Chris Couch

Building AI that collects money so humans don't have to. | I2C · Payments · Unreasonably competitive

[Chris Couch @chriscouch]

Agent-native without context is just a wrapper. An AR agent is only as good as what it knows before it touches a customer: payment terms, dispute history, and who can approve a credit. The vendors that last will own that context. The buyers that win will hand it over.

[Chris Couch @chriscouch]

@elliot_garreffa For finance teams, the bigger news runs the other way. If the AR and billing tools you already pay for expose an MCP, an agent can work collections from inside Gemini. The vendor's job shifts to making every action reviewable before it touches a customer.

[Chris Couch @chriscouch]

@Techmeme Accounting firms carry a receivables problem of their own. They close the books for clients every month, while their own invoices often go out late and get paid later. The first test for software that runs firm operations is how fast the firm gets paid for work it already did.

[Chris Couch @chriscouch]

Agent checkout makes the card number the least interesting part of a payment.

With a token in place of the card, the agent holds nothing worth stealing. Fraud review moves to intent, meaning whether the customer asked for this purchase. Agents leave the trail to answer that.

[Mastercard @Mastercard]

Agents are now shopping. Agentic tokens are helping keep the card details to themselves. 
With Mastercard Agent Pay, Shop Pay is helping bring more secure, seamless payments to agentic commerce. Welcome to the next era of checkout. 💳

[Chris Couch @chriscouch]

The signed contract is becoming the billing system.

Invoicing has meant retyping terms both sides already signed. When an agent drafts the invoice from the contract, the terms live in one place. A price change or renewal reaches the next invoice with no rekeying.

[Mercury @mercury]

It's 2026. You shouldn't have to create your own invoices. Upload a contract and Mercury's agent, Command, finds the relevant details to build one for you. Just review and hit send.

[Chris Couch @chriscouch]

@sytaylor @Muse Finance teams will pick the short answer with a source every time. A controller can't book a number pulled from three paragraphs. Give one figure and the ledger entry behind it, and the reviewer approves it in seconds. That answer design wins the close.

[Chris Couch @chriscouch]

@Mars_DeFi When an agent pays a B2B invoice, the supplier pays the interchange, so the supplier must pick the rail. The invoice is the API. List the accepted rails on it, and the agent routes to the cheapest one. Cards keep the purchases where dispute rights earn the fee.

[Chris Couch @chriscouch]

@BoringBiz_ Growth is the headline, and cash is the scoreboard. Revenue grew 25% while free cash flow went negative, because capex and TPU inventory get paid for before customers pay. In this buildout, the company that shortens the path from backlog to billing to cash wins.

[Chris Couch @chriscouch]

Instant, free B2B transfers will turn net 30 into a pricing decision.

Terms grew up when moving money was slow and cost something. When a transfer is instant and free, the wait is just credit. Suppliers can price it openly, and buyers can trade the float for a better price.

[Stripe @stripe]

Introducing 'stripe pay', for instant business-to-business transfers. Find companies using their Stripe profile handle and send funds for free. Available in the Dashboard and CLI.

[Chris Couch @chriscouch]

Finance teams have been writing the answer key for reinforcement learning for years without knowing it.

Every payment a cash application clerk matched to an invoice is a graded example. The agent proposes a match, and the history says right or wrong. Reconciliation works the same way, with a bank line on one side and a ledger entry on the other. Most corporate work can't score an agent's output that cleanly. That is why the back office will produce the first strong RL results in the enterprise.

[Chris Couch @chriscouch]

Collections work is mostly dispute work.

A short payment usually traces back to a wrong price, a missing PO, or a delivery the customer never confirmed. Reminder emails do nothing for any of those.

An AR agent that resolves the dispute and then fixes the record that caused it changes the math. The next invoice goes out right the first time.

That feedback loop is where order-to-cash actually gets faster.

[Rohan Paul @rohanpaul_ai]

Tarek Alaruri (@realtarek), CEO and co-founder of Stuut, on raising $52.5M to automate how companies get paid with AI

Stuut builds an AI-powered order to cash platform that automates the receivables work between sending an invoice and having cash in the bank.

----
Full video on "SiliconANGLE theCUBE" YouTube channel, (link in comment)

[Chris Couch @chriscouch]

Profit you've booked but not collected is still a collections problem.

Profits up 14.4% while receivables rise 10%, accelerating for six straight months, means a growing share of those earnings sits with customers as unpaid bills. At the company level, DSO shows this first. Revenue and margin look healthy right up until cash conversion slips. Rising receivables are the early signal, and the oldest balances are where the work starts, before they turn into write-offs.

[Chris Couch @chriscouch]

Every model price cut is a quiet test of how your finance software charges you.

Per-seat vendors keep the savings, because the seat costs the same whether the model got cheaper or not. Per-invoice pricing passes the drop through, since the unit tracks the actual work.

Buyers will start asking which side of that line a vendor sits on. That question favors vendors who price on the work done.

[Chris Couch @chriscouch]

The chat window is the least interesting part of this. The real shift is that the ERP now gets judged on exceptions closed, not screens visited.

Oracle's Ledger app works the unreconciled entries and anomalies that eat the last days of every close. Once that queue clears on its own, close stops being a calendar event and becomes a running state.

Finance teams get that time back for the calls only a person can make.

[Chris Couch @chriscouch]

@shiffgil The two answer different questions. Tasks are what the customer buys, so price on those, since nobody budgets in tokens. Tokens are what you watch internally, to catch the task type that costs more to serve than it earns. What unit does Reevol bill on today?

[Chris Couch @chriscouch]

Your AI product's average margin is lying to you.

Illustrative case: a $400 flat plan with a 65% margin. Ten accounts cost $900/month each to serve. They're 5% of customers and 32% of cost.

Here's how to find them, and why the fix is rarely a price hike. https://t.co/dDlFZiNQiL

[Chris Couch @chriscouch]

Managing the entire AP process still means two jobs: preparing a payment and releasing it. An agent can own the first end to end, from capture and coding to matching and chasing the missing PO. Releasing cash is where a named person signs, and the agent's job is to make that signature fast and well informed. Where does Tipalti draw that line?

[Chris Couch @chriscouch]

The highest-value fraud check in AP is the one on a vendor bank-detail change. A lot of supplier-payment fraud starts with one routine-looking email asking to update where the money goes. An agent is well suited to this: hold the next payment, confirm through a contact already on file instead of the one in the email, and compare the request against the vendor's history. Do the firms stopping 90% of attempts treat a bank change as an event rather than a field edit?

[Chris Couch @chriscouch]

The scary part isn't the deleted invoice. It's that the agent's own account was the only record. Every action an agent takes in finance must land in a log you can replay, kept apart from what the model says it did. Drafting the vendor email and sending it must be two separate steps, and only the draft can happen without a person seeing it. In the stories you collected, did any team catch it from the log before the denial?

[Chris Couch @chriscouch]

Great map. Almost every one of these sits on the paying side. The receiving side is still empty. When an agent pays an invoice, the supplier still has to match that payment to the right invoice, and agent payments rarely carry clean remittance data. Cash application is where this breaks first. Is anyone in your 30 building for the supplier?

[Chris Couch @chriscouch]

Serving it cheaper comes before any of the three. Route easy requests to a smaller model, cache repeated calls, trim the context. After that, a fair-use cap at the top of normal usage with a clear path to buy more. That turns a cost problem into an upsell. Metered billing comes last, because finance teams budget in fixed amounts and a bill that swings goes to procurement. In your studio work, is the cost concentrated in a few heavy users or spread across everyone?