1. Courses
  2. AI Product Management
  3. Capstone

Capstone project for AI Product Management

Ask Ops: an AI assistant for 320 store managers

Mangrove Market's store managers send head office about 3,000 questions a week, such as "why is this GRN stuck?" and "why was my order cut?". Eight people on the ops desk spend about 300 hours a week answering them, and answers take about three hours. Propose an AI assistant that answers from Mangrove's procedures and systems, and later drafts fixes, such as supplier corrections, for a manager to approve. Compare it honestly with two other AI candidates, specify how it behaves and what it may do, build the golden set you would grade it with, estimate what it costs, and plan how it reaches all 320 stores without a privacy incident or a wrong refund along the way. Use the fictional procedures from the course (RF-2 refunds, RC-1 receiving, EX-3 expiry, ST-4 stock counts, PR-1 prices, CD-1 customer data and others), or write your own.

What you hand in

  1. A one-page memo comparing the assistant with two other AI candidates, each with its boring alternative, cost of being wrong and starting level
  2. A design spec with the autonomy level per task type, approval steps, answer card, failure states, and privacy and safety guardrails
  3. A 50-case golden set with a pass/fail rubric, and an error-analysis table from a test run of at least 30 answers
  4. An estimate of tokens and cost per answer and per month across 320 stores, at normal volume and festival peak, with a hosting recommendation
  5. A staged rollout plan with gates, the metrics that would make you pause it, and an ROI model for the first agent task type

Check your work

Tick each item when your project meets it. Be strict: this is the rubric the reviewer uses.

  • The memo starts from measured pain and names a non-AI alternative for every candidate
  • Cost of being wrong is priced separately for at least two types of question, and customer data is handled by a rule in code
  • The answer card shows sources, data with its time and limits, and the assistant says "I don't know" with a route to a person
  • Every guardrail is enforced in code and has a test case; permissions live in retrieval and tools, not in the prompt
  • The golden set covers every category, over-represents risky cases, and has labels a domain expert owns
  • Any LLM judge is checked against human labels, with its catch rate reported, not only agreement
  • The cost estimate counts every input token, including tool rounds, and the hosting choice uses break-even volume
  • Autonomy is set per task type with written gates and step-down triggers, and ROI is net of review time, errors and upkeep against comparison stores

0 of 8 checked

Model answer

The model answer opens once your capstone passes the review, so you can compare your work with an outline of a strong submission.

Hand in your capstone

Your links are checked automatically, then the work is scored against the rubric above, item by item, with comments. A pass counts toward the course certificate and shows on your profile.

Loading…