# IF.GOV casefile (demo input) — “Acme Support Copilot” (fictional)

**Audience:** reviewers (demo input)  
**Reality gate:** this is a fictional casefile used to demonstrate IF.GOV deliverables.  
**Privacy note:** no real customer data.

---

## 0) Decision question

Should Acme deploy an LLM-assisted “support copilot” to draft internal ticket replies and summarize customer conversations?

Target decision date: 2026-01-05 (demo)

---

## 1) Proposed system (as described by the requester)

Workflow:
1) Customer opens a support ticket (web form + email).
2) The copilot drafts a reply and a short internal summary.
3) A human support agent reviews/edits and sends the reply.

Scope (requested):
- Drafting + summarization only (no autonomous sending).
- No access to payment systems.
- No access to production databases.

---

## 2) Data involved (high level)

Inputs:
- customer-provided text (freeform; may include PII)
- prior ticket thread (may include account identifiers)
- internal support notes (may include incident details)

Outputs:
- draft reply text
- internal summary + suggested next action

---

## 3) Constraints / requirements (as stated)

Operational:
- Must not block the support team if the model is unavailable.
- Drafts must be clearly marked as drafts.
- Escalations must route to humans (no “AI decides”).

Security / governance:
- Explicit “do not paste secrets” guidance for agents.
- Vendor must support data-handling options (no-retention preferred).
- Evidence must be reviewable by a third party (receipt-first).

---

## 4) Evidence provided (demo)

The requester provides:
- a one-page internal memo describing intended use
- a list of “known bad outcomes” (hallucinated refund promises, leaked internal incident notes)
- a shortlist of 2 model vendors (unnamed in this demo)

Not provided (missing):
- evaluation dataset
- model red-teaming results
- documented incident response for “bad draft shipped”

---

## 5) Assumptions (explicit)

- Human review is mandatory and enforced in tooling (not just policy).
- The support team will adopt a “reject / edit / escalate” playbook.
- The initial rollout can be constrained to a subset of ticket types.

---

## 6) Non-goals

- No claims of compliance.
- No claims of factual correctness guarantees.
- No claims that “red teaming” eliminates risk.