How it works
What everything here means
Arrearly makes decisions on their own and explains each one. This page is the key to reading those explanations - what the screens are for, and what every score, status, and number stands for.
Decides, not just sends
The agent chooses the next move for every overdue invoice.
Guardrails in code
A deterministic policy engine checks every decision the model makes.
Auditable by line
Every action is logged with its reasoning and confidence.
About
What Arrearly is
Arrearly is an autonomous receivables agent. It runs the chase for a small business's overdue invoices from start to finish: it decides which account to press today, writes the message, negotiates a payment date, stays quiet when a promise is still in date, and hands the hard ones to a person. Every decision is recorded with its reasoning and confidence, so nothing it does is a black box.
Most tools in this space automate sending a fixed sequence of reminders on a timer. Arrearly automates deciding, which is the part that actually needed a person. A model proposes each action; a deterministic policy engine then checks it against the escalation ladder and the operating limits, and steps back anything that breaks a rule. The model reasons, the rules keep it safe, and the audit trail keeps it honest.
Its goal is simple: recover the money without spending the relationship that produced it, and turn stranded receivables back into working capital.
Try it
See the agent decide, live
ABOUT THIS PREVIEW
This is a live preview, not a demo recording. When you start the conversation, the real agent reads the scenario, asks Gemini for the next action, and runs that proposal through the same deterministic policy engine as the rest of the product. Reply as the customer and watch the decision change: a promise to pay makes it hold, a dispute hands the case to a person, a plea to split the balance makes it offer a plan. Nothing is delivered here, no email, WhatsApp, or Telegram is sent. If Gemini is ever unreachable, the offline policy engine answers instead and the result is clearly flagged.
Try the agent yourself
Set up an overdue invoice, then play the customer. The real agent drafts every move with Gemini and scores your replies live. This is a preview of the decision, never a sent message.
Start the conversation:
Add data
Putting your own accounts in
The Add data screen lets a real business put its own book into Arrearly. Enter a customer first, then an invoice against them: amount, due date, and how much has been paid so far. The agent scores the new invoice and works it exactly like the seeded accounts, so you can watch it decide on your own overdue money, not just the demo data.
One thing to know at this stage: Arrearly is a new product, and its store keeps data in memory rather than in a database yet. That means anything you add lives for as long as the server is running and is cleared when it restarts or redeploys. It is enough to run one business's book end to end and to demonstrate the agent on real accounts. Durable storage in a database, so nothing is ever lost, is the next step on the roadmap.
Your data
Adding accounts, and what reset does
The Add data screen lets you put a real account into the book. Enter a customer first, then an invoice against them, and the agent scores and works it exactly like the seeded accounts. This is how a real business points Arrearly at its own receivables rather than the sample data.
Because Arrearly is still early, the book lives in memory rather than a database. That has one honest consequence worth knowing: anything you add lasts for the session and is cleared when the server restarts, and the Restore the sample data button on the Limits screen brings the book back to the starting set of thirty customers and thirty-six invoices. Nothing you add is permanent yet. Durable, per-business storage is the next step on the roadmap, and when it lands your entries will simply persist.
The loop
What happens on every run
It reads the whole file
For one invoice, the agent takes in the balance, how many days it is past due, the customer's payment history, the entire conversation so far, and every decision it has made on this account before. Nothing is judged in isolation.
It decides which rung to act on
It picks a single step on the escalation ladder - from staying quiet, through a gentle reminder, up to handing the account to a person. It may climb one rung at a time and never skips ahead.
The decision is checked against your limits
Before anything is sent, the choice is tested against the boundaries you set on the Limits screen. If it breaks one - chasing inside the grace window, or jumping a rung - it is stepped back, and the correction is written into the log.
It sends, or it asks you
If the agent is confident enough, the message goes out on its own and is logged as autonomous. If it is uncertain, or the account needs a human, it drafts the message and waits for you instead.
The screens
Where to find what
Today
The morning view. Money at risk, what came back in the last two weeks, and the work queue ranked by how much is at stake on each account.
Ledger
Every invoice the agent watches, oldest first. Filter it down to what is past due, what needs a person, or what has already been recovered.
Invoice
One account in full. Run the agent live and watch it decide, read the whole conversation, and see every past decision on that invoice.
Decisions
The complete audit trail across all accounts - the answer to the only question anyone asks about an autonomous system: why did it do that?
Limits
The boundaries the agent cannot argue past, and a live indicator of which model is currently answering.
The numbers
What every score means
Risk score
0 to 100How likely this balance is to go unpaid without intervention. Higher is worse.
How it is built
Age drives most of it. The customer's reliability and their average lateness adjust it, the size of the balance nudges it up, and a long relationship or a live promise to pay pulls it back down. It is built to be taken apart - a finance team should never have to trust a number it cannot explain.
Confidence
0.0 to 1.0How sure the agent is that a competent human collections manager would send exactly this message. It sets whether the agent sends on its own or waits for you.
How it is built
The model scores its own certainty. It is told to drop below 0.7 when the account is ambiguous, the amount is unusually large, or the customer's last message could be read more than one way. You set the exact line it must clear on the Limits screen.
Reliability score
0 to 100How dependably this customer has settled invoices with you in the past. Higher is better.
How it is built
A standing property of the customer, drawn from their history of settling with you. The agent leans on it heavily: the same days-late number means very different things for a customer who has never missed and one who always does.
The states
What each status tells you
Issued, but the due date has not arrived. The agent watches and stays quiet.
Past the due date and unpaid. This is where the agent does most of its work.
The customer has committed to a specific payment date. The agent goes quiet until it passes.
Some money has arrived and a balance remains - usually a payment plan in progress.
The customer contests the amount or the work. This never returns to automated chasing.
Handed to a human on the finance team, because the ladder is exhausted or the stakes are high.
Settled in full. The file is closed and the recovery is counted.
The ladder
The order the agent climbs in
The single most important rule in the whole system: a credible promise to pay outranks the ladder entirely. A customer who has committed to a date is never chased before it, because chasing someone who has already said yes is the fastest way to lose the yes.
The claim
What “without a human” actually means
The headline figure on the dashboard is the share of decisions the agent made and carried out with no person involved. It is the honest measure of how much work is genuinely being taken off your plate, so it is worth being precise about what counts.
Counts as autonomous: the agent chose the action, cleared your confidence threshold, passed every policy check, and the message went out exactly as written.
Does not count: anything drafted and held for your approval, anything routed to a person, and anything the offline engine decided while Gemini was unreachable.
A decision to stay quiet is still a real decision, and it is logged with full reasoning - but it is counted honestly as a hold, not dressed up as a recovery.