Your Payables Have a Yield

On August 21, 2026, PYMNTS ran a story about one Coupa customer. The customer's Navi Payment Batch Creation Agent had spent five weeks in production. It ran fourteen payment batches containing 2,395 payments worth $20.1 million, and Coupa says it used about $27 in AI credits to do it. The customer estimated it saved around $2,000 of accounts payable staff time in one week.

Everyone who repeated the story quoted the ratio. Twenty-seven dollars against two thousand. It is a lovely ratio, and it measures the wrong thing.

The $2,000 is a clerk's week. The $20.1 million is a stream of cash leaving a company on dates that somebody chose. For forty years the back office has been run, budgeted and judged as a place where labor turns paper into ledger entries. The agents arriving this autumn make that labor almost free. Once the labor is almost free, the only thing left worth optimizing in AI in back-office operations is the thing the labor used to hide, which is time. That means when cash goes out, when it comes in, and what each of those days is worth.

This essay argues that your payables and receivables carry a yield, that your back office has been leaving it on the table because humans were the scheduler, and that the companies who notice first will book the gains in treasury, far from the headcount report.

The ratio that fooled everyone

The labor case for back-office automation is real and small. An accounts payable clerk in the United States costs perhaps $60,000 to $80,000 a year fully loaded. Automate ten of them and you have saved something under a million dollars. For a company with a billion dollars of revenue, that is a rounding error in the annual plan, and it arrives slowly, because people are reassigned rather than removed and the savings leak into other budgets.

This explains a figure Sidetrade quoted when it launched its SAFE agent framework on September 10, 2026. Citing McKinsey, the company said 88 percent of organizations have deployed AI while 81 percent have seen no significant financial impact. There are many explanations for that gap. One of them is simple accounting. If you measure an AP agent by the clerk hours it removes, you are measuring the smallest number it touches.

The Coupa release itself hints at the bigger one. Among its 250-plus updates, reported by ITBrief on August 24, 2026, is an agent built for "identifying early payment discount opportunities." That line got less attention than the $27, and it is where the money lives.

Why the back office runs in batches

Every rhythm in a finance department was set by the speed of people. The weekly payment run exists because a team can prepare, check and release payments about once a week without drowning. The month-end close exists because reconciling everything every day would need three times the staff. Dunning letters go out on fixed cycles, 30 days, then 45, then 60, because a collections team works a list from the top.

None of these rhythms was designed around the value of money. They were designed around the capacity of a room.

The payment run is a fossil

Take the weekly batch. A company with standard terms receives invoices all week and pays them in a lump on Thursday. Some invoices get paid days before they are due, which gives away float for nothing. Some miss an early-payment window by a day, which forfeits a discount. Some get paid on a date that makes no sense against the company's cash position, because the batch date was fixed and the cash position was not.

A human AP team knows all of this and cannot fix it, because fixing it means making a separate timing decision on each of thousands of invoices, every day. That is exactly the kind of work that costs $27 in credits now.

When preparation becomes nearly free, the batch loses its reason to exist. You can decide each invoice on its own day, against its own terms, against the company's actual cost of capital that morning. The Coupa agent still bundles payments into batches, because the banks and the approvers still expect batches. The economics underneath have already moved.

The yield on a payable

Here is some plain arithmetic, offered as illustration rather than as anyone's reported result.

The classic early-payment term is "2/10 net 30." Pay within ten days and take 2 percent off, or pay the full amount on day thirty. Taking the discount means paying twenty days early in exchange for 2 percent. Annualized, that works out to roughly 37 percent. Very few treasurers can find a 37 percent return anywhere else. Most companies with the cash to take these discounts should take every one they can reach.

They do not, and the reason is mundane. The discount window is ten days. The invoice has to be received, coded, matched to a purchase order and a receipt, routed for approval, and scheduled into a payment run, all inside that window. Human processes routinely take longer than ten days. So the discount expires, and the company pays full price on day thirty, or on day twenty-two because that is when the batch ran.

Now invert it. When an invoice carries no discount, the right move is usually to pay as late as the terms allow, keeping the cash working for the company. Paying on day twenty-two instead of day thirty costs eight days of float on that invoice. On a single invoice that is nothing. On $20.1 million in five weeks, it adds up.

Put the two errors together and you see the shape of it. A human-scheduled back office pays too late where paying early is lucrative and too early where paying late is free. The losses from those two errors never show up as a line item. They sit in the spread between what the company paid and what it could have paid, and nobody's bonus depends on that spread.

Run a hypothetical against the Coupa customer's volume. Suppose a quarter of that $20.1 million came with 2/10 terms. Capturing 2 percent on $5 million is $100,000. The labor saving in the same stretch, at $2,000 a week, is about $10,000. The discount alone could be ten times the headcount saving, and the software already has an agent for finding it. None of those numbers are Coupa's, and your mix of terms will differ. The order of magnitude is the point.

Receivables are the other half

Payables are the half of working capital you control directly. Receivables are the half where you have to persuade someone else, and that has always made them a labor problem. Collections teams call, email, reconcile disputed invoices, match partial payments to open items, and escalate. The work scales with headcount, so companies ration it. Big accounts get attention. The long tail waits for the 60-day letter.

Sidetrade is the clearest example this month of a vendor that has stopped pretending this is a labor story. Its SAFE framework, launched on September 10, 2026, runs under the Aimie Cash Collection Agent, which the company says is already in production at Securitas, Accor and Sodexo. CEO Olivier Novasque framed the enterprise worry as a question: "How do we put AI to work on cash flow generation without handing control of our customer data to someone else?" The words that matter there are cash flow generation. He is describing a yield engine.

On September 22, Sidetrade reported first-half 2026 results. Revenue was €34.8 million, up 21 percent at constant currency. AI-native products made up 33 percent of second-quarter bookings, and 80 AI agents were placed on firm order, 26 of them Aimie agents. Those are small numbers in absolute terms. They matter because they show buyers paying for a collections agent as a line item, and because of one detail in the SAFE launch that most finance leaders will skip past.

The vendor guaranteed the compute price

Sidetrade says SAFE runs on its own private data centers, on GPUs it bought rather than rented, with fine-tuned open-weight models. It also says it will guarantee compute pricing to customers through multi-year contracts, shielding them from swings in per-token pricing.

Read that as a treasury person would. A collections agent is an asset that produces a return in days of sales outstanding. Its running cost is compute. Sidetrade is offering to fix the cost side of that trade for years, so the customer's return becomes a spread between a fixed input and a variable output. That is how you would structure a financial product. It is a strange thing for an order-to-cash software vendor to do, and it tells you how the vendor sees what it is selling.

The arithmetic on receivables is as blunt as the arithmetic on payables. A company with $1 billion in annual revenue carries about $2.74 million in receivables for every day of DSO. Take five days out and you have released roughly $13.7 million in cash, once, permanently, without borrowing a cent. At a 7 percent cost of capital, that is close to a million dollars a year in carrying cost. That single result is worth more than automating a sizable collections team, and the collections team was the constraint on getting it.

Two trillion dollars of badly timed cash

On August 5, 2026, CFO.com reported The Hackett Group's latest North American working capital survey. Hackett put the working capital opportunity at North America's largest companies at a record $1.94 trillion, with the performance gap up 12 percent even as revenue and profitability improved.

That number deserves a moment. Companies got more profitable and worse at timing their cash in the same stretch. The gap widened because the processes that set the timing are the processes nobody funds. They live in the back office, they are staffed to a cost target, and their managers are rewarded for throughput and accuracy. Nobody in that chain is paid to care whether an invoice went out on day twenty-two or day thirty.

Hackett has published versions of this number for years, and every year it is enormous and every year it barely moves. That is the signature of a problem that was structurally unsolvable at human prices. You could see the trapped cash. Freeing it meant making millions of small timing decisions correctly, and small timing decisions were the most expensive kind of work in the building, because each one needed a person.

Payment and collections agents change the price of a small timing decision. That is the whole shift, and it is bigger than anything in the labor case.

Where AI in back-office operations should report

If the return from back-office agents shows up in days and basis points, then the org chart has to change to match. Most companies run AP and AR under the controller, whose job is accuracy, compliance and a clean close. Treasury sits elsewhere, watching cash positions and funding. The two meet at the weekly cash forecast and then return to their corners.

That split made sense when the back office could only execute and treasury could only plan. An agent that decides the payment date for every invoice is doing treasury work at clerk speed. The policy it follows (when to take a discount, when to hold cash, which receivables to chase and how hard) is a treasury policy whether or not anyone admits it. Right now, at most companies that deploy these agents, that policy will be set by default settings chosen by a software vendor and approved by an AP manager who has never been asked to think about the cost of capital.

Move the policy to the treasurer. Leave the process with the controller. That is the org design change, and it costs nothing but a meeting and some pride.

Change the scoreboard

The metrics have to follow. Stop reporting the back-office program in hours saved and invoices per FTE. Report it in four numbers:

Discount capture rate, as the share of available early-payment discounts actually taken.
Average days paid against terms, split between discounted and non-discounted invoices.
DSO, by customer segment, with the long tail reported separately.
Cash released, in dollars, since the agent went live.

The first number will embarrass most finance teams. That is useful. An embarrassing baseline is how you get a budget.

The last click becomes a rule

Coupa's payment agent does everything up to the release and then stops. A human, or a separate system, has to approve before money moves. Pablo Fourez, Mastercard's chief digital officer, told PYMNTS: "As autonomy increases, trust cannot be implied. It must be proven."

He is right, and the implication runs further than most readers take it. When the agent prepares 2,395 payments, the person clicking release cannot meaningfully review 2,395 timing choices. What that person can review is the policy that produced them. The approval that matters moves upstream, from the batch to the rulebook. The rulebook needs an owner who understands what a 37 percent annualized discount is and why paying on day twenty-two is a small, silent loss. That owner is a treasury person, and the rulebook is where the yield gets decided.

The counterparty has agents too

There is an obvious objection. If your AP agent is optimizing when to pay, your suppliers' AR agents are optimizing when to collect. Sidetrade is selling Aimie to the same kind of companies Coupa is selling payment agents to. Put the two in the same trade and the spread you were harvesting starts to close.

That is correct, and it strengthens the argument for moving early. Today most counterparties still run human-scheduled back offices. An agent-scheduled company trading against human-scheduled suppliers and customers wins nearly every timing decision, because it is the only side making them deliberately. That advantage is temporary and real. It lasts as long as the other side's back office still runs on a Thursday batch and a 45-day letter.

Then comes a second phase, which is already visible in the Coupa release. Coupa's Navi Connect exposes more than 30 procurement and finance tools through the Model Context Protocol, so outside systems such as Microsoft Copilot can reach into spend data. Once both sides of a trade expose their terms and cash positions to software, payment terms stop being a clause negotiated once a year and start being a price that can move per invoice. A supplier that needs cash this week can offer a deeper discount for payment today. A buyer flush with cash can take it. Dynamic discounting has existed for years as a product. It becomes the default when both sides have agents that can accept or decline in seconds.

In that world, the edge goes to whoever knows their own cost of capital most precisely and can express it as policy fastest. A back office run as a cost center has no view on cost of capital at all. It will be the side that always accepts the default.

What the clerks are for now

Nothing here argues for firing your AP team next quarter. The labor saving is small, as we have seen, and the people who know why a supplier's invoices always arrive without PO numbers are the people who can teach an agent the exceptions. The failure mode is the one the McKinsey figure hints at. A company deploys agents, cuts heads, reports the saving, and never touches the timing policy. It books $100,000 of labor and leaves ten times that in discounts and float.

The better move is to redeploy the most experienced people onto the exceptions and the policy. Every rejected match, every disputed invoice, every supplier who changed bank details on a Friday afternoon is a place where the agent needs a human judgment and where fraud tends to hide. Emburse, launching Emburse AP on September 30, 2026, is pitching it at "finance teams with limited headcount" and says its AI will flag anomalies and route decisions. That is the right division of labor. The software handles volume. The people handle the anomalies and the rules.

The strange consequence is that your back office staff become more valuable per head as the head count falls, because each remaining person governs far more cash. Pay them accordingly, or the best of them will leave for the vendors.

Architect the yield, do not buy the tool

Every vendor in this story will sell you an agent. Coupa has hundreds of customers running them. Sidetrade will fix your compute price for years. Emburse ships next week. Buying any of them is easy, and buying any of them alone will get you the $2,000 week and very little else.

The yield comes from architecture. You have to decide who owns the timing policy, write that policy in terms of your actual cost of capital, rebuild the scoreboard so the back office is measured in cash released, and set up the controls so that a human approves the rulebook rather than rubber-stamping batches no one can read. You have to plan for the day your counterparties run agents too, when payment terms turn into live prices and the side with no view on its own cost of money loses every trade. None of that ships in a product release. It sits between your controller and your treasurer, in a gap no vendor is paid to close.

Agor AI Advisory works in that gap. We map where your cash timing leaks today, put a dollar figure on it before anyone signs a contract, design the policy and ownership model that turns agents into a treasury function, and choose tools that fit the architecture rather than the reverse. The companies that do this in the next two quarters will spend 2027 collecting a spread their rivals cannot see on any report they currently run.

Your payables have a yield, and so do your receivables. Right now a room full of busy, careful people is giving it away one Thursday batch at a time.