"], "alternates": {"html": "https://wpnews.pro/news/i-tried-to-build-a-company-where-ai-autonomously-finds-wins-and-delivers-work-s", "markdown": "https://wpnews.pro/news/i-tried-to-build-a-company-where-ai-autonomously-finds-wins-and-delivers-work-s.md", "text": "https://wpnews.pro/news/i-tried-to-build-a-company-where-ai-autonomously-finds-wins-and-delivers-work-s.txt", "jsonld": "https://wpnews.pro/news/i-tried-to-build-a-company-where-ai-autonomously-finds-wins-and-delivers-work-s.jsonld"}}