Use ChatGPT to Turn a Policy Exception Request into a Review Brief

Policy exceptions arrive as short messages. The missing details are usually the important part: which rule applies, who can approve the exception, and what evidence supports the request.

This is an illustrative ChatGPT workflow with fictional text. It is not a compliance review and does not claim a ChatGPT execution.

The fictional input

Create two files. Keeping the policy and request separate makes it easier to spot claims that came from the request rather than the rule.

travel-policy.txt

Travel reimbursement policy, version 4.2, effective 2026-09-01

Employees may claim standard rail travel for approved business trips. A manager must approve the trip before booking. Premium rail tickets require Finance approval. The policy does not state an approval route for same-day bookings or trips booked after travel has started. Receipts are required for reimbursement.

Policy owner: Finance Operations

exception-request.txt

Requestor: Elena Rossi
Request date: 2026-10-14
Request: Reimburse a premium rail ticket bought this morning for a client meeting.
Reason supplied: The standard train was cancelled and the client meeting could not move.
Evidence supplied: Ticket receipt and a screenshot of the cancellation notice.
Requested decision: Approve reimbursement.

The request says why the ticket was bought. It does not say that the manager approved the trip before booking, or who can approve a premium ticket. Those gaps should remain visible.

A copyable prompt

Open a new ChatGPT conversation, upload both files, and paste this prompt:

Review travel-policy.txt and exception-request.txt as separate sources.

Create a review brief with these sections:
1. Request summary, using only facts stated in the request.
2. Policy rules that clearly apply, with the exact supporting sentence.
3. Evidence supplied and evidence still needed.
4. Authority questions, including any approval route the sources do not state.
5. Conflicts or gaps between the request and the policy.
6. A decision box containing only: APPROVAL ROUTE NEEDED, EVIDENCE NEEDED, or READY FOR OWNER REVIEW.

Rules:
- Do not approve or reject the request.
- Do not invent a threshold, deadline, approver, exception, or legal conclusion.
- Keep the policy and request separate. Label each statement with its source.
- If a fact is absent, write NOT STATED.
- Quote the exact policy sentence for each rule.
- Treat the requested decision as a request, not as an approved outcome.
- Do not turn the cancellation screenshot into a claim about policy eligibility.

The prompt asks for structure and traceability. It does not ask the model to decide whether the exception should be granted.

The checkable artifact

Ask for a table in this shape:

Question	Source	What the files say	Owner check
Was the trip approved before booking?	Policy and request	NOT STATED	Confirm with manager
Does premium travel require Finance approval?	Policy	Yes	Confirm Finance route
Is there a rule for a same-day cancellation?	Policy	NOT STATED	Policy owner decision
Is reimbursement already approved?	Request	No, it is requested	Finance review

This table is an illustrative target, not a captured model output. Compare each row with the two files before anyone uses it.

Acceptance checks

The draft passes when:

The premium-ticket rule includes its exact policy sentence.
The cancellation is listed as supplied evidence, not as an automatic exception.
Missing manager approval stays NOT STATED.
Missing approval authority stays an open question.
The requested decision is not rewritten as an outcome.
A named policy owner still reviews the brief.

Troubleshooting

ChatGPT merges the files: ask it to label every sentence as POLICY, REQUEST, or NOT STATED, then compare the result with both originals.
It invents an approver: repeat the rule that absent authority must be written as NOT STATED.
It treats evidence as permission: separate evidence supplied from eligibility under the policy.
It cannot read a file: check the upload status and ask it to quote the file title before reviewing the result.

Limits and safer use

OpenAI documents file uploads for tasks such as comparing documents and applying a framework from one document to another. Plan limits and account settings vary, so check the current File Uploads FAQ. Use fictional or approved low-risk material until your organisation has reviewed data handling. This workflow does not replace Finance, compliance, legal, or management approval.

TRY: one low-risk internal request with a named policy owner.

SKIP: requests involving legal rights, safety, regulated decisions, or confidential data until the responsible team approves the process.

USE: the brief as a source-linked handoff for a human decision, not as the decision itself.

For practical AI-for-work tasks with an honest TRY / SKIP / USE decision, get The Weekly Verdict.