Study guides / CCAR-P / Domain 5

Stakeholder Communication & Lifecycle Management · Lesson 1 of 5

5.1 - Structured Discovery and Requirements Gathering

Run a structured discovery session that surfaces use case, data, constraints, risk and measurable success criteria, so that the model and architecture you recommend follow from evidence rather than from the first idea in the room.

Discovery is the phase in which the architecture is actually decided; the design document just writes it down. A stakeholder rarely arrives with a requirement. They arrive with a solution (“we need an agent that approves refunds”, “we want RAG over our wiki”) wrapped around a problem they may not have fully described. Your job in discovery is to separate the two, find the constraints that will eliminate options later, and agree on how success will be measured before you recommend a model, a pattern or a platform.

Scenario questions here usually hinge on sequencing: an option that recommends a model tier, a multi-agent design or a build plan before use case, data, constraints, risk and success metrics are pinned down is almost always the distractor.

Why premature recommendations fail

Three failure modes recur. Anchoring: the first architecture you sketch becomes the one everyone defends, even when later facts contradict it. Solution smuggling: the stated requirement is really an implementation (“use a vector database”), which hides the underlying need and shuts out cheaper options. Capability-first thinking: choosing the most capable model or most autonomous pattern because it impresses, not because the task needs it. Anthropic’s guidance on agents is to find the simplest solution that works and add complexity only when it demonstrably improves outcomes, since agentic systems typically trade latency and cost for task performance. Discovery is where you learn whether that trade is worth it, and sometimes that a deterministic program or one well-prompted call is enough.

A session structure that covers the ground

Organise the conversation around six areas and keep a running list of assumptions and open questions alongside the answers. The table shows what each area is really buying you, that is, which design decision the answers feed.

AreaQuestions that matterDesign decision it feeds
Use case and baselineWhat job is being done today, by whom, how often, at what cost and error rate? What does the user do with the output?Whether an LLM is needed; the baseline any solution must beat
DataWhere does the input live, how clean is it, how fresh must answers be, how sensitive is it, who may see it, how long may it be kept, where may it be processed?Retrieval design, redaction, platform and data-handling arrangement
ConstraintsLatency budget, peak volume, budget ceiling, integration points, languages, availability expectations, team skills, deadlineModel tier, sync vs. asynchronous processing, capacity planning
RiskWhat is the worst plausible wrong output? Who is harmed, how fast, and can it be reversed? What review capacity exists?Human-in-the-loop placement, verification steps, autonomy level
Success metricsWhat would make you call this a success in six months? What number, measured how, against what baseline?Eval design, launch gates, SLOs (see lesson 5.3)
Stakeholders and ownershipWho decides, who is affected, who supports it, who is not in the room yet?Governance, escalation, change control

Key concept

The output of discovery is not a model choice. It is a short, evidence-backed set of requirements, each traceable to a stakeholder answer or a real sample, plus explicit assumptions and open questions. A model and architecture recommendation is then a function of that set, and you should be able to say which answer would change it.

Turning a wish into a measurable success criterion

“High accuracy” and “better customer experience” are wishes. Anthropic’s documentation on defining success says criteria should be specific, measurable, achievable and relevant, and that most use cases need several criteria at once. Its example categories include task fidelity (with edge cases), consistency, relevance, tone, privacy preservation, context use, latency and price. Ask stakeholders to trade these off explicitly: is a slower, more accurate answer acceptable, and is a wrong answer worse than no answer? A usable criterion looks like “at least X% of tickets in the agreed test set are triaged to the correct queue, measured against senior-agent labels, with no cases in the ‘legal threat’ category misrouted.” The threshold is a business decision; your job is to make it measurable and to ask for real, messy examples so the test set mirrors the actual task distribution, including irrelevant, overlong, ambiguous and adversarial inputs.

Because Claude output is probabilistic, also ask how much variation is tolerable; that seeds the error-budget conversation in lesson 5.3.

Surfacing hidden requirements

Most expensive surprises are requirements nobody stated because they seemed obvious. Techniques that reliably expose them:

Several of these have concrete architectural consequences on Claude. Anthropic documents data-handling arrangements for its API (zero data retention and HIPAA readiness, with a feature-eligibility table), geographic controls for where inference runs, and organisation-level usage tiers, spend limits and rate limits that the docs describe as maximums rather than guaranteed minimums. On partner clouds the cloud provider is the data processor under its own terms. So “where may this data go?” and “what is the peak load?” can rule out a platform, a feature or a launch date. Treat these as architecture-level implications to confirm with legal, security and current vendor documentation; this is not legal advice.

Common exam distractor

Watch for options that answer a vague request by immediately selecting the most capable model, proposing a multi-agent design, or building a demo “to get buy-in”, and for options that defer data-sensitivity and compliance questions until a prototype works. Both skip the questions that decide platform, feature set and acceptance test. Equally weak: asking the sponsor to pick a model tier from a menu. The strong option gathers missing information in a structured way, states assumptions, and defines measurable success first.

Closing discovery

End with a written summary the sponsor can correct: problem and baseline, scope, data inventory and sensitivity, constraints, risk appetite and review assumptions, success criteria with thresholds and measurement, decision owner, and ranked open questions with owners and dates. Include a plain go/no-go statement: sometimes the honest finding is that rules, search or a workflow tool solves the problem, or that the data or review capacity does not yet exist. Discovery is not one-shot; new answers reopen old ones, so version the summary. It later feeds the decision records in lessons 5.2 and 5.4.

Exam traps

Practice question

An operations director at a retail bank asks you to design a Claude-based agent that will automatically approve customer dispute refunds up to a set amount. The request is one sentence, a demo is expected in three weeks, and nobody has mentioned data handling, error tolerance or how success will be judged. What is the best next step?

  • A Hold a structured discovery session on baseline, data sensitivity, volume and latency, the cost of a wrong approval and review capacity; collect real disputes; agree measurable success criteria; then recommend with stated assumptions. Correct

    This gathers exactly the information that decides autonomy level, human review, platform and eval design, and it converts the vague goal into measurable criteria before any recommendation. Sample cases also expose messy inputs early.

  • B Propose a multi-agent design on the most capable model tier so that accuracy is unlikely to be a concern, and start building the demo immediately to secure sponsor buy-in within the three-week window and show visible progress early.

    This is capability-first thinking. It commits to a costly architecture before you know the error cost, latency needs, data constraints or baseline, and it cannot be justified against any measurable criterion.

  • C Build a quick prototype on synthetic data to show the director what is possible, then use the director's reaction to the demo to decide the requirements and refine the design in later iterations once expectations are clearer.

    A demo on synthetic data anchors expectations on best-case behaviour and defers the risk, data and success questions. Requirements should drive the prototype, not the reverse.

  • D Send the director a comparison of the available model tiers and ask them to choose which one to use, since they own the budget, then design the agent around whichever tier they pick from the list and confirm the cost.

    Model selection is a technical decision that follows from requirements and evals. Asking a business sponsor to pick a tier without evidence transfers the decision to someone without the information to make it.

Build exercise: Design a discovery session and turn the answers into architecture implications

Intermediate · 45 minutes

You'll practice:

  1. Pick a realistic scenario (for example: 'an insurer wants Claude to summarise claim files and suggest a next action'). Write a one-page discovery guide with 4 to 6 questions under each of the six areas in the lesson table. Mark, for each question, the design decision its answer could change.

    Questions that cannot change a decision are noise. Forcing the link between question and decision is what makes discovery structured rather than a chat.

    You should see: A guide of roughly 25 to 30 questions where every question has a short 'changes: ...' annotation, such as 'changes: human review placement' or 'changes: platform choice'.

    Hints
    1. Which questions would you regret not asking if the system failed in production three months from now?
    2. Start from the decisions you will eventually have to make (model tier, sync vs. async, retrieval, human review, platform) and work backwards to the question that informs each.
    3. Example line: 'Q: What is the worst plausible wrong summary and who acts on it? -> changes: whether a human must approve before any action, and how strict the eval threshold must be.'
  2. Write down five hidden requirements the sponsor is unlikely to volunteer in this scenario (for example audit trail, retention of prompts and outputs, multi-language claims, adjuster review capacity, PII in attachments). For each, write the single question that would surface it.

    Hidden requirements are where projects overrun. Practising the extraction technique makes it reusable.

    You should see: Five requirement and question pairs, at least two of which involve people who are not the sponsor (support, security, downstream system owner, reviewers).

    Hints
    1. Who touches the output after Claude produces it, and what do they need that the sponsor may never have thought about?
    2. Use the prompts from the lesson: worst failure, real inputs, who is not in the room, what happens when it is wrong today, and the routine non-functionals.
    3. Example: 'Regulator may ask why a claim was routed this way' -> question: 'If a customer or auditor asks for the reason behind a decision six months from now, what do you need to be able to show them?' -> implication: store inputs, model ID, prompt version and outputs with retention rules.
  3. Take the vague goal 'summaries should be accurate and useful' and rewrite it as three measurable success criteria that cover at least two different categories (for example task fidelity, consistency, latency, cost). For each, state the baseline you would need to collect, the measurement method and who decides the threshold.

    This is the step that turns discovery output into an eval design and launch gate. Thresholds are business choices; measurability is your responsibility.

    You should see: Three criteria, each with a metric, a method (labelled sample, reviewer scoring, automated check), a baseline source and a named decision owner, with no adjectives like 'good' left un-quantified.

    Hints
    1. How would two different people, looking at the same 100 outputs, arrive at the same score?
    2. Pick a labelled test set drawn from real cases, a comparison against current human performance, and a metric for the cost of the worst-case error rather than only the average.
    3. Example: 'Recommended next action matches the senior adjuster's label on at least the agreed share of a 200-claim sample, with zero cases where a claim flagged for fraud review is auto-closed; baseline = current adjuster agreement rate; decision owner = head of claims.'
  4. Produce a requirements-to-architecture table with one row per requirement: requirement, evidence or source, assumption status (confirmed or assumed), design implication, and 'answer that would change this'. Include at least one row where the implication is 'an LLM may not be needed' or 'keep the workflow deterministic here'.

    Traceability lets you defend the eventual recommendation and revisit it when an assumption fails, and it forces honesty about where the simple solution is enough.

    You should see: A table of 8 to 12 rows, several marked 'assumed', and at least one row that argues for a simpler mechanism than an LLM call.

    Hints
    1. For each design implication, could you point to the exact stakeholder statement or sample that justifies it?
    2. Group rows by design decision (autonomy, platform, capacity, evaluation) and note which assumptions would flip the decision.
    3. Example row: 'Claims over the threshold amount need adjuster sign-off | Confirmed by head of claims | Human approval step before any payment action | If the threshold is raised, revisit review capacity and error budget.'
  5. Write the closing discovery summary in under one page: problem and baseline, scope, key constraints, risk appetite, success criteria, decision owner, and a ranked list of open questions with owners. Then critique it: what is the one open question that, if answered badly, would most change your recommendation?

    A short, correctable summary is the artefact stakeholders sign off on, and naming the most dangerous open question is how you prioritise follow-up.

    You should see: A one-page summary that a sponsor could correct in five minutes, ending with a single highlighted open question and the action to resolve it.

    Hints
    1. Which item on your list is least backed by real evidence and most expensive to get wrong?
    2. Rank open questions by (impact if wrong) times (how uncertain you are), and give each an owner and a date.
    3. Example: 'Highest risk: volume at month-end unknown. Owner: claims ops analyst. Due: before design review. If peak volume exceeds assumed capacity, revisit asynchronous processing and rate-limit headroom.'

Sources