Build vs Buy Worksheet As of 2026-05

Score one use case across 6 axes → ranked recommendation: Claude direct / Claude via hyperscaler / OpenAI / open-source / packaged SaaS. Anchored on Claude — alternatives appear only as the right call when they are. Anchor heuristic: build where you have a moat, buy where you don't.

1. Describe the use case

2. Score 6 axes

Regulatory + data sensitivity

Regulated PII, PHI, financial, sovereign, classified. High = workload requires BAA, residency, or sovereignty constraints.

Low — public/internalHigh — PHI/sovereign

Latency requirement

How time-sensitive the response must be.

Low — overnight okHigh — sub-second

Customization depth

How much your domain shapes the answer. Generic Q&A vs deep proprietary procedures + style + edge cases.

Low — generic chatHigh — proprietary domain

Scale (request volume)

Sustained monthly traffic.

< 100k/mo> 50M/mo

Internal AI/ML expertise

Engineers who can ship + own a Claude-based system, including evals, safety, ops.

None / 1 dabblerDedicated AI team

Strategic moat depth

How much this use case is a differentiator vs. a commodity. High = your version wins on proprietary data, workflow, or expertise that competitors can't buy. Low = anyone can buy a near-equivalent SaaS today.

Pure commodityCategory-defining moat

Ranked recommendation

How recommendations are scored

Each option has a fit profile across the 6 axes. The worksheet computes a fit score per option = how closely your inputs match each profile's strengths. TCO bands are order-of-magnitude annual ranges, not commitments.

No option is universally best. The dominant signal is usually regulatory + scale; customization depth flips Skills/MCP from optional to essential; moat is the tiebreaker between build and buy — moat ≥ 4 invalidates packaged SaaS, moat ≤ 2 makes building hard to defend.

Decision frames (cheat sheet)

Regulated workload + need latest models → Claude direct. Hyperscaler routes lag on model versions.
Regulated workload + already deep on AWS / GCP → Claude via Bedrock or Vertex. Procurement path matters more than version lag.
Generic chat / RAG, no proprietary edge → Packaged SaaS likely cheaper unless volume is high.
High-volume, low-margin, latency-tolerant batch → Run the math both ways. Claude Haiku 4.5 + batch is hard to beat. Open-source self-hosted only when volume + ops capacity are both real.
Deep proprietary domain → Claude direct + Skills + MCP + plugins. The customization stack is what makes "build" win over "buy SaaS."
No internal AI/ML expertise → Packaged SaaS or Claude direct with a small build team. Skip open-source self-hosted entirely.
Multiple teams adopting → Claude direct + plugins distribution. Standardize the platform; let teams build on it.
Moat-defining workload (moat ≥ 4) → Build it. Claude direct + Skills + MCP + plugins (or open-source if sovereignty mandates). Packaged SaaS is wrong by construction — your competitors buy the same product.
Pure commodity (moat ≤ 2) → Buy it. Packaged SaaS or thin Claude-direct integration. Your engineering hours are better spent on the moat-defining workload above.

1. Describe the use case

2. Score 6 axes reset

Ranked recommendation

Decision frames (cheat sheet)

2. Score 6 axes