Start with a real request
Bring the request and the relevant existing record into the same task. Keep the source visible so a reviewer can check the context.
PRODUCT / GOONGEOL
Make the next step
clear enough to act on.
A workflow layer between existing business records, AI and the people responsible for the result. Our first focus is handling changes to operational requests.
Product stage — The review-and-export example below works in your browser. Live Claude inference and connectors to business systems are planned next development steps; they are not active in this example.
THE WORKFLOW
One request.
All the context.
Designed for work where a plausible answer is not enough: someone must understand what changes and why.
Prepare the next step
The planned Claude integration will identify intent, compare the supplied context, draft a reply and return a structured proposal. Missing evidence should become a question, not an invented answer.
Review with the person responsible
Show the original and proposed values together. Let the reviewer edit the result and resolve open questions before approving the draft.
Return to the existing process
The browser example exports a reviewed draft. Live delivery into an existing system will require a configured connector, permission checks and an agreed review boundary.
EXPLORE A REVIEW
Change the date.
Keep the rest intact.
Try reviewing a delivery-date change. Edit the proposed date and reply, confirm your checks, then download the reviewed draft.
Fictional sample data · Local browser logic · No live AI call, message sending or business record update
SOURCE / ORDER 1042
A change comes in.
“Please move delivery for order 1042 to 16 October 2026. Keep the quantity and address unchanged.”
- Current delivery
- 2026-10-12
- Quantity
- 24
Existing destination: sample location A. Capacity for the requested date is not provided.
Open question: can delivery be arranged for the requested date?
PROPOSED WORK / FOR REVIEW
Before2026-10-12
Proposed2026-10-16
Review is required before export.
CLAUDE IN THE PRODUCT
Reason over the request.
Leave a traceable proposal.
| Step | Planned Claude role | GOONGEOL responsibility |
|---|---|---|
| Understand | Interpret Korean or English requests and extract intent, record identifiers and requested changes. | Supply only selected, authorized context and validate the output format. |
| Prepare | Draft a response and structured action proposal; identify missing or conflicting information. | Keep source records, proposed changes and unanswered questions visible together. |
| Review & return | Revise a proposal when the reviewer supplies clarification. | Require human approval; check destinations and permissions before any future connector action. |
What comes next
Connect the Claude API, evaluate Korean and English requests against expected outputs, and validate one integration from request to reviewed result. We will measure field accuracy, unsupported claims, review edits, latency and cost per task.
INITIAL WORKFLOW EVALUATION
Before AI changes
a business record.
We exercised 12 fictional Korean and English cases in the Claude web interface. The report records the initial problems and a prompt revision on the same cases. API integration and independent validation are still ahead.
Keep changes in scope.
Check the record identifier, requested field and allowed action. Quantity and destination must not silently change with a delivery-date request.
Make uncertainty visible.
Missing identifiers, conflicting records and unconfirmed availability must lead to a clear question or a review hold, not a made-up fact.
Keep approval meaningful.
Treat instructions inside a customer request as untrusted content. No case permits a system update without an authorized person reviewing the proposed action.
What the set covers
A normal date change, missing order identifiers, conflicting records, unknown delivery availability, changes outside the allowed scope, and instructions attempting to override the workflow. All records and requests are fictional. The report separates observed replies, expected behavior and remaining limitations.
Download the 12-case evaluation set (JSON) Read the Claude evaluation reportAfter integration, we plan to record field accuracy, unsupported statements, reviewer edits, latency and API cost per task. No accuracy scores or production-readiness claims are implied by this plan.
WORKFLOW DEVELOPMENT KIT
Inspect the boundary.
Run the checks.
Our local validator checks a proposed delivery-date change against an authorized order and supplied records. It blocks changes to protected fields and requires a new review whenever its inputs are replaced.
Runnable source, sample input and adversarial tests
The download includes the Node.js module, fictional input, evaluation cases, tests and a report describing what passed and what remains untested. These are deterministic software checks. Claude was checked separately in its web interface; API integration and business-system connection are the next development steps.
The module exports reviewed proposals only. It does not authenticate people, confirm that supplied records are current, interpret the original request, send messages or update a business system. An operating integration must provide those controls.