AI Progress Update

We built an agentic system: a small set of core agents, each backed by a growing set of specialized skills, working together inside a single conversation. When a member messages us, the system figures out what they actually need and routes to the right agent — grounded in CRWD's real rules and real data, not generic guesses. It's built to know its limits, too: frustration, disputes, or anything sensitive gets handed straight to a human.

Before writing any agent behavior, we analyzed over 400,000 real member messages to find what members actually ask, where they get stuck, and what a good response looks like. Every agent and skill below was curated from that real usage, not designed in a vacuum.

Reviews every piece of proof a member submits for a completed gig — the receipt, the screenshot, the content — and decides whether it's valid. Its underlying model is trained on millions of receipts, so it recognizes real receipt formats and purchase patterns rather than checking surface-level details — the goal is a foolproof, human-like audit on every submission.

• Receipt Review — checks submitted proof of purchase against what the gig actually requires (right product, right store, readable, complete).
• Duplicate Detection — flags receipts that have already been submitted before, so the same purchase can't be claimed twice.
• Fact-Checking Against the Gig — pulls the real details of the specific gig/campaign a receipt was submitted for and verifies the receipt actually matches it (product, store, purchase window) instead of approving on appearance alone.
• Proof Completeness Check — catches missing pieces before they become a rejection (e.g. a required second receipt, a missing content link).
• Submission Status Lookup — tells a member whether their submission passed, failed, or is still pending review.

Runs underneath the Auditor to protect payouts — scoring submissions for legitimacy so invalid or fraudulent claims get caught before money moves.

• Legitimacy Scoring — is being built out to score each submission for trust/fraud signals, flagging anything suspicious for review.
• Pattern Detection — watches for repeat or coordinated abuse across submissions rather than judging each one in isolation.
• Escalation on Risk — anything flagged goes to a human for a final call rather than being auto-approved or auto-rejected.

The member-facing front line — the agent members actually talk to for everything else, start to finish.

• Gig Discovery — finds gigs for a member, explains payout/deadline/requirements, and locates the exact nearby store (address, phone, hours) via live web search.
• Gig Execution — walks a member through completing a gig correctly — what to buy, what content to create, what proof to submit.
• Payment Status — explains and tracks a member's real, live payout status end-to-end.
• App Navigation — guides members around the app — finding gigs, tracking progress, getting to the right screen.
• Troubleshooting — resolves common technical issues with quick, proven fixes.
• Reminders & Follow-ups — proactively schedules nudges so members don't miss a deadline or drop off, and checks back in to help them find their next gig.
• Conversation Labeling & Reporting — automatically categorizes every conversation, powering internal reporting on what members need.
• Escalation / Handoff — recognizes frustration or anything sensitive and hands off to a human cleanly.
• Live Web Search — answers real-world questions it can't know on its own — store hours, phone numbers, product details — instead of guessing.

The system carries a running understanding of each member across every conversation, not just within one chat. It builds a picture over time: gig history, past questions, what they get stuck on, how they like to be helped. In practice that means no repeated questions, faster and sharper answers, personalized coaching, and an experience that keeps improving the more a member uses it — this isn't a stored log of messages, it's a model of the person that keeps learning.

• Scales instantly — with no ramp-up time, and stays consistent every time.
• Compounding — every agent, skill, and member profile gets sharper with more real usage.
• Retention & efficiency — members feel understood, and support cost per member drops as re-explaining goes away.
• A durable data asset — the longer a member is active, the more valuable their relationship history becomes.