AI Agent for Customer Inquiries

Job Title: AI Agent Developer — Claude API + Acumatica ERP + meta message Whtap. Integration Job Description: Overview: I need an AI agent ('Markbot') built on the Claude API (Anthropic) that handles inbound customer inquiries via meta message Whtap Business, looks up inventory and pricing in our Acumatica cloud ERP in real time, and generates formal sales quotations — all automatically. Version 1 is human-in-the-loop: the agent drafts everything and I approve before it sends. You build the technical plumbing; I supply all business rules and domain content. What you are building: A meta message Whtap-to-Acumatica AI agent that: receives inbound meta message Whtap messages via the Meta Cloud API; classifies the inquiry type (availability, quote request, order status, escalation); queries Acumatica via its REST API for live inventory, pricing, and customer data; generates a formatted quotation and creates the quote record in Acumatica; and returns a natural language response to the customer via meta message Whtap — all with a human-approval checkpoint before any message is sent in v1. Important API constraint: Our Acumatica license allows 3 concurrent web service API connections and 500 requests per minute. The agent MUST be architected to behave like a single human user: one active session at a time, sequential API calls with confirmed responses before proceeding, session closed upon task completion, and a request queue for concurrent inbound messages. Accuracy and coexistence with human users takes absolute priority over throughput. Architecture: meta message Whtap Cloud API (Meta) for message in/out; Python backend on cloud hosting (Railway or AWS); Claude API (claude-sonnet-4-6) as the agent brain with tool-calling; Acumatica REST API (OAuth2, scoped MARKBOT service account) for ERP data; Acumatica sandbox available for all development and testing — zero access to live production data until final go-live. What I supply: Completed Business Requirements Document; Acumatica sandbox credentials (Phase 2); Meta Business Manager access; all business logic, pricing rules, and knowledge base content; prompt engineering for the agent persona; domain expertise on demand throughout the build. Deliverables: Working deployed v1 system passing all acceptance tests; full source code in a Git repo I own; documentation for maintaining and extending the knowledge base without a developer; handover session walking me through the architecture; clear upgrade path documented for v2 enhancements. 2. REQUIRED SKILLS & EXPERIENCE Non-negotiable (must demonstrate prior work): • Claude API or OpenAI API — tool-calling / function-calling agent loops • Acumatica REST API — OAuth2, endpoint mapping, reading and writing ERP records • meta message Whtap Business Cloud API (Meta) — webhook setup, message send/receive, templates • Python backend development — async handling, webhook servers, error recovery • API rate limit management — request queuing, session management, backoff patterns Strong advantage (will influence selection): • Prior experience integrating AI agents with ERP systems (Acumatica preferred) • Familiarity with Acumatica's contract-based REST API and scoped user roles • Experience building human-in-the-loop approval workflows • Understanding of meta message Whtap message template approval process (Meta) 3. ACUMATICA INTEGRATION DETAILS tenant — current cloud version Deployment Cloud-hosted (Acumatica SaaS) Sandbox Full snapshot of production tenant available for development — no live data risk Service account Dedicated 'MARKBOT' user with scoped role: read Inventory, Customers, Price Lists; create Sales Orders; cannot advance to Shipment or Invoice Email domain XXX@ggg Concurrent API limit 3 connections / 500 requests per minute — agent must operate within 1 concurrent session to preserve slots for human users Key endpoints needed StockItem (inventory/availability), Customer (lookup/create), SalesPrice (price lists), Opportunity or SalesOrder (quote creation) Developer Note — API Throttling Requirement • The agent must hold only ONE active Acumatica API session at a time • Every API call must await a confirmed response before the next call fires • Inbound meta message Whtap requests must queue — not process in parallel • The session must close immediately upon task completion • This is a hard architectural requirement, not optional optimization 4. meta message Whtap INTEGRATION DETAILS API type Meta meta message Whtap Business Cloud API (free, Meta-hosted) Number of divisions 4 divisions, each with a dedicated meta message Whtap Business number Division 4 status Number exists but is not yet under main Meta Business Account — migration to unified MBA is part of this project Routing Single webhook endpoint; agent routes by phone_number_id to correct division logic and Acumatica company/branch Message types in scope Text messages only for v1 (no voice, image, or document processing) Template messages Developer must assist with Meta template approval for outbound quote summaries Test environment Meta-provided test phone number for full development — live numbers not touched until go-live 5. AGENT CONVERSATION FLOWS (v1 SCOPE) Inquiry Type Agent Action Mode Availability inquiry Customer asks if item is in stock → agent queries Acumatica StockItem → returns qty + lead time Auto-reply Quote request Customer requests pricing → agent collects item(s) + qty → checks inventory + price list → creates quote in Acumatica → sends formatted summary to customer Human-in-loop v1 Order status Customer references an order → agent queries SalesOrder by customer/number → returns status + ETA Auto-reply New customer Unrecognized caller → collect contact info → create lead/customer record → alert sales team Escalate Complaint / dispute Negative sentiment detected → acknowledge, collect details → alert team immediately Escalate Unknown / out of scope Graceful fallback message → escalation triggered Escalate 6. PAYMENT MILESTONES Milestone Deliverable Payment M1 Project kickoff: architecture review, dev environment setup, Meta test number connected, Acumatica sandbox credentials received. Developer confirms understanding of API throttling requirements. 20% M2 meta message Whtap webhook live (test number); agent receives and classifies inbound messages; Acumatica sandbox connected; inventory availability query working end-to-end. 25% M3 Full quote flow working in sandbox: customer identification, item + qty collection, price list query, quote created in Acumatica, formatted summary returned via meta message Whtap test number. 25% M4 All 4 division meta message Whtap numbers routed correctly; Division 4 migrated to main Meta MBA; all 8 acceptance test scenarios passing in staging environment; human-approval checkpoint operational. 20% M5 Go-live cutover to production; monitoring confirmed active; handover session completed; documentation delivered; Git repo transferred to client ownership. 10% Budget Guidance • Expected total: $1500-$2000 depending on developer rate and complexity • Prefer fixed-price milestones over hourly — this protects both parties • Request itemized estimate broken down by milestone before awarding • Include a 30-day post-launch support period in your negotiation