MIT research shows vendor AI partnerships succeed 67% of the time. Internal builds succeed 33%. Yet most engineering teams default to building. Here’s a framework for making the right call.
The decision matrix
Score each factor 1-5 for your use case:
| Factor | Build (score high if true) | Buy (score high if true) |
|---|---|---|
| Differentiation | AI is your core product | AI is a feature, not the product |
| Data sensitivity | Proprietary data can’t leave your infra | Standard data, DPA is sufficient |
| Customization | Need deep control over model behavior | Standard behavior is fine |
| Speed to market | Can wait 3-6 months | Need it in weeks |
| Team expertise | Have ML/AI engineers on staff | No AI expertise in-house |
| Maintenance budget | Can dedicate 20%+ of eng time | Want zero maintenance |
Score 18-30 for Build: You have the team, the data, and the differentiation to justify building. Score 6-17 for Buy: Buy a solution and focus engineering on your core product.
When to build
Your AI IS the product
If AI is your competitive advantage — your recommendation engine, your coding assistant, your diagnostic tool — you need to own it. Buying means your competitor can buy the same thing.
Examples: AI coding tools, AI-powered search, custom document processing.
You have proprietary data
If your training data or fine-tuning data is your moat, you need to build. Sending proprietary data to a vendor’s API means trusting their data handling and potentially training their models.
Solution: Self-host open-weight models (DeepSeek, Qwen, Devstral) and fine-tune on your data.
You need deep customization
If you need the model to behave in very specific ways — custom output formats, domain-specific reasoning, integration with internal tools via MCP — building gives you full control.
When to buy
AI is a feature, not the product
If you’re adding AI to an existing product (chatbot for support, AI-powered search, document summarization), buy. The AI isn’t what makes your product special.
Examples: Customer support chatbot, internal knowledge search, email summarization.
You need it fast
Building an AI system from scratch takes 3-6 months minimum. Buying takes days to weeks. If time-to-market matters, buy first, build later.
You don’t have AI expertise
Building AI systems requires understanding of prompt engineering, evaluation, observability, and security. If your team doesn’t have this expertise, buying is safer.
The hidden costs
Hidden costs of building
| Cost | Typical range |
|---|---|
| AI engineer salary | $150-250K/year |
| GPU infrastructure | $500-5,000/month |
| Observability tooling | $100-500/month |
| Evaluation pipeline | 2-4 weeks to build |
| Ongoing maintenance | 20% of initial build effort/year |
| Model updates | Quarterly re-evaluation and testing |
Total first-year cost: $200K-500K for a serious AI system.
Hidden costs of buying
| Cost | Typical range |
|---|---|
| API costs at scale | $500-10,000/month |
| Vendor lock-in | Migration costs if you switch |
| Limited customization | Workarounds for missing features |
| Data privacy risk | GDPR compliance depends on vendor |
| Dependency risk | Vendor changes pricing, features, or shuts down |
Total first-year cost: $10K-120K depending on usage.
The hybrid approach (recommended)
Most successful companies don’t purely build or buy. They:
- Buy for speed — use Claude API or GPT API to ship v1 fast
- Evaluate for fit — run for 3-6 months, understand your actual needs
- Build what differentiates — replace vendor components where you need customization
- Keep buying commodities — use APIs for non-differentiating features
Example timeline:
- Month 1-3: Ship with Claude API ($500/month)
- Month 4-6: Identify which components need customization
- Month 7-12: Self-host the custom components, keep API for the rest
- Year 2+: Fully custom where it matters, vendor APIs where it doesn’t
Decision by use case
| Use case | Recommendation | Why |
|---|---|---|
| Customer support chatbot | Buy | Commoditized, many good vendors |
| AI coding assistant | Buy (Claude Code, Cursor) | Extremely hard to build, frontier models needed |
| Internal document search | Hybrid | Buy RAG framework, build custom retrieval |
| AI-powered product feature | Build | Your differentiation |
| Content generation | Buy | API call + prompt, no need to build |
| Compliance/audit AI | Build | Regulatory requirements demand control |
FAQ
Should I build or buy AI?
Buy if AI is a feature (not your core product), you need it fast, or you lack AI expertise. Build if AI is your competitive advantage, you have proprietary data that can’t leave your infrastructure, or you need deep customization. Most companies should start by buying, then selectively build where differentiation matters.
When is building AI worth it?
Building is worth it when AI is your core product, when you have proprietary training data that creates a moat, or when you need behavior that vendor APIs can’t provide. You also need the team (ML engineers) and budget ($200K-500K first year) to sustain it. If you score 18+ on the decision matrix above, building makes sense.
What are the hidden costs of AI?
For building: ongoing maintenance (20% of initial build effort per year), GPU infrastructure ($500-5K/month), observability tooling, quarterly model re-evaluation, and AI engineer salaries ($150-250K/year). For buying: API costs that scale with usage, vendor lock-in migration costs, limited customization workarounds, and dependency risk if the vendor changes pricing or shuts down.
Related: Self-Hosted AI for Enterprise · AI Governance for Startups · Evaluate AI Vendors · How to Reduce LLM API Costs · AI Risk Assessment Template