Inspiration

A business can have a full order book and healthy margins and still struggle to pay its suppliers this week. In project work, money goes out before the next customer payment comes in.

The idea came from building cashflow tools for a real project business. What mattered most wasn't the balance; it was the connections between supplier payments, project progress and customer invoices. Pay the wrong bill and a job stalls. Pay the right one and a milestone invoice gets released.

We wanted Reckon to answer one practical question: "What should we do next to keep the business moving?"

What it does

Reckon is an AI finance agent for project businesses. It works from the finance team's own spreadsheet, finds the cash gap, and closes it instead of just deciding who waits.

In the demo, a customer's email says a $9,000 milestone will be a week late, which leaves the business $3,220 short. Reckon works out that a $5,890 freight bill releases a $24,000 milestone on another job, so it ranks that payment first. Because forecast confidence has dropped, that payment is now over what Reckon may send alone, so it goes to the MD for approval with the exact terms bound to it. Reckon defers a software renewal that's in its grace period, asks a supplier who has split before to split again, and sends another customer a PayPal invoice offering 1.5% off for paying today. An unbudgeted $8,175 deposit is held for the MD. The gap closes for $200.

PayPal moves the money. Customer invoices go out through Invoicing, suppliers are paid by Payouts, and PayPal webhooks tell Reckon when a customer has paid or a payout has changed state, so it replans on its own.

How we built it

Reckon has three parts: reading, deciding and paying. Only the first uses the model.

Claude reads customer and supplier emails into one typed, checked fact each (a delay, a split agreed, a short payment explained) and drafts the messages Reckon sends. Code checks every number and link in those drafts against the books before anything goes out.

A policy engine does the deciding: what each payment unlocks, budget checks, the reserve floor, autonomy limits, and the cost of each lever. It's pure and deterministic, so every decision can be replayed from its stored inputs to the identical answer, and a tampered record fails its hash check.

The PayPal integration uses Invoicing v2, Payouts, the Reporting API and webhooks. Each payout uses Reckon's request id as its sender_batch_id, so PayPal itself refuses to pay the same decision twice. Every webhook delivery is verified with PayPal before Reckon acts on it, and duplicates are ignored. All of it was tested against the real PayPal sandbox: 9 of 9 calls passed.

Decisions are written back to a new tab in the finance team's workbook, so nobody has to change how they work. The app is Next.js and TypeScript on Vercel, with Supabase holding runs and webhook events. 96 tests cover the engine, the PayPal client and the webhooks.

Challenges we ran into

Working out what a payment unlocks. The oldest or most urgent invoice is often not the one that best protects the business, so Reckon ranks by what a payment releases on its project, not by due date.

Deciding how much authority the agent should have. We settled on limits that tighten when confidence drops: a payment Reckon would send alone on a normal morning goes to the MD once a receipt slips, and the approval is void if any term changes.

Payments are messy in practice. In our first live sandbox run the customer paid the invoice and PayPal sent the "invoice paid" webhook, but the money was still pending because the receiving account held no US dollars. Reckon didn't count that cash until PayPal marked it fully paid, which is what a finance team would want. Payouts to unregistered emails sit unclaimed, so Reckon tracks every payout state and only sends a replacement once PayPal confirms the first one came back.

Accomplishments that we're proud of

It works end to end on the real PayPal sandbox: real supplier payouts, a real invoice paid by a test customer, and webhooks updating the run as PayPal reports what happened.

The model never touches an amount, a limit or an approval. That split, Claude reads and writes words while code decides and PayPal moves the money, is what makes Reckon something a business owner could trust with real cash.

Replay. Any decision re-runs from its stored inputs to the same answer, and a tampered copy is caught.

What we learned

Cashflow decisions depend on context. A small supplier payment can matter far more than a large one if it releases a bigger customer invoice.

The model's job needs to be defined precisely. Language models are good at reading messages and explaining options; explicit rules are what give financial execution consistent boundaries.

Trust comes from visibility. Owners need to see what changed, why the plan changed, and which actions still need them.

What's next for Reckon

Treat PayPal fees and pending funds as first-class parts of the cash plan, since the sandbox showed us both matter.

Test Reckon on more realistic project scenarios and get it in front of small finance teams.

Longer term: accounting integrations, scenario comparison, and finer control over approval policies, so owners spend less time piecing together their cash position and more time acting on a clear plan.

Built With

Share this project:

Updates

Submission history