CertSafari
    Anthropic Claude Certified Architect – Professional (CCAR-P)· Lessons

    Domain 6 · Lesson 33/38

    Stakeholder expectations: SLAs, limits and pilot feedback

    Manage stakeholder feedback loops and expectation alignment

    19 min read
    2.8% of exam
    13 sources
    Published 19 Oct 2026
    Docs as of 14 Oct 2026

    What you will be able to do

    • State what the standard and Priority service tiers promise about availability, and what they leave out
    • Explain how organisation rate limits and spend caps limit throughput and cost, and recognise each error a stakeholder will see
    • Collect the model-selection criteria stakeholders should sign off before an SLA, and reopen them after launch
    • Name the right channels for Anthropic-side incidents and for model retirement notices, including how partner platforms differ
    • Set up a pilot feedback loop that collects, triages and reports back from day one
    • Tell a real quality drop from normal variation before reacting to a complaint

    Key concept

    Ceilings and targets, not guarantees — The service-level facts a Claude deployment relies on are maximums and targets: rate limits and spend caps cap usage without promising capacity, and only Priority Tier comes with an uptime target. Any expectation you set with stakeholders has to be built on these ceilings, not on numbers nobody is committed to.

    1.What standard and Priority Tier actually promise

    Stakeholders rarely ask about service tiers directly. They ask "will it be up?", and the honest answer depends on which tier the traffic runs on. Anthropic offers three tiers. Standard is the default for every API request. Priority Tier needs an existing capacity commitment. Batch is for asynchronous work that can wait. Promising availability you have not paid for is the most common way an SLA conversation goes wrong later.

    Standard promises best effort and nothing more. Its requests are prioritised alongside everyone else's, so a busy period can return "server overloaded" errors and there is no uptime figure to quote. Priority Tier puts your requests ahead of all other traffic, which helps minimise overload errors even at peak times, and it targets 99.5% uptime. Two words in that sentence matter. "Targets" is not "guarantees". And the target covers only the capacity you committed to: requests beyond it automatically fall back to standard tier.

    What each service tier offers when you set availability expectations
    TierWho gets itWhat it promisesWhat it does not cover
    StandardEvery request by defaultBest-effort availability, prioritised alongside all other requestsNo uptime target; overload errors are possible at peak
    Priority TierOrganisations with a capacity commitmentTargets 99.5% uptime with prioritised computeTraffic above committed input/output tokens per minute falls back to standard
    BatchAsynchronous workflowsWork that can wait or run outside normal capacityNot for interactive, latency-sensitive paths

    A Priority Tier commitment is specific. It names input tokens per minute, output tokens per minute, a duration of 1, 3, 6 or 12 months, and one specific model version. Cache reads and writes, along with US-only inference on newer models, count against that capacity at their own weightings, so a heavily cached workload uses less commitment than its raw token count suggests. Check which model you plan to run before you promise Priority Tier. It is not supported on several current models, including Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. A stakeholder who has been told "we'll get the 99.5% tier" for an Opus 5.5 workflow has been told something that cannot happen.

    Checkpoint 1 of 9· Match them up

    A stakeholder asks "will it be up?" Match each case to what you can honestly promise.

    Tap a term, then the definition that fits it.

    Sources1

    2.Rate limits and spend caps are ceilings

    Two organisation-level limits set how much a deployment can do: rate limits cap requests and tokens per period, and spend limits cap monthly cost. Both come from your usage tier. Organisations are placed on a tier automatically from their usage history and move up over time, and new organisations may start in a lower Evaluation tier. These limits are enforced across the whole organisation. You can set lower limits of your own on individual workspaces, which is how one team's pilot is kept from using up another team's headroom.

    Rate limits on the Messages API are measured per model class as requests per minute (RPM), input tokens per minute (ITPM) and output tokens per minute (OTPM). They are enforced over shorter windows than their names suggest. A 60 RPM limit can behave like one request per second, so a burst can fail even when the minute's total is under the limit. Caching is the lever to explain to stakeholders: for most models only uncached input tokens count toward ITPM, so a high cache hit rate raises effective throughput without a tier change. A higher limit can be requested in the Console once you are using at least 50% of your current limits.

    Monthly spend cap by usage tier
    Usage tierMonthly spend cap
    Start$500 USD
    Build$1,000 USD
    Scale$200,000 USD
    CustomNo cap; limits arranged with the account team

    Hitting the spend cap does not slow usage down. It stops it. When the organisation reaches its tier's cap, API usage pauses until 00:00 UTC on the first day of the next month. A limit you set yourself below the cap behaves differently again. Stakeholders should know in advance which of these three responses they could see and who is able to fix each one.

    How each limit shows up, and who can clear it
    Limit hitResponseSignalWhat restores access
    Rate limit (RPM, ITPM or OTPM)HTTP 429 rate_limit_errorretry-after header presentWait the stated seconds, then retry
    Tier monthly spend capHTTP 429 rate_limit_errorerror_code enforced_spend_limit_reached, no retry-afterNext month at 00:00 UTC, or a higher tier
    Spend limit you setHTTP 400 invalid_request_errorMessage begins You have reached your specified API usage limitsRaise or remove the limit

    Checkpoint 2 of 9· Check yourself

    Your organisation's usage tier gives a pilot team a high RPM and ITPM rate limit on the Messages API, and all of its traffic runs on the standard tier. The product owner wants to tell users that throughput up to that limit is assured. What can you honestly commit to?

    Sources23

    3.Criteria stakeholders sign off before an SLA

    An SLA is only as good as the criteria behind it. Write them so they can be measured, as covered in Measurable success criteria, and then have stakeholders sign them. The work in a sign-off session is turning "it should be fast and cheap" into choices stakeholders own. Choosing a Claude model means balancing three factors, and each one is a question to put to the business, not one for engineering to answer alone.

    Ask these three questions in the session:

    • Capabilities: which features the model must have to meet the workflow's functional requirements.
    • Speed: how quickly the application must respond, and whether premium pricing is acceptable for it. Fast mode on Claude Opus 5.5, Claude Opus 5 and Claude Opus 4.8 delivers up to 2.5x higher output speed at premium pricing, which is a business decision.
    • Cost: the budget for both development and production usage, which also sets the usage tier and spend cap you will need.

    Architecture details such as parameter counts or where the weights are hosted are not selection criteria. Stakeholders can't evaluate them and the model guidance doesn't ask for them. Uptime and response time, on the other hand, belong in the signed criteria as operational metrics, worded as targets that match the tier you are actually buying.

    Signed criteria need to be reopened after launch, because the model underneath them changes. Prompts and guardrails tuned for one model can hold back a more capable one, so teams retest tasks as models change and improve. An eval suite keeps that renegotiation honest. Early on it makes the team define what success means, and later it holds the system to a consistent quality bar. A request to relax latency in exchange for a stronger model, or to accept a new model after a retirement, is then decided against numbers rather than impressions.

    Checkpoint 3 of 9· Check yourself

    Six months after stakeholders signed the success criteria, the team moves the workflow to a newer, more capable Claude model and keeps the original prompts and guardrails. A stakeholder asks whether the signed criteria still hold. What should you do?

    Checkpoint 4 of 9· Exam question

    Before committing to an SLA for a new agentic workflow, a delivery lead wants to run a structured feedback session with business stakeholders to align on model selection criteria, following Anthropic's guidance on choosing a model. Which factors should the lead explicitly gather from stakeholders during this session? (Select all that apply)(Select 3)

    0 of 3 selected

    Sources4567

    4.How stakeholders hear about outages and model retirements

    Every customer's incident runbook should name one public source for Anthropic-side incidents: status.claude.com. It gives real-time updates on the scope, impact and resolution progress of confirmed incidents. Account and billing questions go through a separate channel, Support, which you reach from the message icon in the Help Center or from "Get help" in the Claude Console. Keep the two apart in the handbook. One reports what is happening to everyone. The other handles your organisation's own case.

    Some signals look like outages but are not. Capacity constraints, where demand is high but the system is working normally, do not appear on the status page, because they are normal load management. A 429 means your organisation hit a limit; it is not evidence of an incident. Internal dashboards, rate-limit charts and model system cards all have their uses, but none of them is where Anthropic announces incidents.

    Model retirement is the other planned event that changes a committed timeline. Anthropic notifies affected customers by email and in the documentation, giving at least 60 days' notice before a publicly released model is retired. Deprecated models still work but have a replacement and a retirement date. Retired models fail every request. You can find which keys still call a deprecated model by exporting the Console Usage page as a CSV broken down by API key and model.

    Which platform the traffic runs on decides which retirement date applies. Anthropic's published dates cover the Claude API, Claude Platform on AWS and Microsoft Foundry. Amazon Bedrock and Google Cloud run their own lifecycle, so the same model can be retired on the Claude API and still be available on a partner platform. Direct API traffic has to move to a supported model before Anthropic's date. Traffic on a partner platform follows that platform's timeline. Existing traffic gets no exemption, and retirement blocks API requests, not only the Console. How to run the migration itself is covered in Lifecycle phases and model migration.

    Checkpoint 5 of 9· Check yourself

    At peak hours, a pilot running on the standard tier starts returning "server overloaded" errors. The product owner wants to tell users that Anthropic is having an outage and point them to status.claude.com. What should you tell the product owner?

    Checkpoint 6 of 9· Exam question

    A customer support tooling vendor still runs Claude Haiku 3.5 in a legacy component and receives Anthropic's notice that the model is retired outside of Bedrock and Google Cloud. Their engineering stakeholders want to know how this affects their committed timeline. What is the accurate expectation to set?

    Checkpoint 7 of 9· Exam question

    An account manager is drafting an incident-communication annex for a customer's stakeholder handbook, describing how they will learn about Anthropic-side outages affecting their production deployment. Which resource should the annex point to as the authoritative, real-time source, separate from the account-specific support channel?

    Sources8910

    5.Building the feedback loop into the pilot

    Decide how feedback will be collected and used before the pilot starts. Feedback gathered as an afterthought arrives as scattered complaints with no context. The pilot invitation is the first part of the loop. Tell the cohort they were chosen because they will use it on real problems and tell you the truth about it. Then give them a specific request: use it on at least one real task, and post what worked, what was annoying and what surprised them. Close by telling them what their input is for, so they know it has consequences.

    Collection needs one place and one owner. A single channel gives questions one place to land. Name an owner to watch it for the first 48 hours, because unanswered launch-day questions kill momentum. Answer in public so each answer benefits everyone watching. A weekly prompt keeps feedback coming without a meeting: each Friday, ask what the tool helped with this week, and accept screenshots or a couple of sentences.

    Triage and reporting back keep the loop from becoming a suggestion box. Sort each item before acting on it. Is it a capability gap, a limit or quota problem, a prompt or configuration fix, or a quality complaint that needs checking against evals? Then report the outcome back to the people who raised it. Usage data adds a signal nobody has to type. Analytics show adoption trends and dips in usage that may indicate friction, which tells you where to ask questions. One Anthropic team kept the human side manageable with a weekly "lessons & missteps" report and by batching questions so a reviewer could answer them in one pass. Reviewer attention runs out, so the loop should use it carefully.

    A pilot feedback loop: each stage and what it produces
    StageMechanismOutput stakeholders see
    CollectDedicated channel with a named owner; ask for what worked, what was annoying, what surprisedEvery question has an answer, in public
    PromptWeekly show-and-tell threadA running record of real wins and failures
    ObserveUsage analytics: active users, sessions, spendDips that flag friction before anyone complains
    Report backWeekly summary of decisions taken from feedbackEvidence that feedback decides the wider rollout

    Checkpoint 8 of 9· Match them up

    You are designing the feedback loop for a Claude pilot before it launches. Match each mechanism to the problem it solves for stakeholders.

    Tap a term, then the definition that fits it.

    Sources1112136

    6.Separating real degradation from normal variation

    "It feels worse this week" is the complaint every pilot gets eventually, and responding before checking is how teams lose a week. Without evals, debugging is reactive. You wait for complaints, reproduce them by hand, fix something, and hope nothing else regressed. The underlying problem is that a team without evals can't distinguish real regressions from noise. A model gives different answers to similar inputs, so a few bad outputs are not evidence that anything changed.

    Check causes in a fixed order, cheapest first:

    1. Confirmed incident: is there an active entry on status.claude.com? If so, the cause is outside your system and the status page tells you scope and progress.
    2. Capacity or limits: are errors overload responses or 429s? Capacity constraints are not outages, and limit errors appear in the table above. Neither is a quality problem.
    3. Measured regression: does the eval suite, run on the current build, score below its baseline? Only this result justifies a rollback or a prompt change.
    4. Normal variation: if none of the above applies, add the reported examples to the eval set and tell the stakeholder what you measured.

    Set the expectation for variation in the signed criteria, before anyone complains. Consistency is a criterion of its own: how similar responses need to be for similar inputs, and how much it matters that the same question asked twice gets a semantically similar answer. A stakeholder who signed "answers may vary in wording but not in substance" has a standard to judge a complaint against. Without one, every unexpected answer looks like a regression.

    Checkpoint 9 of 9· Check yourself

    On Wednesday, after a prompt change shipped on Monday, several pilot users say answers "feel worse". status.claude.com shows no active incident, and there are no overload errors or 429s. The team lead wants to roll back Monday's prompt change straight away. What should decide whether you roll back?

    Sources785

    Sources

    (13)

    Every claim above is drawn from one of these pages, quoted as it was written on the date shown.

    1. 1.
      “The API prioritizes these requests alongside all other requests with best-effort availability.”
      ↩︎ What standard and Priority Tier actually promise
      “This prioritization helps minimize "server overloaded" errors, even during peak times.”
      ↩︎ What standard and Priority Tier actually promise
      “Priority Tier targets 99.5% uptime with prioritized computational resources.”
      ↩︎ What standard and Priority Tier actually promise
      “A commitment duration (1, 3, 6, or 12 months)”
      ↩︎ What standard and Priority Tier actually promise
      “Priority Tier is supported on all available Claude models except Claude Fable 5.1, Claude Haiku 5.5”
      ↩︎ What standard and Priority Tier actually promise
      “Requests beyond your committed capacity automatically fall back to standard tier.”
      ↩︎ Exam trap 1
      “Priority Tier targets 99.5% uptime with prioritized computational resources. Requests beyond your committed capacity automatically fall back to standard tier.”
      ↩︎ Checkpoint
      “The standard tier is the default service tier for all API requests. The API prioritizes these requests alongside all other requests with best-effort availability.”
      ↩︎ Checkpoint
      “The API prioritizes these requests alongside all other requests with best-effort availability.”
      ↩︎ Checkpoint
    2. 2.
      “The API enforces service-configured limits at the organization level, but you may also set user-configurable limits for your organization's workspaces.”
      ↩︎ Rate limits and spend caps are ceilings
      “For most Claude models, only uncached input tokens count toward your ITPM rate limits.”
      ↩︎ Rate limits and spend caps are ceilings
      “When usage reaches a spend limit you set, requests return HTTP 400 with error type invalid_request_error.”
      ↩︎ Rate limits and spend caps are ceilings
      “All limits described here represent maximum allowed usage, not guaranteed minimums.”
      ↩︎ Key concept
      “Short bursts of requests can exceed the limit and trigger rate limit errors.”
      ↩︎ Exam trap 2
      “Once you reach your tier's spend cap, API usage pauses until 00:00 UTC on the first day of the next month”
      ↩︎ Exam trap 3
      “the response has no retry-after header. Retrying, including the SDK's automatic retries, fails until access resumes.”
      ↩︎ Prediction
    3. 3.
      “If you need higher limits, you can request them there once you're using at least 50% of your current limits.”
      ↩︎ Rate limits and spend caps are ceilings
    4. 4.
      “Choosing a Claude model means balancing capabilities, speed, and cost.”
      ↩︎ Criteria stakeholders sign off before an SLA
      “Capabilities: What specific features or capabilities will you need the model to have to meet your needs?”
      ↩︎ Criteria stakeholders sign off before an SLA
      “which delivers up to 2.5x higher output speed at premium pricing”
      ↩︎ Criteria stakeholders sign off before an SLA
      “Cost: What's your budget for both development and production usage?”
      ↩︎ Criteria stakeholders sign off before an SLA
    5. 5.
      “Operational: Response time (ms), uptime (%)”
      ↩︎ Criteria stakeholders sign off before an SLA
      “If a user asks the same question twice, how important is it that they get semantically similar answers?”
      ↩︎ Separating real degradation from normal variation
    6. 6.
      “retest tasks as models change and improve”
      ↩︎ Criteria stakeholders sign off before an SLA
      “batch questions to be answered in a single pass”
      ↩︎ Building the feedback loop into the pilot
      “Prompts may need re-wording and guardrails that used to be helpful may constrain a smarter model from pursuing more creative solutions.”
      ↩︎ Checkpoint
    7. 7.
      “later they help uphold a consistent quality bar.”
      ↩︎ Criteria stakeholders sign off before an SLA
      “Absent evals, debugging is reactive: wait for complaints, reproduce manually, fix the bug, and hope nothing else regressed.”
      ↩︎ Separating real degradation from normal variation
      “distinguish real regressions from noise”
      ↩︎ Separating real degradation from normal variation
      “Absent evals, debugging is reactive: wait for complaints, reproduce manually, fix the bug, and hope nothing else regressed. Teams can't distinguish real regressions from noise”
      ↩︎ Checkpoint
    8. 8.
      “To check for confirmed incidents, visit status.claude.com, where you'll find real-time updates on scope, impact, and resolution progress for any active incidents.”
      ↩︎ How stakeholders hear about outages and model retirements
      “Capacity constraints are not outages.”
      ↩︎ Separating real degradation from normal variation
      “Capacity issues will not appear on our status page because they represent normal load management rather than technical problems.”
      ↩︎ Exam trap 5
    9. 10.
      “Impacted customers will always be notified by email and in the documentation.”
      ↩︎ How stakeholders hear about outages and model retirements
      “providing at least 60 days' notice before model retirement for publicly released models”
      ↩︎ How stakeholders hear about outages and model retirements
      “Retired: The model is no longer available for use. Requests to retired models will fail.”
      ↩︎ How stakeholders hear about outages and model retirements
      “Partner-operated platforms (Amazon Bedrock and Google Cloud) set their own retirement schedules, so a model's lifecycle status and dates can differ.”
      ↩︎ How stakeholders hear about outages and model retirements
      “The dates on this page apply to Anthropic-operated platforms: the Claude API, Claude Platform on AWS, and Microsoft Foundry.”
      ↩︎ Exam trap 4
    10. 11.
      “because you will put it on real problems and tell us the truth about it.”
      ↩︎ Building the feedback loop into the pilot
      “That feedback decides how we roll it out to everyone else.”
      ↩︎ Building the feedback loop into the pilot
      “Unanswered launch-day questions kill momentum”
      ↩︎ Building the feedback loop into the pilot
    11. 12.
      “answer questions publicly so each answer benefits everyone watching”
      ↩︎ Building the feedback loop into the pilot
      “No preparation, slides, or meeting are required; screenshots and short descriptions are sufficient.”
      ↩︎ Building the feedback loop into the pilot
    12. 13.
      “Dips in usage that may indicate friction or issues”
      ↩︎ Building the feedback loop into the pilot

    Ready to test yourself?

    Practise the 12 questions on this subdomain.

    Spotted a mistake, or was something unclear? Tell us.