Project · Beginner · about 8 hours

From a vague ask to a pilot plan

Play the forward deployed engineer for an online store's support team, analyse their message backlog, and write the discovery notes and a one-page pilot plan with a measured baseline.

Skills:Communication, Customer-facing delivery, SQL, Stakeholder management

Part of: Forward Deployed Engineer

Data and licences

Brief

The head of support at an online store says: "We want AI to help with support." Their team handles questions about orders, refunds, invoices, delivery and accounts. The Bitext dataset stands in for their message history: 26,872 messages labelled with 11 categories and 27 intents. It is synthetic, which is useful for practice and a limit you must state.

Your job is not to build anything yet. It is to find out what would be worth building, and to write a plan the sponsor can say yes or no to.

What to build

  1. Ask a friend to play the head of support using a short persona sheet you write (goals, pains, constraints). Run a 30-minute discovery call and write notes in the five areas: workflow, pain, data, people, success.

  2. Load the dataset into DuckDB or pandas. Report volume by category and intent, and pick the two or three intents that are the best pilot candidates. Explain why.

  3. Define a baseline you could measure at a real customer (handling time, first-response time, reopen rate) and how you would collect it. Mark which numbers are assumptions.

  4. Write the pilot plan on one page: scope in and out, success criteria, guardrails, weekly plan, owners on both sides and a decision date.

  5. Record a three-minute walkthrough for the sponsor.

Deliverables

  • Repository with discovery-notes.md, analysis/, pilot-plan.md and a link to the recording

How it is graded

Reviewers read the plan as the sponsor would. The key question: would they know exactly what they are agreeing to, and how success will be judged?

Stretch

Write the security questionnaire answers you expect to need (data flows, storage region, retention, access) for the pilot.

Deliverables

  • Discovery notes from a role-played call, in the five-area format
  • A baseline analysis of message volume by category and intent, with SQL or pandas you can rerun
  • A one-page pilot plan with scope, baseline, success criteria, guardrails, owners and a decision date
  • A three-minute recorded walkthrough of the plan for the sponsor

How a reviewer grades it

CriterionWeightWhat good looks like
Discovery notes separate facts, assumptions and open questions25%
Baseline analysis is correct, reproducible and answers the sponsor's question25%
Success criteria are measurable and tied to the baseline30%
Plan and walkthrough are clear enough for a non-technical sponsor20%