Microsoft, Google, and Meta launch competing products monthly now. Not yearly. Monthly.
The question isn't "can we build this?" anymore. It's "what do we build and why will it win?"
That question got harder. Moats that lasted 6-12 months now survive 2-3 weeks. The answer is treating AI like Lego blocks - your advantage comes from how you assemble them with your unique stuff, not from the blocks themselves.
I learned this lesson the hard way. The temptation to build custom models is real - fine-tuning on your corpus, custom preprocessing, optimised inference. It feels like a moat. But foundation models improve on a different cadence than product features. What's impressive today is baseline next quarter. The competitive advantage becomes a maintenance burden that costs more to run than the API alternatives.
The mistake isn't technical. It's strategic. Confusing "can we build this?" with "should we build this?" The moat isn't in the model. It was never going to be in the model.
The Lego Framework
Three layers:
AI Capabilities (commodity) -----
\
Your Data (differentiation) -------> Differentiated Product -> Unique Output
/
Your Functionality (control) -----
AI Capabilities: Everyone Has These
Standard models from OpenAI, Anthropic, Google. Transcription from Deepgram. Text, image, audio processing. Universally available. Not a moat.
Your Data: The Actual Moat
Proprietary data gives context for unique outputs. What matters is marginal advantage over what's already in foundation models.
- Real-time data not in training sets
- User behaviour data specific to your product
- Domain expertise (legal, healthcare, finance)
- Human judgement data around curation and quality
- Reinforcement data from user interactions
Marginal value matters more than volume. Ask: "How much extra value does our data provide over what's already in the base model?"
Your Functionality: How It Behaves
This determines how AI acts inside your product. The "superpowers" that create unique experiences.
- Specialised workflows for your users
- Business rules specific to your domain
- Integrations with existing tools
- UI patterns that control AI interaction
Granola: Finding a Seam in a Saturated Market
Granola entered AI note-taking dominated by Otter, Fathom, Fireflies, Zoom, Teams. Crowded as hell. They still got 40% awareness among tech professionals and AUD $30 million in funding within 12 months.
The Insight Everyone Missed
Competitors focused on replacing note-taking ("we'll take your notes for you"). Granola saw that users wanted to be empowered to take better notes, not have the task removed.
Gap between automated transcription and human-guided documentation.
Data Flywheel
- User-generated notes during meetings
- AI-enhanced transcription from system audio
- Contextual metadata from calendar integrations
- Improvement data from user edits
Unique repository that gets better over time. Enables cross-meeting search, project workspaces, automated actions.
Functionality That Differentiates
- Guided note capture, not full automation
- Human-in-the-loop refinement with AI suggestions
- Cross-meeting intelligence connecting related discussions
- Action item extraction tied to calendar and task systems
Two Failure Modes
| Failure Mode | Symptom | Why It Happens |
|---|---|---|
| Reinvention trap | Building custom models when off-the-shelf works | Engineering pride, NIH syndrome |
| Copy-paste trap | Same Lego stack as competitors | No unique data or functionality, competing on UI polish |
Reinvention wastes cycles on solved problems. Copy-paste creates undifferentiated products competing on price.
The Reinvention Trap in Practice
The reinvention trap usually looks like this: a team spends months building a custom model, proud of the accuracy numbers, the custom feature engineering, the hyperparameter tuning. Then someone runs the same data through a commercial API and matches or exceeds the results with zero training, zero maintenance, zero infrastructure.
The demoralisation is real. But the mistake happened earlier - nobody asked "what's the marginal value of building this ourselves?" They'd assumed custom meant better. Sometimes it does. Usually it doesn't. The question isn't whether you can build it. It's whether building it creates value that buying it doesn't.
Platform Primitives in Practice
When building an AI platform, we faced the reinvention question directly. Do we build a custom gateway, or buy one? Do we write our own model routing, or use an existing solution? The temptation to build was strong - we understood the problem deeply, and off-the-shelf solutions never fit perfectly.
We chose a hybrid: buy the gateway, build the SDK. The gateway handled model routing, failover, cost tracking - commodity capabilities where buying beat building. The SDK abstracted sessions, memory, tracing, and agent patterns - our unique layer that absorbed vendor pivots without rewriting applications.
This proved valuable when we later switched gateway vendors. The applications on top didn't change. The SDK absorbed the migration. The primitive - a consistent interface for AI interactions - outlasted the implementation.
Build where your unique context matters. Buy where the problem is already solved. The boundary moves as the market matures.
The Copy-Paste Trap in Practice
Copy-paste competition is everywhere in AI right now. Same GPT-4 backend. Same "rewrite this paragraph" feature. Same Chrome extension. Same pricing. The only differentiation is landing page colour and founder Twitter following.
The ones that survive don't win on AI. They win on workflow integration - the specific CMS their users can't live without, the approval chain that maps to how the team actually works, the data format that removes a manual export step. The AI is identical. The moat is the workflow.
Copy-paste competition is a race to the bottom on price. The winner is whoever has the most runway, not whoever has the best product. Avoid it by finding the seam that everyone else missed - the workflow your users spend 80% of their time in, the integration that removes friction they didn't know they had.
Moat Timeline
Month 0-1: Launch with unique data + standard AI (temporary feature advantage)
Month 2-3: Competitors copy (moat compresses)
Month 3-6: Data flywheel creates improvement (defensible if data is unique)
Month 6-12: Workflow integration deepens (switching costs emerge)
Month 12+: Platform effects from user content (network moat)
Each layer depends on the previous. No unique data means workflow integration is just UI preference. No workflow integration means network effects never happen.
Summary
- AI capabilities are commodities. Same models, same APIs.
- Your data is the first moat. Focus on marginal value, not volume.
- Your functionality is the control layer. Workflows and integration create switching costs.
- Avoid reinvention. Don't build what you can buy.
- Avoid copy-paste. Same assembly as competitors means price competition.
What's Next
Audit your product: what percentage of AI differentiation comes from unique data and workflows versus UI polish on commodity capabilities? If it's mostly polish, your moat is thinner than you think.