What 150,000 invoices actually looks like
An aerial view doesn't do it justice. 150,000 unpaid invoices is not an abstraction. It is a daily flow.
It is a Fortune 500 treasury team uploading 8,000 acquisition-target invoices on a Friday afternoon, and 200 paid back by Monday morning without a single human collector touching the file.
It is a logistics enterprise dropping a 2,500-invoice batch every Monday at 06:00 PT for the past three weeks, with the AI calling, emailing, and SMS-ing each debtor in their local timezone, at the time they previously opened email or made payments.
It is a payment infrastructure company white-labeling our agent to run collections on behalf of their own SMB merchants, with the agent introducing itself under the merchant's brand and never breaking character.
It is 50 states, 7 languages, 24 hours a day. Every invoice in production has its own AI agent following its own playbook based on the debtor's history, the company's industry, the time of day, the device they read email on, and the hundred other signals we have learned matter.
How the AI actually does the work
People ask if "AI agent" means a chatbot. It does not. A real AI collection agent is a five-stage operational pipeline running per debtor, in parallel, every minute.
Stage 1 — Find the right contact. 30 to 40 percent of enterprise invoice files have an obsolete contact. The AI runs identity enrichment, public-record matching, and verification before it ever sends a message. If the email bounces, the agent does not give up. It searches LinkedIn, the company directory, and inbound payment data for the next reachable human.
Stage 2 — Choose the channel and timing. Email at the right hour for that debtor's timezone. SMS if email opens but does not click after 36 hours. Voice call if SMS goes unread for 72 hours. The order is decided per debtor, not per template.
Stage 3 — Hold the conversation. The voice agent picks up the phone with the debtor's full context: who they are, what they owe, what they paid before, what disputes they raised, what their company is doing this quarter. No human collector working 200 accounts can carry that depth.
Stage 4 — Resolve. Pay now. Set up an installment plan. Open a dispute. Confirm the invoice already paid the prior week. Each path closes the loop with the right document, the right ledger entry, and the right notification to the client's accounting system.
Stage 5 — Learn. Every conversation, every payment, every dispute is feedback. The next invoice for the same debtor benefits from everything the system learned on the last one.
The five stages used to take a 200-person collection floor. Today, two humans oversee them across 150,000 cases.
The 99.7% positive interaction rate
Every conversation our AI has with a debtor is scored on three dimensions: did the debtor stay engaged, did the debtor end the call without anger, and did the debtor express any negative sentiment.
99.7 percent of conversations come back clean across all three.
That number is not a marketing rounding. It is the ratio that lets a CFO sleep at night when their company brand is on the line. Aggressive collection agencies routinely score in the 70 to 85 percent range, and the bottom 15 percent is what generates regulatory complaints and lawsuits.
Zero legal complaints in 150,000 cases. Not a single bar association inquiry, FDCPA filing, or attorney letter. That is the operational floor on which Microsoft, Dell, Plaid, Checkr and other enterprise clients agreed to put their brand on the line.
The 0.3% — when humans step in
The two humans are not collectors. They are exception handlers.
They look at the conversation log when an unusual signal appears: a debtor mentioning bankruptcy filing, a contact disputing the underlying contract, a fraud red flag in payment data, a mass duplication across a client's account.
Every escalation is closed within hours, not days. And every escalation feeds back into the pipeline so the AI handles the next case of that pattern without a human in the loop.
Why collections has been broken for 50 years
The collection industry was built on a wage-labor economics that AI just reset.
A junior collector at a US agency carries 150 to 250 accounts. Their paycheck is variable: a percentage of dollars collected. So they gravitate to the largest balances. Small balances and complex balances stay on the desk and rot. Industry-wide, less than 20 percent of placed accounts ever get recovered, even at the agencies that look the best on paper.
The CFO of a large enterprise has known this for decades. The economics did not allow another option.
AgentCollect changes the unit economics. The agent is indifferent to balance size. A $200 invoice gets the same playbook as a $200,000 one. The cost of running an extra conversation is near-zero. The marginal return on every additional contact attempt is positive.
That is what 500x leverage actually means. It is not that humans got faster. It is that the entire economic constraint that defined the industry no longer applies.
The shape of what comes next
150,000 is a number we did not pick. It crossed today. The number itself does not matter. The shape it points to does.
The shape is every B2B receivable in the United States running through an AI agent the moment it ages past due. Every late invoice that today sits as a write-off line in a CFO's quarterly. Every Sunday a small business owner used to spend chasing customers who would not pick up.
That is what we are building toward. The two-humans-overseeing pattern is not a phase. It is the architecture.
Where this actually started
When I was a student in Paris, a hotel didn't pay me for a piano gig. I never asked for the money back. I was too embarrassed to remind them. That feeling never left.
My grandfather René, gone a few weeks ago at 102, ran a furniture store back when banks didn't extend credit. He gave the credit himself, and walked door-to-door every Sunday to collect payments. What stayed with me: every Sunday post-war, family time burnt chasing money that should have been paid.
He would have been proud.
