Problems, users, data, constraints, success measures — before any design.
Domain 6 · 14% of the exam
The best architecture
ever explained badly
lost to a mediocre one explained well.
Half of architecture is systems. The other half is people: hearing what stakeholders mean, translating trade-offs into their language, and carrying the system through its whole life — not just its launch.
Objectives
What you’ll be able to do
- Run structured discovery and requirement gathering
- Communicate decisions and trade-offs per audience
- Align expectations and negotiate honest SLAs
- Document architectures for the people who inherit them
- Manage the full lifecycle: discovery to iteration
The story
Three languages, one decision
The claims system needed a second model tier — complex cases deserved deeper reasoning, at four times the cost. Meera had one decision to defend and three rooms to defend it in. In the boardroom, she never said "model." She said: "Complex claims are 9% of volume and 80% of our legal exposure. This spends money exactly where the risk lives." The CFO nodded at a fuel gauge, not a transformer.
With legal, the same decision wore different clothes: "Every complex-case decision will carry deeper reasoning and a full citation trail — here's what that means for discoverability." And with the adjusters, who feared being replaced: "The routine flood you hate goes to the machine. The judgment calls that need twenty years of experience — those come to you, with better preparation."
Same architecture. Three translations. "A decision that isn't understood," she told Theo afterward, "gets re-litigated forever. Translation is how a decision becomes final." Then she wrote it all down in an ADR — because the fourth audience is whoever holds this system in three years, and they weren't in any of the rooms.
Metaphor map
Images that stick
| Concept | Image | Why it sticks |
|---|---|---|
| The architect | The interpreter at the summit | Business and engineering speak different native tongues. Whoever translates faithfully holds the room. |
| Discovery | Archaeology | The request is the surface. The requirement is buried under it — dig with questions, brush gently, date everything. |
| Trade-off talk | The pilot’s briefing | Passengers don’t need aerodynamics. They need the route, the weather, and what happens if it changes. |
| SLA | The contract on the wall | A promise with numbers, dates, and a plan for the misses. Everything else is a mood. |
| Documentation | The map for the next traveler | ADRs, runbooks, baselines — drawn for someone who has never met you and can’t call you. |
| Lifecycle | Seasons | Launch is one day; the system lives for years. Monitoring feeds discovery, and the loop begins again. |
Discovery
Discovery interview simulator
A stakeholder speaks. Choose the question a working architect asks next — then see why.
Vp Of Operations, First Meeting
- They say
- We want an AI chatbot. Everyone has one now.
- Best reply
- “What outcome would make this project a success a year from now?”
- Why
- Exactly. Anchor on the outcome, not the artifact. "A chatbot" is a solution someone saw at a conference; the outcome is what you’ll actually be measured against.
Same Meeting, Ten Minutes Later
- They say
- Fine — the real problem is our agents spend hours answering the same questions.
- Best reply
- “Which questions, how many per day, and where do the answers live today?”
- Why
- The archaeologist’s question: volume, shape, and data source in one stroke. The answers determine pattern, model tier, and retrieval design — this is requirements becoming architecture.
Executive Sponsor, Closing
- They say
- One condition — it has to be 100% accurate. We can’t afford mistakes.
- Best reply
- “Let’s define what a mistake costs, target a measured accuracy on real cases, and route the highest-stakes ones to human review.”
- Why
- The architect’s move: replace an impossible absolute with a measurable target plus a safety design. The sponsor gets something stronger than 100% — a system with a plan for its own errors.
Trade-off translator
One decision, four languages
The decision: "Route complex cases to a more capable, 4× more expensive model tier, with human review on denials." Pick who's across the table.
CFO
Money in, risk out. Unit economics, not architecture. The CFO hears every technical word as either a cost or an excuse.
- Never say
- model tier, tokens, inference
- Say this
- Complex claims are 9% of volume but 80% of our legal exposure. This routes four times the analysis budget to exactly those cases — total spend rises 12%, projected disputed-claim losses fall by more than that alone.
General Counsel
Defensibility. Can every decision be reconstructed, cited, and defended in front of a regulator or a court — years later?
- Never say
- black box, the AI decides, roughly
- Say this
- Every complex case gets deeper reasoning with a full citation trail — which regulation, which policy clause, which precedent. Denials always carry a named human signature. If we’re ever asked "why", the answer is one query away.
Head of Support
Their people. Workload, morale, and the quiet fear in every AI meeting: are you automating my team out of a job?
- Never say
- headcount efficiency, replace, automation rate
- Say this
- The repetitive flood — status checks, routine approvals — goes to the machine. What comes to your team is the work that actually needs twenty years of judgment, arriving pre-researched. Your seniors become the escalation tier, not the paperwork tier.
CTO
Now you can speak architecture — but the CTO’s real questions are operational: failure modes, rollback, monitoring, and who gets paged.
- Never say
- it just works, trust me, temporary workaround
- Say this
- A router classifies complexity and sends the top decile to the capable tier; both paths share one eval harness and one trace pipeline. If tier-two degrades, we fail over to tier-one plus mandatory human review — same SLA, higher review load, zero downtime.
Lifecycle
Five phases, one loop
Architecture, pattern, model, and governance choices — each traced to a requirement, captured in ADRs.
Diagrams, runbooks, eval baselines, limitations — understanding transferred, not just access.
SLA dashboards, drift alerts, feedback loops with real users and stakeholders.
What monitoring reveals becomes the next discovery. The loop is the lifecycle.
Storyboard
The film, shot by shot
Two delegations enter the room. Neither speaks the other’s language.
Between them sits one chair. The interpreter’s.
"We want a chatbot," says the surface. The interpreter picks up a brush and starts to dig.
One decision, four rooms. Watch it change clothes without changing shape.
"One hundred percent," demands the sponsor. The interpreter doesn’t flinch — she reframes.
And for the person who inherits all this in three years — a map, so they never have to find us.
Flashcards
Recall drill
Select a card to reveal the answer.
Quiz
Exam drill
Question 1
In the first discovery meeting, a stakeholder says: "We need an agentic AI platform with RAG and MCP." What is the architect’s best response?
Question 2
A sponsor demands a guarantee of perfect accuracy before approving launch. The architecturally sound response is to:
Question 3
Your team hands a production Claude system to a client’s engineering group. Which handoff package is complete?
Revision
Close your eyes and remember
Dig for outcomes, not features. Translate every trade-off into the listener's language. Promise numbers, never perfection. And write for the person who inherits the system — they're the audience that matters longest.
Domain 6 · 14% of your exam