Rollout & adoption
Take it from demo to pilot to production, and make sure people actually use it.
Roadmap: Client simulation & ownership
Why it matters
Go-live isn't the finish line. A deployment that works but isn't used has created no value. FDEs own the path from a demo to a controlled pilot to broad production, with clear gates at each step. Not knowing the difference between a demo, a pilot and production is a red flag in interviews.
What it's made of
- Shipping to a live URL: CI/CD, environments, secrets
- Pilot cohorts behind feature flags
- Readiness gates and launch checklists
- Capacity and cost estimates with labelled assumptions
- Training, support handoff and runbooks
- Adoption metrics and feedback loops
- Clear expand / narrow / pause / stop criteria
Demo, pilot, production: know which you're building
- Demo: proves something is possible. Happy path, sample data, no SLAs.
- Pilot: a small real group on real data, with monitoring, evals and a rollback. Its job is to produce evidence.
- Controlled production: wider rollout behind flags, with support processes and on-call.
- Broad production: everyone, with SLOs, cost budgets and a team that owns it day to day.
Estimate capacity before you promise anything
Use a simple chain: users × requests per day × tokens per request × price. 2,000 users × 10 questions × 6,000 tokens is 120M tokens a day. That number decides whether you need caching, a smaller model for simple questions, or a batch job instead of live calls. Say your assumptions, and say which one would change the design if it's wrong.
When nobody uses it
If the integration works but usage is flat, go and watch people work. Usually the tool lives in a separate tab, adds a step instead of removing one, or produces output people don't trust enough to send without rewriting. Fixes are usually about fit: put it inside the tool they already use, cut a step, or show sources so people can verify quickly.
Mid-level looks like
Proposes a sensible pilot with monitoring and a rollback plan.
Senior looks like
Defines proceed, narrow, redesign, pause and stop criteria up front, names who makes each call, and treats low adoption as a product problem to diagnose.
Common pitfalls
- Big-bang launches to the whole company.
- Scaling up before the pilot's edge cases are handled.
- Treating low adoption as a technical bug rather than a workflow or trust problem.
- No support handoff, so the customer's team can't operate it without you.
Practice drills
Drill 1
Write a pilot-to-production plan with gates for an AI support-reply assistant at a 300-agent call centre.
A strong answer: A small pilot cohort, measurable gates (quality, edit rate, handle time), a rollback trigger, a support handoff and an expansion schedule.
Drill 2
Estimate daily token volume and cost for an internal policy assistant at a 5,000-person company. Which design choice does the number change?
A strong answer: A clear formula, labelled assumptions, a sensitivity check, and a concrete design consequence (caching, routing, batch).
Drill 3
Six weeks after launch, the integration works but weekly active users are 12%. Diagnose it.
A strong answer: Going to observe users, separating access, fit and trust issues, and proposing small changes with a metric for each.
Best resources for this skill
Repos, courses, articles and podcasts, ranked by how much they teach.
OpenAI Academy
OpenAI's learning hub: workshops and guides on deploying models in real organizations.
PodcastThe Inner Loop (Prototype to Production)
In this episode, we dive into Chapter 8 of The Forward Deployed Engineer: Architecting the Last Mile of Enterprise AI by Sho Shimoda. We explore the "Inner Loop"—the critical, iterative phase of an engagement that bridges the gap between the end of customer discovery and the first production milestone [1]. Tune in as we break down the tactical frameworks FDEs use to turn a generalizable platform…
PodcastThe Last-Mile Problem in Enterprise AI
In this episode, we dive into Chapter 2 of The Forward Deployed Engineer: Architecting the Last Mile of Enterprise AI by Sho Shimoda. We explore the concept of the "last mile"—a term borrowed from telecommunications and logistics to describe the hardest, most expensive part of a deployment. You will discover why, contrary to popular belief, AI actually makes this last mile longer and explodes the…
AWS Certified Solutions Architect – Associate
The most requested cloud cert in FDE job posts. Proves you can design deployable systems in customer environments.
Read next
Solutions Architect's Handbook
Saurabh Shrivastava, Neelanjali Srivastav
A broad tour of the architecture choices you will be asked to defend on customer calls (cloud patterns, security, data, migration, cost), written from the solutions architect seat, which overlaps heavily with the FDE one.
The Forward Deployed Engineer
Paden Gayle · 2027
An upcoming O'Reilly book dedicated to the FDE role, by Paden Gayle, a founding FDE. It isn't released yet; watch for the early-release chapters.
The Philosopher in the Valley
Michael Steinberger
A reported history of Palantir, the company that created the forward deployed engineer. Read it to understand why the role exists, how it was meant to work, and what tradeoffs Palantir made by embedding engineers with customers.
The Mom Test
Rob Fitzpatrick
A short, practical guide to asking customers questions that produce facts about their current behaviour instead of polite encouragement. Use it before every discovery session so you scope the real problem rather than the one they pitch.