Process Agent

AI reads the customer case, fixed rules decide what runs. Paste a customer message: Claude works out what the customer wants, which policies apply and which actions would solve it. Then deterministic rules, not the AI, set the risk and decide what happens:

Also: if the AI is less than 75% sure, or the customer is upset, a person reviews the case. A safety rule adds a refund or cancellation that the message plainly asks for, even if the analysis missed it. The demo never executes anything; "approve" only shows what approval would unlock.

The demo uses Northwind Daily, the same made-up news subscription as ReplyDesk. Its policies and actions live in src/company.js, and the rules in src/decision.js.

Built by Frits (fizzl.eu) with Claude.

• Analysis: Claude with structured output (intent, language, sentiment, urgency, policies, proposed actions from a fixed list, confidence, open questions, a draft reply).
• Decision: plain code. The risk of each action is fixed in code; the model can only pick from the list. Every reason and step text comes from the rules, never from the model.
• Guard rails: the system prompt is fixed on the server, and the customer message is wrapped as data. The browser can't send its own prompt.
• Limits: 10 cases per visitor per hour and 300 per day. Messages are capped at 4,000 characters.

POST /api/process-case with { "message": "…" } returns { caseId, analysis, decision: { risk, confidence, humanRequired, reasons }, actionPlan: [{ action, status, reason }], draftReply, executed: false }.

The widget on fizzl.eu uses the same API; CORS allows the fizzl.eu sites only.

npm install
echo "ANTHROPIC_API_KEY=sk-ant-..." > .env   # never commit this
npm run dev                                  # http://localhost:3000
npm test                                     # no API key needed

Deploy on Render with the included render.yaml (Blueprint); set ANTHROPIC_API_KEY in Render only.