---
type: "synthesis"
tags: ["synthesis", "tension", "model-strategy", "contradiction"]
sources: ["execution"]
id: "cross-build-vs-buy-model-strategy"
sourceVaultSlug: "hbr-seg-execution"
originDay: 8
articleStem: "hbr-seg-execution"
sourceUrl: "(unified vault: 7 sources)"
sourceTitle: "HBR — Firm Ⅱ-C · Execution quality — correct execution of AI"
---
## Two articles, opposite model advice

- **A054** argues [[claim-public-llms-low-value]] — public LLMs add little for real business tasks; deploy [[action-use-proprietary-slms|proprietary Small Language Models]] grounded in proprietary data for insight, and relegate ChatGPT/Claude to formatting.
- **A093 (Moody's)** argues [[claim-proprietary-models-not-competitive-advantage]] and [[contrarian-off-the-shelf-over-proprietary]] — commercial LLMs are 'ready-to-use tools'; advantage comes from *application* to proprietary data, not from owning a model.

## The real (narrower) agreement underneath

The contradiction is mostly about the word 'model.' **Both agree the value lives in proprietary data + workflow, not in generic public prose.** A054 gets there by tuning/owning a small model; A093 gets there by wrapping commercial models in a secure [[concept-ai-orchestration-layer]] that routes prompts across OpenAI/Anthropic/Meta/Google while keeping data inside its perimeter. A089 sits between them: leaders use commercial vendors ([[claim-partnership-ecosystem-maturation]]) but invest in [[prereq-meticulous-data-management]] and [[concept-unstructured-data-utilization]].

The unresolved axis is **regulatory/entropy risk vs. speed.** A054 fears public models pollute processes and drift ([[concept-knowledge-entropy]]); A093 accepts vendor dependence ([[question-long-term-vendor-lock-in]]) for velocity. A pragmatic reading: use orchestration + grounding to get commercial speed, but govern the perimeter and preserve human ground truth — the two articles are less opposed than they sound. See [[cross-preserving-human-judgment]].