Avenger

Pitch. Spinning up a new company in a procurement platform takes weeks. You upload separate files for purchase orders, suppliers, categories, units, templates, rules — and half of them have broken data nobody catches until something blows up. We built the Onboarding Agent, and it gets a company live in hours instead of weeks. It opens by asking the admin one simple question — what do you want to do today? — and offers four use cases: manual RFQ, RFP, Intake, or integrated PR. Then comes the magic trick: instead of asking for six different files, it asks for one — the PO file. From that single upload, the agent builds a column mapping for human review, then dispatches a team of specialist sub-agents — PO, supplier, category, template, rule — each one cleaning its own slice of the data. The output is sharp: on one side, golden data ready to push to the company; on the other, a plain-English list of errors the admin can fix. A chat agent sits beside every error to explain what's wrong and how to fix it. Restart, and you're live. The problem. Standing up a new company on a procurement platform is a multi-week ordeal. Admins juggle separate uploads for POs, suppliers, categories, units of measure, templates, and rules — and customer files are routinely incomplete, mislabeled, or just plain broken. The errors surface late, blocking go-live and burning implementation hours. The solution. A conversational Onboarding Agent that takes the admin from empty workspace to ready in hours. One conversation, one file, one clear go-live signal. How it works. 1. Use case picker. The agent opens by proposing four onboarding paths — manual RFQ, RFP, Intake, or integrated PR — so the admin chooses an outcome, not a configuration checklist. 2. One file, not six. Instead of separate uploads for every entity, the agent ingests a single PO file and derives everything else from it. 3. Column mapping with human in the loop. The agent maps the customer's columns to standard schema columns and presents the mapping for human review and confirmation — the only manual step that matters. 4. Specialist sub-agent fan-out. Once mapping is confirmed, the orchestrator dispatches a team of specialist agents in parallel: a PO specialist, supplier specialist, category specialist, template specialist, and rule specialist. Each one owns its data domain end-to-end — extract, normalize, validate, enrich. 5. Two clean outputs per dataset. Every specialist produces (a) golden data — clean, schema-conformant rows ready to push to the company, and (b) a structured error list with human-readable explanations. 6. Chat agent for repair. A grounded chat agent sits next to the errors. The admin asks "why is this row failing?" and gets an answer pinned to the actual data — not a generic LLM guess. 7. Fix and restart. The admin resolves errors, re-runs, and the loop closes. Why it wins. Weeks become hours. Six fragile uploads become one. A black-box import becomes a transparent assembly line of specialists, each one accountable for its own slice. And when something does break, the admin gets an error they can read, a chat agent that explains it, and a restart button that just works. Resources. Main https://fairmarkit.atlassian.net/wiki/spaces/FUSE/pages/4647649281/Customer+Onboarding+Agent Live App https://customer-onboarding-afa021.pages.fmdev.io/ GIT https://gitlab.fmdev.io/integrations/customer-onboarding https://gitlab.fmdev.io/fairmarkit/frontend/-/tree/hack27-onboarding_avenger