See the structure, then implement and verify with the notes.
Place requirements against product capability, tenant extensions and product gaps before committing.
What situation are you walking into?
The first email asks for AI to improve repeat purchases. It names no baseline, owner or workflow. Your first deliverable is a defensible investment decision, not a technical architecture.
01 / Translate AI ambition into a broken job
Reconstruct a recent incident: actor, input, decision, delay and consequence. Break retention into identification, segmentation, offer design, outreach and measurement. Select a tractable job with accessible evidence and a real operator. Apply the same logic to prospect research and support knowledge.
02 / Four gates for project fit
Assess task value, data readiness, organizational ownership and delivery feasibility. Scores help compare candidates; hard blockers remain hard blockers. A valuable opportunity without authorized data access is not ready for an execution pilot.
03 / The delivery–product boundary
An FDE combines engineering, problem definition, validation and result interpretation. This does not replace domain experts, sales, product or operations. Classify capabilities as standard, configurable, adapter work, product gaps or unsupported. Keep tenant-specific semantics outside the core.
04 / Make the first commitment testable
Commit to validating a bounded opportunity on authorized samples, with explicit coverage limits. Compare manual improvement, rules and a single model call before proposing autonomy. A well-chosen SQL query can be the right delivery decision.
Put the method to work
Exercise: compare an autonomous business brain, evidence-backed sales briefs and a generic approval assistant. Sales briefs may be a better pilot because evidence is available, users can score them this week and sending remains controlled. Change the choice if those assumptions fail.
Practice this week
- Describe three candidate jobs, a recent failure and the cost of inaction.
- Complete four fit gates and the five-way fit/gap map.
- Run kit tests and document environment and capability boundaries.
Acceptance checks
- No unsupported revenue or efficiency promise.
- Business, data and acceptance owners are named by role.
- Deferral criteria and next validation are executable.
90 minutes: turn an ambition into a bounded validation
Work in groups of three: sponsor, operator and FDE. Synthetic setup: the sponsor wants growth, the operator spends four hours weekly on reports, and IT can only export CSV. These are exercise assumptions, not client facts.
0–20 min: map the current job, a failure and unknowns. Mark each assertion F (given fact), H (hypothesis) or U (unknown).
20–55 min: compare manual improvement, scripts, a single model call and an agent. Decide by evidence needs and maintenance burden, not perceived sophistication.
55–90 min: exchange fit memos. Reviewers challenge one prerequisite; the team narrows scope, adds a test or defers.
Artifact excerpt · synthetic example
Decision: proceed conditionally
Job: reconcile two authorized reports; produce evidenced questions
Boundary: no budget changes or outbound messages
One-week evidence: five examples, blind review, comparable time logs
Prerequisites: data rights, metric reconciliation, a real operator
Defer if: only screenshots exist, without definitions or samplesInstructor review
Do not claim all four hours as savings. Include review, exceptions and maintenance, then examine whether released capacity is used productively.
An unseen field challenge ↗
If manual process improvement addresses most of the problem, what should you sell first? Explain the system boundary you would preserve for later automation.
Case studies for this module
Separate tenant adapters, domain assets and the product core. Feed learning back without embedding one tenant in the core.
High value, but sample access is unauthorized. What next?
Masking does not create permission. Synthetic data can validate a technical path, separately from real-world effectiveness.
References and further reading
OpenAI ↗ · OpenAI Deployment Company2026-05-11 · OpenAIAnthropic ↗ · Building effective agents2024-12-19 · AnthropicRead alongside the assignment and record design constraints and tests. Verify specifications against your deployment version.