Build vs. buy: the honest trade-offs

Every AI use case has a build/buy decision. Getting it wrong either way is expensive. Here's the framework that consistently produces the right answer.

The three options (not two)

• Buy — off-the-shelf product (Glean, Notion AI, Microsoft Copilot).
• Build on platform — you use a provider's API + your integration / UX / data.
• Build from scratch — you train or heavily customize models.

Most organizations should almost never choose "build from scratch." Almost every successful 2026 AI deployment is "buy" or "build on platform."

The decision tree

Start at the top; if the answer is yes, stop there.

1. Is there a product that covers 80% of the need and has happy customers in our segment? → Buy.
2. Do we have a differentiated data or workflow advantage that a product can't replicate? → Build on platform.
3. Are we an AI-native company where model choice is a core capability? → Possibly build from scratch.

When "buy" is right

• Enterprise search → Glean, Microsoft Copilot, Atlassian Intelligence.
• Meeting intelligence → Fireflies, Otter, Zoom AI Companion.
• Customer support copilot → Intercom Fin, Zendesk AI, Ada.
• Code assistance → GitHub Copilot, Cursor, Claude Code.

These are mature. Buying saves 12-18 months of engineering and gets you a product that keeps improving.

When "build on platform" is right

• You have proprietary data that wouldn't go into a vendor's system (regulatory, competitive).
• You have unique workflows a vendor can't economically serve.
• You need tight integration with your existing products or internal systems.
• The product category doesn't exist yet — you're early.

Build on a platform like OpenAI, Anthropic, or Google. Don't train models. Use their infrastructure.

When "build from scratch" is rarely right

It's right when:

• You are at the frontier of a specific capability (voice, vision, domain-specific reasoning).
• Your scale makes per-inference economics matter more than capability.
• You have regulatory requirements that mandate control of the model.

For everyone else, the cost of training/hosting beats the marginal benefit by a wide margin.

The hidden cost of "build"

• Upfront: 6-18 months to ship v1.
• Ongoing: dedicated team (2-5 engineers, plus product, plus data).
• Maintenance: model updates every 3-6 months require evaluation, prompt tuning, sometimes re-architecture.
• Opportunity cost: that team isn't building other things.

"Build on platform" costs less than "build from scratch" but still costs 3-5x what teams estimate upfront.

The hidden cost of "buy"

• Vendor lock-in. Your data, workflows, and users get wired into their product.
• Roadmap mismatch. Their priorities won't perfectly align with yours.
• Per-seat pricing scales unpredictably. An AI feature that saves 30 min/week/person might cost $30/seat/month — check the math.

The integration question

Either way, you'll probably need some custom integration. The question is scope: are you integrating an AI product (buy) or are you building one (build on platform)? Most organizations underestimate integration effort by 2-3x regardless.