The build vs buy decision for AI in 2026: my framework after advising startups

Q: When should a company build custom AI instead of buying?

When three conditions are met: you have proprietary data that creates a genuine moat (not just 'we have lots of data' but 'our data captures something no public dataset does'), your use case requires accuracy that general-purpose models can't match (domain-specific terminology, unusual data distributions), and you can sustain the engineering investment ($200K-500K/year for an ML team).

Q: What is the most common build-vs-buy mistake?

Companies building custom models for commodity tasks. I have seen teams spend 6 months building a custom text classifier when OpenAI's API handles the same task at 95% accuracy for $50/month. The custom model reached 97% accuracy — a 2% improvement that cost $300K in engineering time and ongoing maintenance. The API was the correct choice.

Q: How do I know if my data creates a real AI moat?

Three tests: (1) Would a competitor with access to your data and a good ML engineer build a meaningfully better product? If yes, your data is a moat. (2) Can you replicate your data's value using public datasets or synthetic data? If yes, it is not a moat. (3) Does your product generate data that makes your model better over time (a data flywheel)? If yes, the moat compounds.

8 minute read

TL;DR

The build-vs-buy decision for AI follows a clear pattern. Buy the API when AI is a feature, not the product. Build custom when you have proprietary data that creates a moat and enough volume to justify the engineering investment. Hire when you need the capability but lack the team. The mistake most companies make is building when they should be buying — and the mistake most AI startups make is buying when they should be building.

Three diverging paths of different materials — circuit traces, concrete, and fiber optics — representing the build, buy, and hire options for AI strategy

Why is this decision different for AI?

Traditional build-vs-buy applies to software: build if it’s a core differentiator, buy if it’s a commodity. AI adds a wrinkle: the quality of the “buy” option is unprecedented. GPT, Claude, and Gemini are general-purpose reasoning engines that handle 80% of text-based AI tasks at 90%+ accuracy out of the box. No previous generation of software offered this.

The result: the bar for “build” is much higher than it was three years ago. Building custom AI only makes sense when you’re solving a problem that general-purpose models handle poorly — and you can prove it, not just believe it.

The framework: three paths

flowchart TD
    Start[AI capability needed] --> Q1{Is AI the<br/>core product?}
    Q1 -->|No, it's a feature| BUY[Buy the API]
    Q1 -->|Yes| Q2{Do you have<br/>proprietary data?}
    Q2 -->|No| Q3{Can you acquire it?}
    Q3 -->|No| BUY2[Buy API +<br/>prompt engineering]
    Q3 -->|Yes, but expensive| HIRE[Hire ML expertise<br/>to evaluate first]
    Q2 -->|Yes| Q4{Does custom model<br/>beat API by > 10%<br/>on your task?}
    Q4 -->|Yes| BUILD[Build custom]
    Q4 -->|No| BUY3[Buy API.<br/>Data moat isn't<br/>translating to accuracy.]
    Q4 -->|Haven't tested| EVAL[Run a 2-week<br/>evaluation first]
    
    style BUY fill:#4a4a8a,stroke:#333,color:#fff
    style BUY2 fill:#4a4a8a,stroke:#333,color:#fff
    style BUY3 fill:#4a4a8a,stroke:#333,color:#fff
    style BUILD fill:#2d5016,stroke:#333,color:#fff
    style HIRE fill:#8b6914,stroke:#333,color:#fff
    style EVAL fill:#8b6914,stroke:#333,color:#fff

Path 1: Buy the API

When: AI is a feature, not the product. Customer support chatbot, document summarization, ticket classification, content moderation, search enhancement.

What it costs: $50–5,000/month in API fees depending on volume. Zero engineering investment beyond integration.

What you give up: Control over model behavior, data ownership (terms vary by provider), and the ability to differentiate on AI quality. You’re using the same model as your competitors.

The 2026 reality: For commodity AI tasks (chatbot, summarization, classification), the API is the correct choice 90% of the time. The team that spends 6 months building a custom classifier to beat GPT-4o by 2% has spent $300K on a 2% improvement. Unless that 2% directly translates to revenue, the API wins.

Path 2: Build custom

When: Three conditions are all true: (1) proprietary data that creates a genuine moat, (2) domain-specific accuracy requirements that general models miss, and (3) engineering team capable of sustaining the investment.

What it costs: $200K–500K/year minimum for a small ML team (1 ML engineer, 0.5 data engineer, infrastructure). Data annotation, training compute, and evaluation infrastructure add another $50K–150K/year.

What you gain: A model that improves with your data over time. Accuracy on your specific task that no competitor can replicate without your data. Independence from API provider pricing and terms changes.

The test: Run a 2-week evaluation. Fine-tune an open-source model on your data. Compare it to the API on 500 examples from your actual production distribution. If the custom model beats the API by more than 10% on the metrics that matter to your business, build. If not, the data moat isn’t translating to accuracy — buy the API.

Path 3: Hire expertise

When: You need AI capability but don’t have the team to evaluate the options, and the investment is large enough to justify external guidance.

What it costs: $3K–10K for a diagnostic assessment. $8K–15K/month for ongoing fractional CTO work.

When this is the right first step: When you’re considering a six-figure AI investment and your team doesn’t include someone who has shipped a production ML system. The cost of the assessment is trivial compared to the cost of building the wrong thing.

The mistakes I see most often

Non-AI companies building when they should buy. A logistics company spent 8 months building a custom route optimization model. Off-the-shelf optimization APIs (Google OR-Tools, commercial routing APIs) handled 95% of their use case. The custom model added 3% accuracy on edge cases. The cost: $400K in engineering time plus ongoing maintenance. The API cost: $2,000/month.

AI startups buying when they should build. A voice AI startup used OpenAI’s API for all inference, paying $15/million tokens. At 100 million tokens per month, that’s $1,500/month — manageable. At their projected 10 billion tokens per month, it’s $150,000/month. Self-hosted inference at that volume would cost $10K–20K/month. They hadn’t modeled the cost curve.

Everyone skipping the 2-week evaluation. The build-vs-buy decision should be evidence-based, not intuition-based. A 2-week evaluation (fine-tune a model, compare to API, measure on production data) costs $5K–10K in engineering time. It prevents $200K+ mistakes. I’ve never seen a team regret running the evaluation. I’ve seen many regret skipping it.

The data moat test

“We have a lot of data” is not a moat. Three questions determine whether your data creates genuine competitive advantage:

Would a competitor with your data build a better product? If a new entrant could license your data type from a third party (public datasets, data brokers, synthetic generation) and match your model quality, your data is not a moat. It’s a head start that erodes.

Does your product generate data that improves your model? A data flywheel — where the product creates data that trains the model that improves the product — is the strongest form of data moat. But it only works above certain volume thresholds and with explicit feedback mechanisms.

Is the data’s value in its content or its labels? Raw data is cheap. Labeled data with domain expert annotations is expensive and hard to replicate. Medical transcription data with physician corrections, legal discovery data with attorney classifications, customer support logs with resolution outcomes — these are moats because the labels encode expertise that’s expensive to reproduce.

Key takeaways

Buy the API when AI is a feature, not the product. For commodity tasks (chatbot, summarization, classification), the API is correct 90% of the time.
Build custom only when you have proprietary data that creates a moat AND your custom model beats the API by more than 10% on production data. Both conditions must be true.
Run a 2-week evaluation before committing to build. It costs $5K–10K and prevents $200K+ mistakes.
“We have a lot of data” is not a moat. Test whether the data translates to accuracy advantage over general-purpose models.
The mistake most companies make is building when they should buy. The mistake most AI startups make is buying when they should build (and not modeling the cost curve).
Hire expertise when the investment exceeds $100K and your team lacks production ML experience.

FAQ

When should a company buy an AI API instead of building custom?

When AI is a feature of your product, not the product itself. The cost of building a custom model ($200K–500K minimum) only makes sense when the AI is your competitive advantage, not a commodity capability.

When should a company build custom AI instead of buying?

When three conditions are met: proprietary data that creates a genuine moat, accuracy requirements that general-purpose models can’t match, and the engineering team to sustain the investment.

What is the most common build-vs-buy mistake?

Companies building custom models for commodity tasks. Teams spend months building custom classifiers when the API handles the same task at 95% accuracy for $50/month.

Should I hire a fractional CTO to help with the build-vs-buy decision?

If your company is not AI-native and you are considering a six-figure AI investment, yes. The cost of assessment ($3K–10K) is trivial compared to building the wrong thing ($200K+).

How do I know if my data creates a real AI moat?

Three tests: Would a competitor with your data build a meaningfully better product? Can you replicate the value using public or synthetic data? Does your product generate data that improves the model over time?