PayCrew

Inspiration

A familiar vendor sends an invoice. Same logo, same amount, same tone. One detail is different: "We've switched banks, please pay this new account."

That one line is business email compromise, and it cost US businesses \$3.046 billion in reported losses in 2025. The defense is embarrassingly simple: call the vendor on the number you already have and ask. Almost nobody does, because the small-business owner is already doing five jobs.

So we asked: what if an AI crew didn't just raise an alert, but actually picked up the phone?

What it does

PayCrew is an autonomous finance crew that reads every invoice, pays the clean ones, and calls the vendor before a single dollar goes to a suspicious account.

Invoices land and a six-stage agent workflow kicks in instantly: extract details, match purchase orders, catch duplicates, spot changed bank details, score the risk, decide the next move.
Clean invoices get paid through Stripe, live, in seconds.
Suspicious invoices get a real phone call, one after another, to the vendor's number on file. The voice agent quotes the exact invoice number and amount and asks one question: did you change your bank details?
Vendor denies it? Payment blocked, owner alerted in Slack.
Vendor confirms it? The owner gets an Approve button in Slack. One tap, and Stripe pays.

Our demo runs five invoices end to end: three routine ones paid instantly, one impersonation attempt stopped cold by a live call, and one legitimate bank change confirmed by phone and approved in Slack. Real calls. Real Slack. Real Stripe.

How we built it

Brainbase runs the agent crew and places the live verification calls, carrying each invoice's details into the conversation.
Stripe handles payments through a connected Link wallet, paying only to trusted vendor checkout links, never to bank details pulled from an invoice.
Claude reads each invoice and powers the analysis, alongside database lookups and explicit rules the model can never override.
Slack is the owner's control room: fraud alerts and one-tap approvals.
Frontend: React, Vite, and Framer Motion. A live invoice board plus an interactive workflow diagram, so you can watch every invoice move through every stage.
Backend: Python and FastAPI, SQLite for invoices, findings, calls, and payment state, and WebSockets pushing every change to the board in real time.
Cloudflare powers our hosting and public tunnel so Stripe, Slack, and the calling agent can reach us from anywhere.

We also built a safe local mode (local extraction, mocked calls, payments disabled) and an offline test suite, so we could iterate at full speed without phoning real vendors or moving money by accident.

Challenges we ran into

Handling uncertainty without cutting corners. A changed bank account is suspicious, not proof of fraud. A vendor's "yes" is evidence, not permission, so the owner still decides. Starting a payment isn't the same as a payment succeeding.

We designed explicit states for every step, including failures and unclear call outcomes. Then we got the board, backend, analysis, calling, and payments to agree on one invoice status, all in under eight hours.

Accomplishments that we're proud of

A full pipeline, from invoice to phone call to Slack to Stripe, working live in one day.
Two invoices with the same red flag reach different outcomes based on what the real vendor says on the phone. That's judgment, not a keyword filter.
Safety baked in: the agent only calls numbers on file, never reads out full account numbers, and never accepts new bank details over the phone.

What we learned

Useful AI automation needs boundaries as much as intelligence. Document understanding is powerful, but trusted records, explicit rules, a human's final say, and plain-English explanations are what make people trust an agent with their money.

What's next for PayCrew

Owner authentication and access controls, stronger independent payment confirmation, recovery when workflows get interrupted, and one-click onboarding from existing accounting tools so vendor records and purchase orders flow in automatically. Then the other side of cash flow: the same crew chasing the money owed to the business.

PayCrew pays your invoices automatically. When something smells like fraud, it picks up the phone.