ProseAI
AI writing assistant that matches your voice and style for blogs, emails, and docs
Building an AI writing tool in 2023-2024 without a defensible moat beyond 'better output quality.' Quality is temporary — foundation model providers improve constantly, and they have billions of dollars of compute behind them. Our fine-tuning advantage lasted about 6 months before the base models caught up. We should have built the moat in distribution (integration into a specific workflow) or data (proprietary training data for a niche industry), not in model quality.
4/10 revival potential
Timeline
The story
What was built
ProseAI was a SaaS writing assistant that went beyond generic AI writing. Users uploaded samples of their existing writing (blog posts, emails, documentation) and the tool learned their voice — tone, vocabulary, sentence structure, level of formality. Then they could use ProseAI to draft new content that sounded like them, not like ChatGPT. We built a Chrome extension for writing in Gmail and Google Docs, a web editor, and an API for power users. The style-matching was our technical differentiator — we fine-tuned models on each user's writing samples for genuinely personalized output.
Why they built it
When ChatGPT launched, we immediately saw the problem: AI-written content sounds like AI. Every blog post, email, and LinkedIn post generated by ChatGPT has the same bland, hedging, corporate tone. We thought there was a real market for AI writing that actually sounds like the person using it. Content creators, marketers, and executives would pay for AI that wrote in their voice. We had ML experience and thought the style-transfer problem was solvable.
What worked
The style matching was genuinely good. After uploading 10+ writing samples, the output sounded noticeably more like the user than generic ChatGPT output. Power users — especially content marketers producing 10+ blog posts per month — loved it and said it saved them 5-10 hours per week. We hit $8,200 MRR in month 5, which felt like product-market fit. The Chrome extension for Gmail was surprisingly popular — people used it for email drafting more than blog writing.
What failed
In early 2024, three things happened in quick succession. First, OpenAI launched Custom GPTs and the 'memory' feature, which let ChatGPT learn your preferences and style over time — for free. Second, Google added AI writing features directly into Gmail and Docs via Gemini. Third, Claude, Gemini, and Copilot all improved their writing quality dramatically, closing the gap with our style-matching. Our $50/month tool was suddenly competing with free features built into the platforms people already used. Churn spiked from 5% to 18% per month. New signups dropped because people tried ChatGPT's Custom GPTs first and found them 'good enough.' We went from 164 paying users to 60 in three months. The unit economics collapsed — our per-user fine-tuning costs were $8/month, and with 60 users at $50 the margin wasn't enough to sustain three founders.
What was validated
The style matching was genuinely good. After uploading 10+ writing samples, the output sounded noticeably more like the user than generic ChatGPT output. Power users — especially content marketers producing 10+ blog posts per month — loved it and said it saved them 5-10 hours per week. We hit $8,200 MRR in month 5, which felt like product-market fit. The Chrome extension for Gmail was surprisingly popular — people used it for email drafting more than blog writing.
Key lesson
Building an AI writing tool in 2023-2024 without a defensible moat beyond 'better output quality.' Quality is temporary — foundation model providers improve constantly, and they have billions of dollars of compute behind them. Our fine-tuning advantage lasted about 6 months before the base models caught up. We should have built the moat in distribution (integration into a specific workflow) or data (proprietary training data for a niche industry), not in model quality.
Failure analysis
What the signals looked like
The style matching was genuinely good. After uploading 10+ writing samples, the output sounded noticeably more like the user than generic ChatGPT output. Power users — especially content marketers producing 10+ blog posts per month — loved it and said it saved them 5-10 hours per week. We hit $8,200 MRR in month 5, which felt like product-market fit. The Chrome extension for Gmail was surprisingly popular — people used it for email drafting more than blog writing.
Where it actually broke
In early 2024, three things happened in quick succession. First, OpenAI launched Custom GPTs and the 'memory' feature, which let ChatGPT learn your preferences and style over time — for free. Second, Google added AI writing features directly into Gmail and Docs via Gemini. Third, Claude, Gemini, and Copilot all improved their writing quality dramatically, closing the gap with our style-matching. Our $50/month tool was suddenly competing with free features built into the platforms people already used. Churn spiked from 5% to 18% per month. New signups dropped because people tried ChatGPT's Custom GPTs first and found them 'good enough.' We went from 164 paying users to 60 in three months. The unit economics collapsed — our per-user fine-tuning costs were $8/month, and with 60 users at $50 the margin wasn't enough to sustain three founders.
Lessons
What the founder learned
If your product is an AI wrapper, your timeline to commoditization is 6-18 months. Every major AI capability gets absorbed into the platforms: writing assistance goes into Gmail/Docs, code assistance goes into IDEs, image generation goes into design tools. Your moat cannot be 'better AI output' because foundation model providers will match you. The only defensible positions for AI startups are: (1) proprietary data the models don't have, (2) deep integration into a specific workflow that platforms won't build, (3) industry-specific compliance/security requirements, or (4) being the platform itself. Also, MRR growth during the AI hype wave is not product-market fit — it's a sugar rush. Watch churn, not growth, to know if you have a real product.
What they’d do differently
We'd pick one vertical and go all-in. Instead of 'AI writing for everyone,' we'd build 'AI writing for real estate agents' or 'AI writing for healthcare compliance documentation.' Vertical-specific means the model needs domain knowledge the general platforms don't have, the output needs industry-specific formatting and terminology, and the buyer (a company, not a consumer) has real budget. We'd also charge 5x more to fewer customers — $250/month to agencies, not $50/month to individuals — because the enterprise buyer is less likely to churn to ChatGPT.
Editorial scorecard
How viable is rebuilding this today?
Did real users or customers want this?
How well was it built and shipped?
Did they have a path to reach users?
Was the business model viable?
How useful is this postmortem for other builders?
Scores are assigned by App Graveyard editors after review. They are directional, not scientific.
Rebuild opportunity
4/10General AI writing tools are a dead category. The opportunity is in vertical-specific AI writing where the content has compliance requirements, industry jargon, or workflow integration that general tools can't match: legal brief drafting, medical documentation, financial reporting, real estate listings. Each vertical is its own business with domain-specific training data, specific formatting needs, and buyers who pay enterprise prices. The general 'write anything' tool is free now. The specific 'write this industry document correctly' tool is worth real money.
Revive this app
The founder is open to revival interest. App Graveyard has not verified ownership, asset claims, pricing, or availability yet. This is an interest signal, not a transaction.
Contact through App Graveyard
We review revival interest before anything is forwarded. The founder's private contact details are never shown publicly.
Related postmortems
Built something that didn't work out?
Every failed app has a lesson. Submit yours and help the next builder avoid your mistake. Anonymous submissions welcome.