The 72-Hour Forward Deployment Discovery Sprint

The client kickoff slide says "AI Strategy Phase 1." Your job in seventy-two hours is to prevent a six-month science project that never reaches production. Forward Deployed discovery is not requirements gathering in a conference room. It is rapid hypothesis testing under customer constraints to produce a ranked backlog with architecture sketches and explicit kill criteria.

Hours 0–8: Shadow the work, not the workshop

Sit with the end user. Watch them search the policy PDF. Time how long it takes. Note the workaround — they Slack Sarah because search returns zero results for "pregnancy" when the document says "maternity leave." The real requirement is "Sarah's brain in a searchable system," not "chatbot."

Record exact utterances. Those strings become your eval set and your intent taxonomy seed data. Executives describe strategy; practitioners describe pain. FDE discovery lives with practitioners.

Hours 8–24: Data archaeology

What data exists? PDFs in SharePoint with broken permissions? APIs behind OAuth nightmares that take three weeks to approve? SQL tables with forty percent nulls in the column that matters? Document the data contract before the model contract.

Ask: What is the refresh cadence? Who owns updates? Is there already a search index? Sometimes the right intervention is metadata cleanup, not embeddings. Credibility comes from recommending the non-AI fix when appropriate.

Hours 24–48: Intent taxonomy from real utterances

Collect fifty real questions — support tickets, Slack messages, call transcript snippets. Cluster into interpretive (policy lookup), transactional (update beneficiary, submit claim), high-risk (medical/legal/financial advice), analytical (compare plans, summarize across docs), navigational (take me to the upload form), and meta (why did you say that, start over).

Each class maps to a different technical pattern. Interpretive gets RAG with citations. Transactional gets agents with OpenAPI-bound tools and confirmation steps. High-risk gets guardrails and warm handoff to humans — never autonomous binding advice.

If two clusters would use the same pipeline, merge them. Taxonomy serves routing, not slide aesthetics.

Hours 48–72: The one-page delivery brief

Problem statement in one sentence. Primary user persona. Intent classes with example queries. Proposed pipeline per class (RAG, agent, handoff). Data sources and known gaps. Success metrics the executive already cares about. Two-week MVP scope with explicit out-of-scope list. Top three risks and mitigations. One decision needed from leadership before build starts.

Share the brief with the customer technical lead before the executive readout. Surprises in executive meetings kill trust.

When discovery reveals you should not use AI

Say it plainly. Recommend a rules engine, better search UX, workflow automation without an LLM, or a human staffing fix. FDEs who recommend not using AI when the data or risk profile does not support it earn the next engagement. FDEs who force LLMs into every problem lose the room when the demo hallucinates in front of the compliance officer.

Connecting discovery to Applied AI architecture

The discovery sprint output is the architecture input. You are not choosing LangChain versus LlamaIndex on hour one — you are choosing whether the dominant user need is retrieval, action, or triage. That choice drives everything downstream: eval rubrics, guardrail strictness, latency budgets, and whether you need human-in-the-loop on day one.

Bring the one-page brief to your platform team if you have one. Reusable patterns — citation-required RAG templates, approved tool schemas, eval CI gates — accelerate week two. Local discovery still matters because the customer's mess is always unique.

Seventy-two hours of disciplined discovery saves twelve weeks of building the wrong system. That is the FDE superpower no foundation model automates away.