Skip to module content
🧠AI Academy
0/91
Sign in
Module 05 · ~9 min

The Adoption Playbook

OpenAI × Anthropic — the frontier labs' interlocking evidence.

Reading progress
0/8 · 0%

The big idea

💡Key idea
OpenAI's seven enterprise lessons are a sequence, not a menu — and the order matters. It starts with measurement (evals), moves through embedding, iterating, and specializing, and only reaches decentralization, platformization, and automation once the earlier steps are solid. Notice what's first: the frontier lab's own #1 organizational lesson is exactly the thing Volume 1 spent pages teaching you to build. Your eval harness isn't a technical nicety — according to OpenAI's field evidence from real enterprise deployments, it's the thing that unlocks adoption.
Quick check
1 question · instant feedback
0/1
  1. OpenAI's lesson #1 from AI in the Enterprise:

Numbers that matter

98%
of Morgan Stanley advisors use the tools daily after evals-gated rollout
OpenAI, AI in the Enterprise
20% → 80%
document access at Morgan Stanley post-deployment
+60%
error detection at Lowe's after fine-tuning on proprietary product data
2,900
custom GPTs built by BBVA employees in 5 months across legal, compliance, service
17,000
developers shipping AI apps on Mercado Libre's Verdi platform with guardrails
Quick check
1 question · instant feedback
0/1
  1. Anthropic's core deployment counsel:

Deep dive

2/2 open

Where OpenAI's enterprise guide pushes ambition, Anthropic's *Building Effective Agents* supplies the restraint that makes ambition survivable. The counsel: find the simplest solution first, and add complexity only when it demonstrably improves outcomes — not because the architecture looks more impressive.

Prefer composable patterns over frameworks. If you do adopt a framework, understand the underlying code — because wrong assumptions about what a framework is doing under the hood are, per Anthropic, a leading cause of production error.

The architecture heuristic: use a single augmented model call for simple predictable tasks; use composable workflow patterns (chaining, routing, parallelization) for medium-complexity tasks; reach for multi-agent only for genuinely specialized, high-complexity problems where the path can't be predetermined. Match the architecture to the business value, not to the state of the art.

Turn the seven lessons into your implementation-phase checklist — literally a slide in every build proposal.

Label it: 'Here is the deployment sequence that the frontier labs' own enterprise evidence supports. Here's where your engagement sits on it.'

This pre-answers 'why are we spending week one on evals?' with Morgan Stanley's 98% adoption rate instead of with your opinion. That's the difference between a recommendation clients accept and one they negotiate around.

Quick check
1 question · instant feedback
0/1
  1. BBVA (2,900 employee-built GPTs) illustrates which durable principle?

How to run it

  1. 1 · Start with evals
    Morgan Stanley gated every application behind rigorous benchmarks before scaling. Measurement before rollout creates the organizational confidence to deploy.
  2. 2 · Embed AI into products
    Indeed used AI to explain why a job was recommended, lifting applications ~20%. Value comes from inside the workflow, not beside it.
  3. 3 · Start now, invest early
    Compounding favors early iterators. Late starters buy the same models but not the accumulated feedback data.
  4. 4 · Customize and fine-tune
    Lowe's tuned on proprietary product data: +20% search tagging accuracy, +60% error detection. Domain specificity is engineered, not prompted.
  5. 5 · Get AI to the experts
    BBVA let 125,000 employees build; 2,900 custom GPTs in five months. The people closest to the process find the use cases central teams can't.
  6. 6 · Unblock developers
    Mercado Libre's Verdi let 17,000 developers ship AI apps with guardrails built in. Platform-with-guardrails beats gatekeeping.
  7. 7 · Set bold automation goals
    OpenAI's internal platform handles hundreds of thousands of support tasks monthly. Aim at whole processes — but earn it with lessons 1–6 first.
Quick check
1 question · instant feedback
0/1
  1. How should you cite these vendor case studies to a skeptical CFO?

In the field

🔬Worked example
OpenAI's lessons in your practice: Lesson 1 (Morgan Stanley: 98% daily adoption after evals-gated rollout, doc access 20%→80%) → sell your eval harness as change management. Lesson 5 (BBVA: 2,900 employee-built GPTs in 5 months) → empower one named process owner first; their peers become the pull. Lesson 6 (Mercado Libre's Verdi: 17,000 developers with guardrails built in) → your mid-market version is n8n skeletons + prompt library handed over with a one-hour enablement session.
🚫When not to reach for it
Both documents are publications by companies selling the underlying capability, showcasing their best customers. The lessons replicate well because they converge with independent academic findings — cite them for the convergence, not as neutral research. And translate scale honestly: your client's version of BBVA's 2,900 GPTs is one process owner empowered with a well-guarded tool and a feedback channel.

Pitfalls & takeaways

Failure modes

  • Reading the seven lessons as a menu to cherry-pick from, instead of a sequence.
  • Adopting a framework whose underlying code you don't understand — Anthropic's own leading source of production error.
  • Skipping evals to move faster — you lose the adoption unlock, not just the QA layer.
  • Central IT ownership of decentralized-adoption lessons.

Durable takeaways

  • Evals are the adoption unlock, not a QA layer.
  • Empower the process owner, not the org chart.
  • Make the safe path the fast path (platform + guardrails).
  • Anthropic's restraint keeps OpenAI's ambition survivable.

Do the work

📦Artifact to produce
Implementation-phase checklist slide for every build proposal.

Sources

  • · OpenAI, AI in the Enterprise (2025)
  • · Anthropic, Building Effective Agents