CertSafari
    Anthropic Claude Certified Associate – Foundations (CCAO-F) - December 7 onwards· Lessons

    Domain 6 · Lesson 21/21

    AI policy compliance: diagnosing gaps and choosing the corrective action

    Policy, Ethics, and Diligence

    17 min read
    5.67% of exam
    12 sources
    Published 18 Nov 2026
    Docs as of 4 Nov 2026

    What you will be able to do

    • Compare observed Claude usage with specific clauses of an organizational AI policy and Anthropic's usage policy, and name each deviation
    • Recognize bias and transparency failures in AI-assisted work, including where no written rule covers the case
    • Classify a deviation as a policy gap, a training gap or a diligence lapse from the evidence
    • Sequence an audit and an incident response so that the corrective action matches the cause and closure is proven

    Key concept

    Diligence — The person who uses AI is accountable for what it produces and for how it was used: which tool, how honestly its role is disclosed, and whether the output was verified before anyone relied on it. A policy breach is a failure of diligence somewhere, either in the rules, in what people were taught, or in an individual's choices.

    1.What diligence and usage policy require

    An organizational AI policy turns good intentions into rules someone can check: which tools are approved, what may be shared, what must be disclosed, and who reviews output before it leaves the building. Every such rule rests on one principle. Whoever uses AI owns what it produces, so 'Claude drafted it' is never a defence for a wrong number in a proposal or an undisclosed post on the company site.

    Diligence splits into three checks, and most policy clauses map onto one of them. When you audit usage, labelling each observation with the check it fails tells you what kind of rule is involved.

    The three Diligence checks and what a lapse looks like in observed usage
    CheckWhat it asksA lapse you might observe
    Creation DiligenceBe intentional about which tools you useCustomer claim files pasted into an AI tool the organization has not approved
    Transparency DiligenceBe honest about AI's roleAI-drafted content published externally with no disclosure where policy requires one
    Deployment DiligenceVerify and vouch for outputsA Claude-drafted proposal sent to a client with an unchecked cost figure

    Anthropic's Usage Policy is the floor that organizational policy builds on. It singles out high-risk use cases that carry additional requirements, and it requires organizations to help their own users understand when they are interacting with an AI system. For sensitive information and decision making, Anthropic recommends that a qualified professional review content before it reaches consumers. Which uses count as high-risk is covered under use-case appropriateness. This lesson starts from what a practice must do once it is already in that category.

    Take a wealth-management team that emails Claude-drafted, individualized investment recommendations straight to retail clients, with no review by a licensed advisor and no mention of AI. The practice fails twice: nobody qualified vouches for the advice, and clients are not told AI was involved. A corrective action has to restore both. A licensed advisor reviews and approves each recommendation before it is sent, and the communication discloses AI's role. Fixing only one of the two leaves the practice out of line.

    Checkpoint 1 of 8· Match them up

    An audit turns up these three observations. Match each Diligence check to the observation that fails it.

    Tap a term, then the definition that fits it.

    Sources1234

    2.Bias and transparency failures

    A compliance gap breaks a written rule. An ethical risk can hurt people whether or not a rule covers it. Two come up repeatedly in AI-assisted work: bias, where outputs systematically disadvantage a group, and transparency failures, where people are misled about AI's role or about where content came from.

    Bias shows up in patterns, not in single outputs. Suppose a support team uses Claude to assign priority to incoming tickets. A monthly review shows that tickets written in non-standard English are consistently ranked lower than tickets reporting the same fault in polished English. No single ticket looks wrong. The harm only becomes visible when outcomes are compared across groups, which is why a policy that requires a fairness check of AI-assisted triage matters, and why skipping that check is itself a finding. The response has two parts. First, run the required check, comparing priority for the same fault across writing styles. Second, put a human reviewer on the affected triage decisions until the pattern is fixed, so customers stop being under-served in the meantime.

    Transparency failures range from omission to fabrication. Leaving out a required disclosure on external content is one end of that range. Posting 25 Claude-written five-star testimonials as if real customers wrote them is the other end: it invents the source itself, and calling it a missing label understates it. For anything external-facing, users should be told they are interacting with an AI system, and customers should be able to tell whether they are dealing with an AI agent or a human.

    Policy cannot list every case, so judgment covers the gaps. If the same judgment call keeps coming up in an audit, for example whether AI-assisted summaries sent to partners need disclosure, the repetition tells you a clause is missing.

    Checkpoint 2 of 8· Check yourself

    A housing association uses Claude to set urgency on tenant maintenance requests. A quarterly review finds that requests sent through the Spanish-language form get lower urgency than English-form requests reporting the same fault. The team had also skipped the fairness check its AI policy requires. What response fits this finding?

    Sources5146

    3.Auditing usage against the policy, evidence first

    An audit of Claude usage follows two rules: every conclusion rests on observed evidence, and every corrective action matches a diagnosed cause. Breaking either rule produces the most common failure, which is picking a fix (usually retraining) before anyone knows why the behaviour happened. The steps run in this order:

    1. Collect examples of actual usage from the audit period, such as shared chats, Project settings and connector permissions.
    2. Compare each observed pattern with the specific policy clause it touches, and mark where behaviour deviates from the written rule.
    3. Diagnose the cause of each deviation as a policy gap, a training gap or a diligence lapse.
    4. Choose the corrective action that matches the cause: revise a clause, train, or add a review step.
    5. Re-audit after the action using the same measure, to confirm the deviation has actually stopped.

    Evidence comes first because memories and opinions about how staff use Claude are unreliable. On Enterprise plans the Compliance API gives compliance teams programmatic access to usage data and customer content, so the audit can sample what really happened. Comparing against a specific clause keeps the finding precise. A well-built policy answers three questions: where AI fits and where it does not, what is disclosed to whom, and when the policy will be revisited. Each observation should point to the clause that answers one of them.

    The last step is the one most often skipped. Imagine a gap on connector use closed with a new policy section and a short staff briefing. Attendance at the briefing proves people heard something. It does not prove behaviour changed. The strongest evidence that the gap is closed is a follow-up audit that applies the same measure as the original and shows the deviation is gone. A policy without revision triggers also goes stale without anyone noticing, so the re-audit doubles as the point where the policy itself is reviewed.

    Checkpoint 3 of 8· Check yourself

    Three weeks after a connector-use gap was found, the compliance lead added a new policy section and ran a staff briefing. Which evidence best shows the gap is closed?

    Checkpoint 4 of 8· Exam question

    A compliance lead is asked to analyze how staff use Claude against the organization's AI policy, and to correct any problems found. Which sequence of steps should the lead follow?

    Sources10111

    4.Policy gap, training gap or diligence lapse

    The same visible deviation can have three different causes, and each needs a different fix. The test is what the evidence shows about the rule and about the people:

    Diagnosing the cause of a deviation and the corrective action that matches it
    CauseWhat the evidence showsCorrective action that closes it
    Policy gapThe rule is missing, unclear, or contradicted by the organization's own templates or standing instructionsRevise the clause or configuration, then tell staff what changed and where to find it
    Training gapThe rule is clear, but people did not know it or could not apply itTrain with concrete examples, then re-audit
    Diligence lapseThe rule was clear, the person was trained, and they knowingly did otherwiseHold the individual accountable, and add a review step where the stakes are high

    A Claude Project is the same case as the template. If its standing instructions say never to mention that content was drafted with AI, and the policy requires disclosure on published content, the staff who obeyed the Project behaved predictably. Rewrite the instructions to match the policy (see persistent-instruction design) and check other shared Projects for the same conflict.

    The reverse case is a product manager who admits the policy forbids fabricated endorsements, that training covered it, and that they generated fake testimonials anyway. A training refresher would teach them nothing they did not already know. The response is individual accountability and removal of the content.

    Gaps can also sit in the parts of a policy that cover raising concerns rather than using AI. Picture an insurance broker where several employees watch colleagues paste full claim files, customer personal details included, into unapproved tools, and nobody reports it. The data rules are clear. What the policy lacks is any owner, inbox or process for raising a concern, so a protocol staff are expected to follow does not exist. That is a policy gap. Closing it means adding a named owner and a simple reporting route, stating that people who report will not face retaliation, and telling staff where to find all of it. Re-running data-handling training repeats rules people already understand. Asking managers to act informally leaves the route unwritten. Monitoring every device detects pasted data but still gives staff no way to raise a concern.

    Checkpoint 5 of 8· Check yourself

    In an interview, a marketing analyst confirms three things. The policy requires disclosure on published AI-drafted content. Their training covered that rule. The team's shared Project has no instruction that conflicts with it. They still left the AI notice off three Claude-drafted blog posts so that the posts would read as more personal. Which corrective action matches the cause?

    Checkpoint 6 of 8· Exam question

    Several employees at an insurance broker notice colleagues pasting full claim files, including customers' personal details, into unapproved AI tools. None of them reports it. When asked why, they say the AI policy names no owner, inbox or process for raising concerns, though its data rules are clear. What corrective action closes this gap?

    Sources786

    5.Responding when a lapse has already reached someone

    Once a lapse has left the building, the order of the response is set by harm. The client may already be relying on the wrong number, so containing that comes before working out how it happened. Then scope, then cause, then prevention:

    1. Contact the client, flag the figure as unreliable, and send a corrected version with the verified number.
    2. Search other proposals from the same author and Project for similar unverified figures that may already have been sent.
    3. Interview the author to find out whether the verification step was unknown, unclear or knowingly skipped.
    4. Update the policy, template or training to remove the cause that was found, and brief the team on the change.

    Scope comes before cause because the same habit has probably produced more than one document, and each of those is still doing harm while you investigate. The interview in step 3 uses the diagnosis above: an unknown step points to a training gap, an unclear step to a policy gap, and a skipped step to a diligence lapse. Step 4 can only be chosen after that answer. Skipping ahead to 'retrain everyone' fixes a cause nobody has confirmed. How to verify figures before they ship is covered under hallucination detection and verification. The governance point here is ownership. The person who sent the proposal is accountable for the number in it, whoever drafted it.

    Checkpoint 7 of 8· Check yourself

    A proposal with a wrong, unverified cost figure reached a client yesterday. Its author points out that Claude drafted the number and that the team's Project template never mentions verification. What should the compliance lead do?

    Checkpoint 8 of 8· Exam question

    A Claude-drafted proposal containing an unverified, wrong cost figure was emailed to a client yesterday. The team lead wants a response plan. Which sequence of steps is the most appropriate response?

    Sources917

    Sources

    (12)

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

    1. 1.
      “You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
      ↩︎ What diligence and usage policy require
      “Access is a design decision. Check who your assumptions exclude before you call something shipped.”
      ↩︎ Bias and transparency failures
      “Make the call: ship it, fix it, or stop it”
      ↩︎ Responding when a lapse has already reached someone
      “Feedback loop — How will you know if it’s working after it ships?”
      ↩︎ Auditing usage against the policy, evidence first
      “You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
      ↩︎ Exam trap 1
      “Responsibility — Could this output be misread or misused? Have you been transparent about AI’s role?”
      ↩︎ Exam trap 2
      “Feedback loop — How will you know if it’s working after it ships?”
      ↩︎ Exam trap 3
      “Feedback loop — How will you know if it's working after it ships?”
      ↩︎ Checkpoint
      “Responsibility — Could this output be misread or misused? Have you been transparent about AI’s role?”
      ↩︎ Prediction
      “You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
      ↩︎ Checkpoint
      “You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
      ↩︎ Checkpoint
    2. 2.
      “Transparency Diligence (being honest about AI's role), and Deployment Diligence (verifying and vouching for outputs)”
      ↩︎ What diligence and usage policy require
      “Creation Diligence (being intentional about which tools you use), Transparency Diligence (being honest about AI's role), and Deployment Diligence (verifying and vouching for outputs)”
      ↩︎ Checkpoint
    3. 3.
      “We’ve broken out some specific “high-risk use cases” that have additional requirements due to posing an elevated risk of harm.”
      ↩︎ What diligence and usage policy require
    4. 4.
      “For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”
      ↩︎ What diligence and usage policy require
      “For external-facing products, disclose to your users that they are interacting with an AI system.”
      ↩︎ Bias and transparency failures
      “For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”
      ↩︎ Exam trap 4
    5. 5.
      “Bias: When might AI reinforce narrow perspectives, stereotypes, or existing inequities?”
      ↩︎ Bias and transparency failures
      “Using AI responsibly isn't just about following rules; it's about ethical judgment in the gray areas where policy gives no clear answer.”
      ↩︎ Bias and transparency failures
    6. 6.
      “Make it clear when they're interacting with an AI agent versus a human”
      ↩︎ Bias and transparency failures
      “Staff need to understand how agents work, when to trust their recommendations, and how to escalate concerns.”
      ↩︎ Policy gap, training gap or diligence lapse
    7. 7.
      “follow your organization's protocols”
      ↩︎ Policy gap, training gap or diligence lapse
      “If something goes wrong, you have options: Delete the conversation, request data deletion through the platform's privacy process, and follow your organization's protocols”
      ↩︎ Responding when a lapse has already reached someone
    8. 8.
      “Which Diligence check did you find yourself wanting to skip, and what does that tell you?”
      ↩︎ Policy gap, training gap or diligence lapse
    9. 9.
      “You remain responsible for all actions taken by Claude performed on your behalf.”
      ↩︎ Responding when a lapse has already reached someone
    10. 10.
      “Three questions anchor the policy: where AI fits and does not, what you disclose and to whom, and when you will revisit.”
      ↩︎ Auditing usage against the policy, evidence first
      “A policy without revision triggers goes stale silently.”
      ↩︎ Auditing usage against the policy, evidence first
    11. 11.
      “compliance teams get real-time programmatic access to Claude usage data and customer content”
      ↩︎ Auditing usage against the policy, evidence first

    Also cited

    Ready to test yourself?

    Practise Anthropic Claude Certified Associate – Foundations (CCAO-F) - December 7 onwards in quiz mode.

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