Why OpenAI’s Forward Deployed Engineering Takes Six Months to Deliver Usable AI

Introduction: The Need for FDE

When ChatGPT first launched, the hype was huge but enterprises struggled to extract value because the traditional "hands‑off" support model failed to deliver reliable production‑grade solutions.

“We found our previous hands‑off technical support model was just not reliably getting these folks to production.” — Colin Jarvis

According to MIT NANDA’s 2025 report, 95% of AI deployments do not generate measurable ROI; OpenAI’s FDE team aims to be the catalyst that moves the remaining 5% to success.

What FDE Tries to Solve

FDE embeds engineers directly into customer organizations to understand business domains, co‑create solutions, and accelerate adoption of probabilistic AI models.

The team’s two goals are to discover product‑hypothesis opportunities and to solve high‑value industry problems that can steer OpenAI’s research.

Case Study 1: Morgan Stanley – 6 Weeks Technical, 4 Months Trust

Background : 40,000 wealth‑management advisors needed fast, reliable access to 70,000 research reports.

Technical Phase (6‑8 weeks) : Built a Retrieval‑Augmented Generation (RAG) pipeline, tuned retrieval, and designed guardrails to ensure answers referenced verified sources.

Trust Phase (additional 4 months) : Ran pilot programs, collected feedback, performed data labeling, and iteratively expanded deployment. After this period, 98% of advisors adopted the system and research‑report usage tripled.

The key insight is the “technology‑ready vs. trust‑ready” mismatch: roughly a 1:3 ratio of weeks to months.

Case Study 2: Semiconductor Company – Eval‑Driven Development

FDE helped a European chip maker identify the biggest waste in its value chain and applied AI to improve efficiency.

Methodology :

Forked Codex and added telemetry to observe model reasoning.

Co‑created a five‑case evaluation set with domain experts.

Adopted an "Eval‑Driven Development" workflow: define evaluation → write code → pass evaluation → deploy.

Traditional flow (code → test → deploy) was replaced with a stricter sequence where code is considered done only when it passes the evaluation suite.

Three automation levels were defined:

L1 – Advisor: model suggests fixes, humans decide.

L2 – Drafting: model writes PRs, engineers review.

L3 – Execution: after extensive evals, model can run tests in a sandbox.

Result: 20‑30% efficiency gains in early departments, with a target of 50% engineer‑time savings overall.

Case Study 3: Automotive Supply‑Chain – Deterministic + Probabilistic Hybrid

Scenario: a tariff increase on parts required rapid re‑routing of the supply chain.

FDE built a three‑layer architecture:

Deterministic guardrails : hard rules (e.g., at least two tire suppliers, delivery‑time limits).

Probabilistic layer : LLM interprets natural‑language queries and generates candidate re‑routing plans.

Verification UI : visual dashboards let users inspect reasoning, check data tables, and compare simulated outcomes.

This hybrid approach ensures zero‑tolerance constraints are enforced by rule‑engine logic while the LLM handles nuanced decision‑making.

Technical Takeaways

1. Eval‑Driven Development replaces the classic "code → test → deploy" with "define eval → code → pass eval → deploy"; the quality of the evaluation set determines the ceiling of LLM performance.

2. Deterministic guardrails + probabilistic reasoning form a “sandwich” architecture that makes AI systems auditable and trustworthy.

3. Progressive automation (L1‑L3) lets teams gradually hand over control as trust is earned.

Business Insights

FDE is positioned as a product‑discovery engine, not a profit‑center consulting shop. Success is measured by reusable product insights rather than service revenue.

The ability to say “no” to lucrative but non‑strategic projects is a key indicator of organizational maturity.

Organizational Lessons

FDE talent must combine deep technical chops (forking and instrumenting models), domain expertise, persuasive communication, and product intuition—an extremely rare skill set.

Resource allocation follows two axes: (1) product‑hypothesis validation (high‑reuse projects) and (2) research‑learning (high‑impact, low‑reuse problems).

Key Takeaways

AI adoption fails mainly due to trust gaps, not model capability.

Compressing the 1:3 technology‑to‑trust timeline is a competitive advantage.

Eval‑driven development and deterministic guardrails are essential for enterprise‑grade AI.

FDE’s success hinges on talent, capital, and strategic resolve—most companies cannot fully replicate the model.

References

Colin Jarvis | Head of Forward Deployed Engineering at OpenAI: Trust. Product. Impact., Altimeter Capital, November 2025. https://www.youtube.com/watch?v=cBD7_R-Cizg

MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, MIT Media Lab, August 2025.

Paul Graham, “Do Things that Don’t Scale”, July 2013. https://paulgraham.com/ds.html