The work, not the slides.

Synthetic example · Branch Manager

Subject: Branch X coverage Wed–Fri — action needed by EOD Mon

Action: Sign up for one shift on the attached coverage sheet.
Deadline: End of day Monday.
Why: Two tellers out on leave; lobby coverage gap 10a–2p Wed–Fri.

If you can't cover, reply to me directly with your role — I will pair
remaining gaps with float staff. Do not paste customer names in any
reply; lobby logistics only.

— Branch Manager X

Synthetic example · Compliance Officer

Vendor claim: "Our LLM is 99.4% accurate at AML alert triage."
Source provided: internal vendor whitepaper, no peer review.

REVIEW
✓ Number cited (99.4%) — flagged for verification, no source link
✗ "AML alert triage" — vendor did not define the alert population
✗ "Accuracy" — no precision/recall split, can't infer false-negative rate
✓ Whitepaper — internal, not third-party audited

VERDICT
Treat as marketing material. Do not pass to AML team without:
  (a) named dataset description
  (b) precision + recall (not just accuracy)
  (c) sample of false negatives reviewed by your BSA officer

Synthetic example · Lending Operations

[ROLE] You are an experienced credit analyst at a community bank.
[INPUT] Loan application denial code: {DENIAL_CODE}
        Borrower file summary (sanitized): {SUMMARY}
        Institution adverse-action standard: {INTERNAL_STANDARD}
[TASK]
  1. Write a 120-word adverse-action notice in plain English.
  2. Cite the FCRA-required disclosures inline.
  3. Use neutral, non-blaming language.
[CONSTRAINTS]
  - Never invent denial reasons not in the input.
  - Never reference race, age, marital status, source of income, or
    other protected characteristics.
  - Include the reviewer's name + date on the second line.
[REVIEW]
  Compliance officer signs off before mailing.

Synthetic example · Credit Analyst

Role: Credit Analyst II
Daily AI uses (sanitized inputs only):
  - Adverse-action letter drafts (FCRA flow)
  - Loan committee memo summaries
  - Member-facing rate explanations

Tools approved by IT: Claude (Anthropic), Gemini (Google Workspace)
Tools NOT approved: ChatGPT (personal accounts), Notion AI

Data classification:
  GREEN — public rate sheets, regulatory text, internal SOPs
  YELLOW — sanitized loan summaries (names redacted, amounts ranged)
  RED — full borrower files, credit reports, SSN/TIN, account numbers

Review checkpoint:
  Every AI-assisted artifact is reviewed by my supervisor before it
  leaves my desk. Logged in our shared review register.

Synthetic example · Compliance Manager

Project: Reg E disclosure refresh — Q3 cycle
Audience: All checking-account holders
Source context: 2026 Reg E updates + our current disclosure language
Output format: Updated disclosure document, redline against current,
               plain-English summary for member service reps

Constraints:
  - Match our institution's voice (formal, no marketing language)
  - Cite the specific Reg E section for each change
  - Flag any change that requires committee approval

Review step:
  Draft → Compliance Officer redline → Legal sign-off → Board ratify
  before member-facing distribution.

Synthetic example · Member Service Rep

Today's data I might paste:

  Member Q's complaint email about overdraft fee
    -> NO. Contains name + account ref. Sanitize first:
       remove name, keep complaint shape, paste sanitized version.

  Our overdraft policy text from intranet
    -> YES (Green). Public-equivalent internal SOP.

  Three sample call-center scripts for a coaching exercise
    -> YES (Green). No customer data, no PII.

  Yesterday's branch deposit total + breakdown
    -> NO. Operational data above member-aggregate; treat as Yellow
       and discuss with my supervisor before any AI use.

ESCALATE TO: BSA Officer for anything related to suspicious activity,
            Legal for anything member-litigation adjacent.

Synthetic example · BSA / AML Analyst

My three weekly AI-assisted tasks:

  1. SAR narrative first draft (from sanitized typology + timeline)
     - Tool: Claude
     - Review: my supervisor + BSA Officer before filing
     - Time saved: ~25 min per SAR

  2. CDD baseline drift check (compare current vs prior period)
     - Tool: Gemini
     - Review: my own verification against source records
     - Time saved: ~15 min per check

  3. Structuring pattern summary (turn raw notes into clean prose)
     - Tool: Claude
     - Review: BSA Officer
     - Time saved: ~20 min per pattern

Never use AI for: customer-facing communications, examiner work product
that hasn't been reviewed, anything with the customer's actual identity
attached to the artifact.

Synthetic example · Senior Lender

Title: Loan committee memo — initial summary
When to use it: After credit analysis is complete; before formal committee
What to paste: Sanitized credit summary (no borrower name, no full SSN)
What NOT to paste: Full borrower file, credit reports, examiner letters

Prompt:
  [ROLE] You are a senior commercial lender at a community bank
  preparing a memo for the loan committee.
  [INPUT] Sanitized credit summary: {SUMMARY}
          Loan amount range: {RANGE}
          Industry: {INDUSTRY}
  [TASK] Draft a 250-word committee memo with:
    - One-sentence recommendation
    - Three strongest points supporting it
    - Two risks the committee should weigh
    - Specific policy exceptions requested (if any)
  [REVIEW] My credit officer reviews before committee.

Example output (sanitized):
  Recommended: approve $500K-$750K LOC for a regional plumbing
  contractor with 12-year operating history and 1.8x debt-service
  coverage...

Safety notes:
  - Never include borrower name in prompt or output
  - Verify dollar ranges manually before committee
  - Refresh once per quarter against current policy