The FDE Playbook: Five Stages of AI Delivery from First Client Meeting to System Handoff

Five Stages Derived from Real Failures

The framework is not forward‑designed but reverse‑engineered from costly AI delivery failures, each stage addressing a specific class of loss such as building on vague requirements, deploying untested prototypes, launching without user adoption, skipping productization, or never handing over operations.

Discovery: Turning Vague "Want AI" into an Actionable Problem

In this phase the FDE works with the client to translate a generic desire for efficiency into a concrete, measurable problem definition—identifying the exact workflow, inputs, outputs, and success metrics. The most common failure is starting development while the requirement remains fuzzy, leading to costly rework. The stage ends when both parties agree on a concise, signed problem statement that can be measured.

Prototyping: Validating on Real Customer Data, Not Clean Test Sets

After discovery, the team builds a prototype that runs on the client’s actual data. The OpenAI FDE team insists on establishing evaluation criteria together before any production code is written. A concrete example is the John Deere project, where hundreds of real agricultural samples were reviewed to define the benchmark. Failure occurs when prototypes use sanitized data and then collapse in production. The stage ends when the prototype meets the jointly defined real‑data standards.

Deployment & Co‑Running: Technical Integration Is Only the Entry Ticket, User Trust Is the Goal

Two parallel tasks are performed: (1) technical hardening—integrating the prototype with production systems, handling permissions, legacy interfaces, compliance, and IT approvals; (2) co‑running—standing by the users, observing usage, collecting feedback, and iterating until users trust and rely on the system. The Morgan Stanley case shows a six‑to‑eight‑week technical rollout followed by extended co‑running, achieving 98% adoption and a three‑fold increase in research report usage. Failure is declaring project completion after technical launch without co‑running, resulting in low adoption.

Productization: Turning Custom Experience into Reusable Capability

This stage extracts deliverable‑specific assets—evaluation frameworks, data connectors, trust‑building processes—and feeds them back into the core product. A practical metric is that within 90 days at least one reusable component must be integrated; otherwise the team is merely a services outfit. Skipping this step leads to knowledge lock‑in and scaling problems. The stage ends when a tangible reusable capability is shipped to the product.

Handoff: Enabling the Client to Operate Independently

Effective handoff begins weeks before the final transition by involving the client’s internal owners early, turning operational procedures into automated workflows rather than static documents. Success is measured by the client’s team independently running the system after the FDE departs. Failure occurs when the FDE remains on‑site, creating dependency and preventing team scaling.

Closing the Loop

Each completed cycle enriches the next: evaluation frameworks accelerate future discovery, connectors avoid rebuilding, and trust‑building methods are reused. This reduces marginal cost and enables the FDE model to scale, unlike traditional custom‑development which lacks such feedback loops.