AnthropicClaude Certified Architect – Professional2026-106.3 Manage stakeholder feedback loops and expectation alignment (including SLAs)
Give stakeholder feedback an owned cadence
Completed by the knowledge check
Learning objective
Establish feedback cadences and channels across discovery, pilot, and production.
Core idea
A feedback channel works only when someone can explain what happens after a report arrives. Establish who submits, who triages, who decides, when they respond, and how the submitter learns the outcome. Change the cadence as the delivery moves from discovering needs to testing assumptions to operating a service.
NIST AI RMF GOVERN 5.1 addresses collecting, considering, prioritizing, and integrating feedback from people outside the development team about the system’s potential individual and societal impacts related to AI risks. GOVERN 5.2 addresses incorporating adjudicated feedback into design and implementation. These are distinct responsibilities: receiving a complaint does not establish that the requested remedy is correct. Your deliverable is a feedback operating plan with named owners, review intervals, escalation routes, and disposition records.
How it works
Match the channel to the phase
Use discovery feedback to expose disagreements about the work, pilot feedback to test the proposed service, and production feedback to detect departures from the agreed service. The following is an illustrative operating plan, not a prescribed NIST schedule:
| Phase | Channels and participants | Cadence and decision output |
|---|---|---|
| Discovery | Facilitated workflow interviews, observed task walkthroughs, and a shared requirements register; operators, sponsor, support, and risk owners participate | Twice-weekly synthesis; confirm disputed requirements and return unresolved decisions to named business owners |
| Pilot | In-product issue reports, reviewer annotations, and short cohort debriefs; include unsuccessful users | Daily triage and weekly acceptance review; classify defects, training gaps, and scope requests |
| Production | Persistent feedback form, sampled case review, support intake, and stakeholder service review | Immediate incident escalation, weekly theme review, and monthly service review; assign fixes and communicate dispositions |
Set the intervals according to exposure and decision speed. NIST GOVERN 1.5 makes periodic review frequency and responsibility explicit; it does not require every organization to hold the same meetings. A high-impact workflow needs a faster route for possible harm than for a preference about wording.
Keep the requirements register stable enough to show disagreement. When an operator says that drafts need citations and a sponsor says that speed matters most, record both needs and their owners. Do not turn the loudest comment into a silently revised requirement. Scope decisions belong to the process in 6.3.4.
Capture enough context to adjudicate
Give each report an application case identifier, workflow stage, release identifier, reported outcome, expected outcome, impact category, and submitter contact where appropriate. These are fields you define in your intake record, not Claude API fields. Restrict sensitive case content to the people permitted to review it; a broad stakeholder dashboard can show themes without reproducing the underlying records.
For a Claude API failure, attach the response’s request-id to the restricted technical record (see 1.2.4). Anthropic exposes it as _request_id on Python and TypeScript SDK response objects and asks for the identifier when contacting support. An API request identifier locates a call; the application case identifier connects that call to a person’s workflow and feedback history. Do not expect either identifier alone to explain whether the business outcome was acceptable.
Offer more than a free-text inbox. A short reason list helps distinguish an incorrect answer, missing information, slow completion, confusing presentation, or an unsupported request. Keep an open-text option so the taxonomy does not suppress unfamiliar failures. Reviewer annotations and sampled cases complement volunteered reports: silence can mean that users abandoned the workflow rather than that it succeeded.
Assign a disposition and return it
The triage owner confirms the case, assigns an accountable resolver, and records whether it is a defect, an incident, a content problem, a training need, or a proposed capability change. A domain owner adjudicates substantive correctness; engineering should not label a disputed answer correct merely because the request completed.
Track a report through received, under review, decided, and communicated states in the feedback register. Those are illustrative application statuses. A decision can be fix, investigate, defer, or decline, with rationale and a next review date where needed. Response commitments should distinguish acknowledgement from resolution: a quick acknowledgement does not promise that every requested feature ships quickly.
Return the disposition to the submitter and publish a summary of recurring themes, decisions, and unresolved issues. NIST MANAGE 4.3 supports communicating incidents and errors and documenting tracking and response. The correction-to-release mechanism belongs to 1.2.4; this Bite owns the stakeholder route that brings a case to that mechanism and returns the decision.
Key distinctions and trade-offs
Feedback versus acceptance: An operator’s report is an input to a decision. Acceptance requires the agreed rubric and authorized decision owner. Anthropic’s success-criteria guidance supports quantitative measures alongside consistently applied qualitative scales; popularity alone is not a quality criterion.
Cadence versus urgency: A monthly service review can decide recurring investment questions. It cannot be the only path for an incident that requires action during the working day. Publish the urgent route alongside the ordinary schedule.
Broad participation versus review capacity: Including affected users improves coverage, but unlimited unstructured intake can overwhelm reviewers. Use one intake register with several accessible channels, explicit triage responsibility, and a visible ageing list. Track acknowledgement delay and undecided reports alongside the number received; volume alone rewards collection without action.
Architecture / decision scenario
An education organization pilots a Claude-assisted course-support service across eight departments. It has two support analysts and a four-week pilot. The sponsor proposes one monthly survey because it is cheap. In the first week, instructors submit 35 reports about confusing wording and three reports about incorrect prerequisite advice; students who abandoned the service are absent from the survey sample.
The architect and academic service owner choose daily pilot triage with a 30-minute analyst allocation, a weekly cohort review including unsuccessful users, and an immediate route for advice that could prevent enrollment. The academic owner adjudicates the prerequisite cases before engineering changes instructions. The wording reports become a grouped theme rather than 35 separate feature promises.
Discovery retains twice-weekly requirement reviews until the departments agree on the supported questions. At production handoff, the service owner keeps weekly theme triage and a monthly sponsor review, while support retains the urgent route. Each submitter receives a disposition, including declined requests. The monthly-only alternative saves meeting time but can leave a consequential defect unresolved for most of the pilot; the chosen plan spends limited analyst time on impact rather than report count.
Common exam traps
- Treating a shared inbox as a complete feedback process.
- Using one meeting cadence for discovery disagreements, pilot defects, and urgent production incidents.
- Letting request volume replace impact-based adjudication.
- Equating acknowledgement with resolution or promising every suggestion a release date.
- Reporting themes to sponsors without returning decisions to the people who submitted them.
What to remember
Define channels, cadence, ownership, adjudication, and return communication for each lifecycle phase. Keep urgent escalation separate from routine review, include people whose workflows failed, and use a disposition register to show that feedback reached an accountable decision.
Knowledge check
An insurer runs a four-week pilot of a Claude-assisted claims helper with three analysts. In the first week, staff file dozens of reports about awkward wording and two reports of advice that could lead a customer to miss a filing deadline. The sponsor proposes one monthly survey because it is cheap. Which feedback plan should the architect propose for the pilot?
During a pilot, an instructor reports that the service gave wrong prerequisite advice. Engineering closes the case as resolved because the logged Claude API call completed without an error and its request-id is on file. The instructor disputes the answer itself, not the call. What should the feedback process require before the case closes?
request-id and waits for its answerIn week three of a service-desk pilot, volunteered reports fall almost to zero. Usage logs show that the share of eligible tickets attempted with the assistant has fallen by half over the same period. The sponsor reads the quiet inbox as a sign that the early defects are fixed. What should the architect do?