QuotePilot AI

QuotePilot AI is a quotation and customer-intake agent for service businesses. Talal Khawaja and Aqeela Urooj built it for Track 4, Autopilot Agent, of the Global AI Hackathon Series with Qwen Cloud on Devpost. The idea is an agent that moves quickly on its own but can't make the commitments a business has to stand behind.

Problem

Inbound requests are rarely quote-ready. They are missing a timeline, a budget or a key requirement. Someone has to work out what's being asked, look up the price list and availability, write a quote, get it approved, reply and log it in the CRM. Automating all of that end to end is tempting and risky. A model that invents a price, or an email that goes out before anyone has looked at it, does more damage than a slow reply.

How it works

The request goes to a server-only route that calls Qwen through its OpenAI-compatible API. Qwen returns schema-constrained analysis: the business objective, requested features, timeline, budget, missing information, whether clarification is needed, an allowed package recommendation and draft email content. Zod validates that output before any business tool touches it.

From there, deterministic local tools take over. The recommended package is checked against the requested service, the quote is calculated from a controlled catalogue, and availability comes from the same kind of code. If Qwen flagged missing information, the workflow stops and asks. Otherwise it prepares the quote and email and stops again at Human Approval. Approving runs the simulated email and CRM tools. Rejecting leaves them untouched and records that nothing was sent or created.

The app is Next.js and TypeScript. It runs as an Alibaba Cloud Function Compute web function in Singapore, using a Debian 12 custom runtime, bundled Node.js 20 and Next.js standalone output. The Qwen credentials live only in server-side function configuration.

Key decisions and trade-offs

• Probabilistic work is kept apart from commitments. Qwen understands and drafts. Code sets prices and availability. A human authorises anything that leaves the building.
• Visible state. Every stage and runtime event is logged, so the approval boundary shows up in the UI and in the logs.
• No fabricated fallback. If the provider fails or returns malformed output, the app reports it rather than inventing a result.
• Honest scope. Email, CRM, calendar and availability tools are simulations. The catalogue is small and lives in code, and there is no authentication, database or quote history yet.

I also wrote up the build on LinkedIn as "Building QuotePilot AI: A Human-Governed Autopilot Agent with Qwen Cloud".

What's next

The roadmap covers production email, CRM, calendar and e-signature integrations with scoped credentials. It also covers authenticated workspaces with persistent quote and approval history, catalogue administration (taxes, currencies, discounts, approval thresholds), immutable audit events, retrieval from company policies, evaluation suites, and multilingual intake.

Service requests arrive as incomplete emails and forms. Staff have to clarify scope, check prices and availability, write a quote, get approval, reply and update a CRM. It's repetitive and slow, and a model-generated price or a premature email can cost trust.

Qwen handles the language work: extracting requirements, spotting missing information, recommending an allow-listed package and drafting the email. Deterministic code calculates the quote and availability from a controlled catalogue. The workflow pauses for clarification when details are missing, and it always stops at a human approval checkpoint before the simulated email and CRM steps.

• Deployed as an Alibaba Cloud Function Compute web function in Singapore behind an HTTPS custom domain. The README records the Qwen workflow, approval checkpoint and simulated email and CRM completion as verified there.

• Reliable agents need narrow tools, validated contracts, explicit stopping conditions and observable state.
• Giving the model less authority made a stronger product: Qwen reasons flexibly while code protects the commitments that matter.