The Great AI Feature Convergence¶
I’ve now seen multiple Series A startups quietly pivot from building AI products to offering AI consulting. Why? Because OpenAI, Anthropic, and Google keep absorbing features faster than startups can differentiate. One day you're building a document AI platform; the next, your core value prop is a checkbox in ChatGPT.
We're even seeing foundation models absorb capabilities that used to be a key differentiator for another foundational model (e.g. reaseach, citations, voice, etc.). Lasting businesses need deep workflow integration, proprietary data, or other moats – AI features alone aren’t enough.
Here's how to spot it coming and what actually survives.
How Convergence Actually Happens¶
Feature convergence in AI follows a predictable pattern, and it's accelerating:
Phase 1: Specialist Solutions Emerge A startup identifies a gap that foundation models can't handle well. They build specialized AI for document analysis, meeting transcription, code generation, or content creation. Early adopters love it.
Phase 2: Foundation Model Providers Take Notice OpenAI, Google, and Anthropic see the demand. They have massive R&D budgets and every incentive to expand their platform capabilities. The writing is on the wall.
Phase 3: Integration and Commoditization The capability gets baked directly into foundation models. What used to require a specialized startup and a separate subscription becomes a standard feature accessible through the main platform.
Phase 4: Market Collapse The specialized startup faces an impossible choice: compete against free, or find a new business model. Most don't survive.
This cycle is happening faster than most founders realize. What used to take years now happens in months.
Case Studies: Features That Got Absorbed¶
Document Analysis and Chat¶
What existed: Startups like ChatPDF, DocuAsk, and dozens of others that let you "chat with your documents"
What happened: GPT-4 added native document analysis, Claude got document upload, Google added file understanding to Bard
Timeline: Document analysis capabilities became standard across foundation models
Result: The standalone document chat market essentially vanished
Real-Time Voice Interactions¶
What's happening now: OpenAI launched real-time voice in October 2024, followed within a few months by Advanced Voice enabling real-time conversations with ChatGPT as thought partner, coach, or interviewer. Claude mobile added voice responses. Perplexity mobile app just rolled out full real-time voice with turn-taking.
The vulnerable companies: Startups building voice-first AI interactions as end-user applications (cf. Wispr Flow, which has a great product as well as additional use cases, but still may face a convergence threat).
The survivors: Companies like Eleven Labs that provide specialized voice infrastructure for developers building systems, rather than competing with foundation models for end-user attention
The timeline: This is convergence happening in real-time across all major foundation model providers
Meeting Transcription and Summarization¶
What's happening now: Zoom added AI summaries, Microsoft Teams integrated M365 Copilot, Google Meet added intelligence features. Then building on their voice technology, OpenAI just launched meeting recording for Pro users in the last few weeks. Guess what happens when ChatGPT Plus (~$20/m) users get this - we're going to see churn at Granola, Otter, Fireflies, Circleback, and many others.
The vulnerable companies: Standalone meeting recorders without deeper workflow integration
The timeline: This convergence is happening right now
Basic Content Generation¶
What existed: AI writing assistants for marketing copy, social media posts, blog outlines
What happened: Every major platform added AI writing features—Canva, Adobe, even PowerPoint
Result: Generic AI writing tools lost their differentiation overnight
The Warning Signals¶
I look for these patterns when evaluating whether an AI startup is in the convergence danger zone:
Signal 1: Single-Feature Value Proposition¶
If you can describe their entire value prop in one sentence starting with "AI that...", they're vulnerable. "AI that summarizes documents" or "AI that writes social media posts" are features, not businesses.
Signal 2: No Proprietary Data Generation¶
They process existing data but don't create new, valuable data through their product usage. No flywheel effects, no network advantages.
Signal 3: API-Dependent Business Model¶
Their core functionality could be replicated by calling the same foundation model APIs they use, just with different prompting or interface design.
Signal 4: Generic Horizontal Appeal¶
They target "everyone" rather than deep vertical integration. "AI for productivity" gets commoditized faster than "AI for surgical workflow optimization."
Signal 5: Feature Parity Competition¶
They compete primarily on AI accuracy or speed rather than workflow integration or switching costs. When the foundation models improve, their advantage disappears.
What Actually Survives Convergence¶
Not everything gets absorbed. Here's what I see maintaining defensible positions:
Deep Workflow Integration¶
Companies that embed AI into complex, industry-specific workflows rather than offering standalone AI tools. The AI becomes part of how work gets done, not just a nice-to-have feature.
Example: Healthcare platforms that integrate AI diagnostics into electronic health records and clinical workflows. The switching cost isn't just losing the AI—it's disrupting the entire practice management system.
Proprietary Data Flywheels¶
Businesses where using the product generates valuable data that makes the AI better over time, and this data can't exist anywhere else.
Example: Manufacturing quality control systems that learn from proprietary sensor data and production processes. This data literally can't be replicated without the same equipment and processes.
Regulatory and Compliance Moats¶
AI applications in heavily regulated industries where compliance requirements create switching costs and barriers to entry.
Example: Financial services AI that's built for specific regulatory frameworks. Foundation models can't just add "banking compliance" as a feature.
Network Effects and Multi-Sided Markets¶
Platforms where AI capabilities improve based on network participation, not just model training.
Example: Marketplace recommendation engines that get better as more buyers and sellers participate. The AI value comes from the network, not just the algorithm.
The Investor Reality Check¶
For VCs evaluating AI investments, here are the questions I recommend:
The Convergence Test: If OpenAI announced they were adding this exact capability to ChatGPT tomorrow, what would happen to this company?
The Differentiation Test: What prevents a well-funded competitor from replicating this in 6 months?
The Integration Test: How deeply embedded is this AI into the customer's existing workflow?
The Data Test: Does using this product create valuable data that can't exist elsewhere?
If you can't get compelling answers to all four questions, you're might be looking at a feature that will get absorbed by foundation models.
The Time Horizon¶
Based on what I'm seeing, the convergence timeline is compressing:
6 months: Basic AI features become standard across foundation models
12 months: Specialized single-purpose AI tools face serious pressure
18 months: Only deeply integrated or proprietary data advantages remain defensible
What This Means for Founders¶
If you're building an AI startup, ask yourself: Are you building a feature or a business?
Features get absorbed. Businesses that happen to use AI can survive and thrive. The key is making AI a means to an end, not the end itself.
The startups that will survive the great convergence are those solving fundamental workflow problems where AI is just one component of a larger solution. They're not AI companies—they're industry solution companies that happen to use AI.
Next in the Series¶
Understanding convergence patterns helps explain why certain AI investments fail, but it raises the crucial question: what actually creates lasting defensive value in the AI era?
Next week, I'll dive into the data moats that survive foundation model disruption—the specific types of data advantages that can't be easily replicated or absorbed.
Evaluating AI investments and need help separating features from businesses? I help VCs and corporate investors navigate these market dynamics.