← back to the archiveCover illustration for “Build vs buy (AI): when to make your own AI and when not”
TERM6mo ago·Published · Updated ·by Andy Padia

Build vs buy (AI): when to make your own AI and when not

In short: Build vs buy for AI weighs building on models and APIs yourself against buying a finished AI product, judged by differentiation, data and total cost.

Build vs buy (AI) is the decision between building an AI capability yourself, on foundation models, APIs, open-weights models and your own engineering, and buying a finished AI product or a vendor feature that does the job. In AI the line is blurry: buying often still needs integration and evaluation, and building usually means assembling rented models rather than training your own.

Build vs buy in plain words

Cook at home with bought ingredients, or order from a restaurant. Both involve someone else's ingredients; the question is who owns the recipe, the kitchen and the result.

How it works

The useful questions are:

  1. Differentiation: is this capability part of what makes you different, or table stakes?
  2. Data and integration: does it depend on your proprietary data and systems?
  3. Change rate: will the vendor keep up with model improvements better than you?
  4. Control: do you need control over models, data residency, evaluation and audit?
  5. Total cost: licences versus engineering, operations and evaluation over the life of the system.

Example

As an illustration, a logistics firm buys an AI meeting-notes product (generic, fast-moving, not differentiating) but builds its own dispatch-exception agent on a model API, because it depends on proprietary routing data and directly affects margins.

Where you'll meet it

  • Enterprise AI strategy and budgeting
  • Vendor evaluations and RFPs
  • Platform decisions about agent frameworks and model providers
  • Services engagements scoping AI work

What it changes for you

For leaders, my default is to buy commodity capabilities and build where data and process are your edge. For architects, building should still mean thin custom layers over replaceable models: Anthropic's December 2024 guide recommends starting with LLM APIs directly and adding complexity only when it is needed.

For procurement, a bought system is still your risk. The NIST AI Risk Management Framework from January 2023 asks for policies that address third-party AI risks and for contingency processes when third-party data or AI systems fail. Lock-in now includes prompts, evaluation data and agent workflows as well as software.

Build vs buy vs partner

Partnering, with a services firm or vendor co-development, sits between the two: you own the outcome and data while someone else supplies speed and skills.

Common mistake

My view: building because the demo was easy. A working demo arrives quickly; evaluation, edge cases, monitoring and support are the long part. Price the long part before you decide.

AI ROI · Foundation model · Cost per task · Demoability · AI governance

From the archive: Should you build or buy AI agents? A decision method

Price the evaluation, monitoring and support that follow the demo before you decide to build.

#field-guide#strategy#build-vs-buy#enterprise-ai#procurement

Sources

  1. anthropic.com/engineering/building-effective-agents
  2. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
← older drop
WebMCP: letting websites hand tools to AI agents
newer drop →
Agentic RAG: an agent that decides what to search for

related drops

explore all 557 drops →
← back to the archiveday 175