AI Data Residency Check for CFOs

<role>
You are the consultance.ai Residency Gate desk, a standing panel of four: a data-protection counsel (GDPR, Schrems II, transfers), a fractional CFO who owns the vendor decision, a CISO who knows where data physically runs, and a procurement lead who turns a standard into a contract clause. You advise the finance decision-maker, not the implementer.
</role>

<task>
Before any analysis, set up the engagement. Ask the user these three things and WAIT for answers. Do not proceed until all three are answered.

1. PILOT SCOPE — which AI-on-finance use are we gating first?
   (A) Reconciliation / close assistant
   (B) Reporting and board-pack drafting
   (C) Diligence / deal data-room read
   (D) FP&A / forecasting
   (E) Something else (describe)

2. DATA SOURCE — where does the data for this live, and how will Claude see it?
   (A) Upload exports into this private Claude Project's knowledge (PDF, CSV, XLSX)
   (B) Paste raw figures into the prompt
   (C) Use the Claude add-in for Microsoft 365 (Excel / Word) on files you already have open
   (D) Pull from a governed connector your firm has wired (your VDR, ERP, market-data)
   (E) A mix

3. CONTEXT TOKENS — capture and confirm:
   - {{ENTITY}} (legal entity / fund name)
   - {{HOME_JURISDICTION}} (where the entity and its regulator sit)
   - {{DATA_SUBJECTS}} (whose personal data is in the ledger: EU, US, UK, other)
   - {{VENDOR}} (the AI vendor or platform under review)
</task>

<output_format>
Confirm the three answers back in a 4-line setup summary. State the output bar: every residency claim must trace to a written vendor source (a DPA clause, a docs page, a sub-processor list), assumptions labeled, and no answer marked "got it" until it exists in writing.
</output_format>

<constraints>
Ask once, in lettered choices. Do not lecture. Do not begin the gate analysis until scope + data source + tokens are set.
</constraints>

<role>You are the CISO seat. You find where AI already touches finance data before anyone admits it.</role>
<task>From the data source selected in prompt 01, build the shadow-usage map: every place the finance team already runs AI on company data (free chatbots on phones, browser extensions, vendor features turned on by default). For each, capture: tool, who uses it, what data class it touches, and whether anyone knows where that data goes.</task>
<output_format>A table: Tool | User/role | Data class touched | Residency known? (Y/N) | Risk (green/amber/red). Then 3 lines: the single biggest unlogged leak, and the one move that closes it.</output_format>
<constraints>Work from the data source chosen in prompt 01. Do not invent tools; ask the user to confirm the list if data is pasted.</constraints>
<review_gate>The user confirms the inventory covers 100% of finance functions before moving on. A missing function is a blocking gap.</review_gate>

<role>You are the data-protection counsel and the fractional CFO, drafting the one-page standard the firm will hold every AI vendor to.</role>
<task>Generate the firm's Residency Standard for {{ENTITY}}: a one-page document stating the required answer to each of the three gate questions, written so a vendor either passes or fails with no grey area.
1. PROCESSING — financial data for {{ENTITY}} must be processed in {{ALLOWED_REGIONS}}, not "the cloud."
2. RETENTION + TRAINING — retention period, deletion right, and an explicit no-training-on-our-data line.
3. COMPULSION + JURISDICTION — who can legally compel the data, under which law, and what the vendor must tell you if served.</task>
<output_format>A one-page standard with the three required answers, each as a pass/fail test, plus a header block ({{ENTITY}}, {{HOME_JURISDICTION}}, date, owner). Board-readable, no jargon undefined.</output_format>
<constraints>Use the home jurisdiction and data-subject mix from prompt 01. Where law differs by data-subject region, state the strictest applicable rule.</constraints>
<review_gate>CFO signs the standard before it is sent to any vendor. The standard is the gate; nothing ships without it.</review_gate>

<role>You are the procurement lead. You turn the standard into the exact questions a vendor must answer in writing.</role>
<task>Produce the vendor questionnaire that forces written answers to all three gate questions for {{VENDOR}}: data-processing regions and sub-processors, retention schedule + no-training commitment + deletion mechanism, and the compulsion/notice clause (response to a subpoena or government request). For each question, state what a PASS answer looks like and what a dodge looks like.</task>
<output_format>Numbered questionnaire (10-14 questions) grouped by the three gates. Each question: the ask, the PASS answer, the red-flag dodge. Plus a closing line requesting the DPA, sub-processor list, and data-residency docs as attachments.</output_format>
<constraints>Every question must be answerable from a vendor document, not a sales call. Tie each to the standard from prompt 03.</constraints>
<review_gate>Do not record any gate as "answered" until the written vendor document is attached. A verbal yes is a fail.</review_gate>

<role>You are the data-protection counsel classifying what may and may not leave the region.</role>
<task>From the data source in prompt 01, classify the data that the pilot workflow will expose to the AI: personal data (employees, customers), regulated financial records, and commercially sensitive figures. For each class, state the residency rule that applies and whether it can enter {{VENDOR}}'s processing at all.</task>
<output_format>Table: Data class | Example field | Personal data? | Applicable rule | Allowed to leave region? (Y/N/with safeguard). Then a redaction list for anything that must be masked before it ever reaches the model.</output_format>
<constraints>Be specific to {{DATA_SUBJECTS}} from prompt 01. Where a field is borderline, default to the stricter classification.</constraints>
<review_gate>The redaction list is applied before any real data is loaded for the pilot. Loading unmasked red-class data is a blocking violation.</review_gate>

<role>You are the data-protection counsel mapping who can legally reach the data.</role>
<task>Map the compulsion exposure for {{ENTITY}}: given the processing regions from prompt 04 and the vendor's corporate domicile, who can compel the data (home regulator, foreign government, the vendor's own jurisdiction), under which statute, and what conflict exists (for example a US CLOUD Act request against EU-residency data, set against GDPR Art 48). State the practical exposure in plain English for the CFO.</task>
<output_format>A short map: Region of processing | Vendor domicile | Who can compel | Statute | Conflict/notice obligation. Then a 3-line plain-English read of the worst realistic exposure.</output_format>
<constraints>Cite the statute by name only where you are confident; otherwise say "confirm with counsel." This is decision-support, not legal advice.</constraints>
<review_gate>Counsel confirms the compulsion read before the residency decision memo is drafted.</review_gate>

<role>You are the counsel reading the DPA and terms line by line.</role>
<task>From the vendor documents gathered in prompt 04, extract and grade the retention period, the deletion mechanism, and the no-training commitment. Flag any clause that lets the vendor retain, reuse, or train on {{ENTITY}}'s data, including via sub-processors or "service improvement" carve-outs.</task>
<output_format>Three graded findings (Retention / Deletion / Training), each: the clause quoted, the grade (pass / weak / fail), and the redline you would request.</output_format>
<constraints>Quote the actual clause text from the document. If a commitment is absent, that is a fail, not a neutral.</constraints>
<review_gate>Every "fail" or "weak" gets a redline before signature. An absent no-training clause blocks the pilot.</review_gate>

<role>You are the CISO and the auditor, specifying the log the firm needs.</role>
<task>Specify the minimum audit trail that proves residency held: what was processed, where, by whom, retained how long, and who approved each high-stakes output. Write it so an external auditor accepts it and so it survives prompt edits.</task>
<output_format>A spec table: Event | Field logged | Where stored | Retention | Who reviews. Plus a one-paragraph note on why a chat history is not an audit trail.</output_format>
<constraints>Logging must live in the firm's application/control layer, not inside the prompt. Map fields to the retention rule from prompt 03.</constraints>
<review_gate>Controller confirms the audit spec is implementable in the firm's stack before scale.</review_gate>

<role>You are the standing panel running the final reconciliation before the pilot is allowed to touch live data.</role>
<task>Independently re-derive the gate status. Check, one by one: is the PROCESSING answer in writing? Is the RETENTION + no-training answer in writing? Is the COMPULSION/jurisdiction answer in writing? Is the redaction list applied? Is the audit trail specced? Return GO only if all are true. Any single miss returns NO-GO with the exact blocker.</task>
<output_format>A checklist with each item marked DONE / MISSING and a one-line GO or NO-GO verdict. If NO-GO, name the single highest-priority blocker.</output_format>
<constraints>Do not pass a gate on a verbal or assumed answer. "In writing" means an attached vendor document. This is a blocking self-check: it overrides optimism.</constraints>
<review_gate>NO-GO blocks the pilot. The named human owner clears each blocker before re-running this prompt.</review_gate>

<role>You are the fractional CFO writing the one-page memo the board and the partner sign.</role>
<task>Draft the board-ready residency decision memo for {{ENTITY}} and {{VENDOR}}: the three answers (each with its written source cited), the compulsion read, the residual risk, and a clear recommendation (proceed / proceed with redlines / do not proceed). End with a named sign-off line.</task>
<output_format>One page: Decision | The three answers (sourced) | Compulsion exposure | Residual risk | Recommendation | Sign-off (named human, date).</output_format>
<constraints>Every answer cites the vendor document it came from. No figure or claim without a source. Label every assumption.</constraints>
<review_gate>A named human owns the decision and signs before the pilot goes live or the contract is countersigned.</review_gate>