What you'll learn: how to produce and defend an AI design from an unseen brief in 30 minutes, how a seven-criterion rubric scores it (0 to 2 each, 14 in total, pass at 10), and how to read a scored model answer.
At a glance
Figure 1. The timed defense: four boxes, then a score
How to read this: the clock runs across all four boxes; scoring comes after.
- One unseen brief, 30 minutes: 5 frame, 15 design, 5 risks and evaluation, 5 defend.
- Seven criteria, 0 to 2 each, 14 in total; 10 passes.
- Attempt first, then open the model answer.
Scenario
Composite scenario, not a real company. Cedar Row Facilities is a facilities-services firm with 1,200 active contracts and 40 account managers (illustrative numbers).
Contracts are PDFs: three templates plus customer edits, about 15 percent scanned, amendments filed separately. Renewal dates and penalty clauses (a notice period before automatic renewal, a charge for early termination) sit in each manager's spreadsheet, and some renewals are noticed late. The firm wants an assistant that reads contracts, flags both, and drafts reminder emails a human approves. No legal advice.
What you will be able to do
- Produce a defensible design from an unseen brief in 30 minutes.
- Score your own design against seven criteria and say where you lost points.
- Defend one decision aloud and concede one weakness.
Prerequisites
- The seven-step design method and the patterns: understand, answer, act.
- Review board simulation, for defending under challenge.
Core ideas
The rubric
| Criterion | Weak (0) | Acceptable (1) | Strong (2) |
|---|---|---|---|
| Requirements quality | Starts drawing; no measure of success | Users, data, one outcome, partly unmeasurable | Measurable outcome, baseline to collect, constraints, out of scope, questions asked |
| Pattern choice | One pattern for everything | A fitting pattern, no alternative | Pattern follows the requirement; one alternative discarded with a reason |
| Architecture soundness | Boxes without flow or boundaries | Components and flow; no failure path or trust boundary | Flow, failure path and trust boundary drawn, plus the human step where the design acts |
| Risk and security | None, or "secure it" | Two or three generic threats | Threats tied to this design, each with a control; outbound actions need approval |
| Evaluation plan | "Test it" | Metrics, no labelled set | Labelled set and its mix, metrics, failure cases, how the outcome is measured |
| Cost awareness | None | Says cost matters | Volume times tokens, labelled placeholder prices, biggest driver, human time |
| Communication | No order, jargon | Ordered, no clear decision | Decision first, plain words, one weakness conceded |
How the steps feed the rubric
Figure 2. Which design step earns which criterion
How to read this: skipping a step leaves its criterion at 0.
What the score means
Pass is 10 of 14. Fix any 0 first: a 0 on risk and security means not ready to ship, whatever the total. The score measures one exercise, not an interview or job outcome.
It can give practitioner-level ability to design, build and defend an enterprise AI solution from a blank page, with a path to deepen it through real projects. The next lesson is that path.
Worked example
Attempt the brief first, then open the model answer.
Open after your attempt: the model answer and its score
Frame. Outcome: no renewal notice deadline is missed. Baseline to collect: late renewals in 12 months, minutes to find a date. Users: 40 account managers, a legal reviewer. Data: 1,200 PDFs (about 180 scans) and amendments. Out of scope: legal advice, automatic sending.
Design. A document-processing pipeline: structured extraction, a human review queue, and code for date arithmetic (renewal date minus notice period gives the deadline). The model only extracts fields and drafts text. Discarded: a chat agent over all contracts (no stored deadlines, no per-field accuracy) and automatic sending (no human check).