ISSUE 05 OF THE FOUNDING FIVE · 8 MIN

Small businesses lose money in the gap between doing the work and getting paid. 43% of US B2B invoices sold on credit were overdue last year, a collection agency keeps 25 to 50% of whatever it recovers, and processing a single invoice by hand costs $10 to $15. AgentCollect, Nibble, Clerked, and Fazeshift put agents in those gaps, and the one that charges 1% is the number to study.

Small businesses lose money in the gaps between doing the work and getting paid. Invoices sit 60 to 90 days because nobody chases them, vendors quote list price because nobody haggles, and bills get paid late, or twice, because one person does accounts payable on Fridays.

Each of those gaps is a job an agent can hold: chase, negotiate, pay. The customer already has the money problem and already knows what it costs, so the pricing that works ties the fee to the cash the agent actually moves.

The problem these builds attack is not that businesses don’t know they are owed money. It is that collecting it, negotiating it, and paying it out has always cost a human's salary or a third of the recovery. A collection agency charges a contingency fee of 10 to 50% of what it recovers, and in small claims, a small business actually has, under $3,000, the typical rate is 35% or more, which means a contractor owed $2,000 by a slow client gets $1,300 of it back if the agency succeeds. The alternative is to do the chasing in-house, and a bookkeeping or accounts payable clerk earns a median of $50,670 a year according to the Bureau of Labor Statistics, while a collections specialist earns about $72,000 with bonuses. Processing one invoice by hand costs $10 to $15 in staff time, according to the industry's own benchmarks, versus $2 to $3 when the work is automated; the manual version also carries a 2% error rate that occasionally pays the same bill twice.

The gap is so large that the fees survive. Atradius found that 43% of US B2B sales made on credit were overdue in 2025, which means nearly half of every dollar a small business invoices arrives late, and late money is money the business has to borrow against in the meantime. The software incumbent in that gap, Bill.com, charges $45 to $89 per user per month for the seat and then a fee on the payment: in its most recent fiscal year it earned $277.1 million in transaction fees on $86 billion of payment volume, about 0.32 cents on every dollar that moved, on top of $68.8 million in subscriptions. That is the two toll model in miniature, a seat and a transaction fee, and it is the most comfortable margin in back office software.

That is the terrain: a 35% contingency on small recoveries, a $50,000 salary to avoid it, a seat plus a transaction fee if you buy the software, and 43% of invoices late anyway. The builds below price against every one of those numbers. AgentCollect charges 1% of what it recovers, which is 1% against the agency's 35% and works because the agent's cost per call is close to zero. Clerked replaces the $50,670 clerk with a three-person company's software. Nibble negotiates the price down before the invoice exists, in a category where the software incumbents claim businesses overpay for vendors by an average of 34%. The new demand isn't for invoices that already get chased. It is the ones that do not, because the business could not afford a person or a third of the money to chase them.

One number from the other side of the ledger belongs in every issue on this subject. The FBI's Internet Crime Complaint Center recorded $3.05 billion in business email compromise losses in 2025 across 24,768 complaints, with 86% of the money leaving by wire, ACH, or crypto, and most of it started with an email that said a vendor's bank details had changed. An agent who pays vendors is a new reader of that email.

LISTED · COLLECTIONS

AgentCollect

Voice, email, and SMS agents that contact debtors under your brand, negotiate payment plans, and resolve disputes around the clock. The debtor pays you directly, same day, and AgentCollect charges a success fee of 1% of what it recovers with no setup fee. The site reports capacity of up to 85,000 recoveries a day and about 50% recovered within 20 days, with Microsoft, GitHub, DoorDash, and Plaid named as clients. Founder John Banner, YC S23.

LISTED · PROCUREMENT NEGOTIATION

Nibble

A negotiation agent for procurement. Your team sets a floor and a ceiling, and Nibble haggles with the supplier's rep by chat, one supplier at a time or across a whole tail spend campaign, with integrations into Coupa and SAP Ariba. The homepage counts 1.8 million negotiations handled. Founders Rosie Bailey and Jamie Ettedgui, London.

LISTED · ACCOUNTS PAYABLE

Clerked

24/7 AP clerks. The agent pulls invoices from email and portals, codes them to the ledger, matches them, flags the suspicious ones, routes them for approval, and pays the vendor once a human signs off, all managed from Slack, Teams, or iMessage. Three people, San Francisco, YC F25. Founders Evan Meyer and Sunjeet Chugh.

LISTED · ACCOUNTS RECEIVABLE

Fazeshift

The funded one. End to end accounts receivable as an agent, from invoicing through dunning to collections, wired into Salesforce, HubSpot, QuickBooks, NetSuite, and Stripe. A $4M seed in January 2025, then a round reported at $22M in 2026. Founders Caitlin Leksana and Timmy Galvin, YC S24.

Four products, and only one of them can send money out. AgentCollect's agent talks to the debtor, and the debtor pays the client's own account, so AgentCollect never holds a dollar and only has to prove the recovery happened before it bills its 1%. Nibble never touches money at all; it moves the price within bounds a human set. Fazeshift has write access to invoicing and CRM systems and sends the reminders, but the customer's own Stripe or bank receives the payment. Clerked pays vendors, and it puts a human approval in front of every payment. If you plan to build in this market, the vocabulary is the same one your customer's accountant uses, so it is worth learning before the first sales call.

Start with days sales outstanding, or DSO, the average number of days between sending an invoice and receiving payment, which for a small business with 43% of invoices overdue can run past 60. Every day of DSO is a day the business is lending its own cash to its customer, and the money it has to keep on hand to survive that lending is working capital. A collections agent is a working capital product before it is anything else: cut DSO from 60 days to 40 on $1 million of annual invoices and you have handed the business about $55,000 of cash it no longer has to borrow, and AgentCollect's 1% on the recovered amount is a rounding error against that.

Next is contribution margin on a recovery, which is what is left of the fee after the variable cost of earning it. A collection agency's 35% exists because a human collector makes perhaps 60 calls a day and gets paid $72,000 a year, so the cost of each recovery is high, and the fee has to cover it. An agent that makes 85,000 calls a day, with compute costs, has a variable cost per recovery near zero, which is why a 1% fee can carry a contribution margin the agency would envy at 35%. The same arithmetic explains Clerked: the cost of revenue on an invoice drops from the $10 to $15 of a clerk's time to the $2 to $3 of software, and the difference is the product's entire price.

The last term is the one Bill.com built a public company on: transaction revenue versus subscription revenue. Bill.com earns four dollars in transaction fees for every dollar it earns in seats, which means the money is in the movement, not the login. The builders here have read that annual report. AgentCollect and Fazeshift are priced on the movement, Nibble is priced on the savings, and none of them is priced on the seat, because a seat is the one thing an agent makes worthless.

Take the AgentCollect model to one vertical nobody has covered: trade contractors with 45-day invoices, dental practices chasing insurance follow-ups, or freight brokers collecting detention fees. One vertical means one set of excuses to learn, one set of payment portals, and one price sheet. Charge 1 to 3% of recovered cash, never a subscription, and route the money to the customer's own account so you never hold it, because the day you hold it, you are a money transmitter, and the compliance bill arrives before the revenue does.

The product is the first 30 days of recoveries you can show the next customer in the same vertical. If you price at the agency's 35% you win on cost by a mile, and if you price at 1 to 3% of the working capital you freed you win on margin and on the customer's loyalty, and the margin is the better business to be in.

invoice-x/invoice2data · Python, MIT, about 2.2k stars. A library and command line tool that pulls the issuer, amount, date, invoice number, and currency out of a PDF invoice using pluggable text extraction backends (pdftotext, pdfminer, and OCR through Tesseract, docTR, or PaddleOCR) and a library of YAML templates, with an AI fallback for the invoices the templates miss, and writes the result to CSV, JSON, or XML.

Why it is worth an hour: the templates directory is a ready made pattern library of real vendor invoices, Amazon, hosting companies, and dozens more, contributed by people who needed them parsed, and it is the extraction layer every AP or dunning agent in this issue has to build before it can do anything clever. The plugin system for line item and table extraction is the natural place to wire a language model in on top of a battle tested parser instead of asking the model to read the whole PDF and hoping. What to add before you let it feed a payment step is a check that the bank details on the invoice match the bank details on file, which is the whole of The Tell in one function.

The oldest fraud in B2B is an email that says "we have changed our bank details." It works on humans about once per company. It cost American businesses $3.05 billion in 2025, by the FBI's count, and an AP agent that pays vendors reads that email too, without the human's memory that the vendor has used the same account for six years.

Clerked's approval gate is the right instinct, and the stronger version is a written list of accounts the agent may pay, changeable only through a channel the agent cannot read, because an agent authorized to settle debts can settle them to the wrong account and the settlement will look correct on every screen it touched. FLINT Scout checks whether the paying agent was allowed to, for one cent.

WHAT COMES NEXT

We’re working on it 🙂