The Last Invoice

It's 9:14 on a Tuesday morning, and Sylvia is staring at an invoice for $11,427.63 from a janitorial supply company in Memphis.

She's been doing this for nineteen years. Not staring at this particular invoice, though it feels like it, but sitting in this chair, in this fluorescent-lit room on the second floor of a regional healthcare company's back office in suburban Ohio, opening envelopes, keying numbers, matching purchase orders, chasing down approvals, and cutting checks. Nineteen years. The software has changed, barely. The screens look a little different now, the ERP has been upgraded twice, and somewhere along the way somebody replaced the dot-matrix printer with a laser one. But the work? The work is exactly the same.

Sylvia pulls up the purchase order in the system. She cross-references the line items: 40 cases of floor cleaner at $127.50 each, 25 boxes of trash liners at $84.30, a delivery surcharge that wasn't on the PO. She flags it. She'll need to email the facilities manager to confirm the surcharge, then wait, probably two days, for him to respond. In the meantime, this invoice sits in a digital pile alongside 847 others in various states of limbo. Some are waiting on approvals. Some have discrepancies. Some are duplicates that slipped through. Some are from vendors who submitted to the wrong email address, or sent a PDF that the OCR couldn't read, or attached a handwritten note that says "Please pay ASAP" as if urgency alone could route a payment through three levels of approval hierarchy.

I know Sylvia's world because I've spent the last fifteen-plus years in it. Not doing Sylvia's job, I don't have the patience, but building the systems that were supposed to make her job easier. I co-founded a payments company because I believed that the way businesses pay each other was fundamentally broken. And it is. It absolutely is. But what took me years to understand: making it easier was never going to be enough.

The AP department I just described isn't a relic. It's not some cautionary tale from a Gartner report about "laggards." It's the median. This is how the majority of American businesses process invoices today. The technology around them has transformed, the phones in their pockets can translate languages in real time, the cars in the parking lot can drive themselves on the highway, but the accounts payable department operates with a workflow that a time traveler from 2005 would recognize immediately. Open the envelope, key the data, match the PO, route the approval, cut the check.

Most companies still use paper checks for B2B payments. Three-quarters. In an era when I can send money to a stranger across the planet in seconds using my phone, the dominant method for businesses to pay each other involves printing a piece of paper, stuffing it in an envelope, licking a stamp, and trusting the United States Postal Service to deliver it within a week.

The numbers behind this inertia are remarkable and, frankly, a little embarrassing for our industry. Manual invoice processing carries heavy costs, ranging from $12.88 to $19.83 per invoice. Each one takes 10 to 30 minutes of human attention. Across the U.S. economy, businesses spend billions of hours per year processing invoices. That's not an abstraction. It's Sylvia, multiplied by millions, in offices across every city in America, doing the same repetitive work that could be done in seconds.

The gap between the best and the worst is enormous. Top-performing AP operations run 3.3 full-time equivalents per billion dollars in revenue. Bottom performers? 14.4. That's a 4x gap. Four times the headcount doing the same work, producing the same output, just with more errors, more delays, more missed early-pay discounts, and more vulnerability to the fraud that targeted 79% of companies in 2024.

I'm telling you all of this not to shame Sylvia or her employer. Sylvia's good at her job. She catches discrepancies that software misses. She knows which vendors are reliable and which ones pad their invoices. She has institutional knowledge that took two decades to build. The problem isn't Sylvia. The problem is that we built an entire industry around selling Sylvia slightly better tools, and then convinced ourselves we'd solved the problem.

For twenty-five years, the technology industry has operated under a single dominant paradigm: Software as a Service. SaaS. You know the model even if you've never thought about it explicitly. A vendor builds software, hosts it in the cloud, and charges you a monthly fee, usually per user, per seat, to access it. Salesforce pioneered it. Workday scaled it. ServiceNow, SAP, Oracle, thousands of others followed. It became the default business model for enterprise technology, and it created hundreds of billions of dollars in market value.

The pitch was always the same: We have a tool. It's better than what you have. It'll make your people more productive. And it did, genuinely. SaaS was a revolution. It democratized access to enterprise-grade software. It eliminated the nightmare of on-premise installations. It let a ten-person startup use the same CRM as a Fortune 500 company. I'm not here to bury SaaS. I benefited from it. My company runs on it.

But SaaS had a fundamental limitation that we papered over for a quarter century: it sold you the tool, not the work.

When you bought an AP automation platform, and I've evaluated dozens of them, you got a nicer interface for Sylvia to key invoices into. You got OCR that could read most documents most of the time. You got workflow routing that automated some of the approval chains some of the time. You got dashboards and reports and analytics that told you, with great precision, exactly how behind you were. What you didn't get was the invoices processed. That was still Sylvia's job.

The SaaS vendor got paid whether Sylvia processed a thousand invoices or ten. They got paid whether the OCR worked or failed. They got paid whether you captured early-pay discounts or missed every single one. Their incentive was to sell you seats, more users, more licenses, more revenue, not to make sure the work actually got done. And for a long time, that was fine. The tool was so much better than what came before, paper ledgers, green-screen terminals, spreadsheets, that nobody questioned the fundamental bargain.

That bargain is now breaking.

I remember the exact moment I understood this.

It was late 2024, and our engineering team was running a pilot with one of our larger platform partners, a procure-to-pay provider that processes payments for hundreds of mid-market companies. We'd been integrating AI models into our payment optimization engine for a while: predicting the best payment method for each supplier, identifying early-pay discount opportunities, flagging potential fraud. Standard stuff, or at least what passes for standard in a world that moves this fast.

But this pilot was different. We'd given the system something we hadn't given it before: agency. Not in the philosophical sense, we weren't debating consciousness in the conference room. Agency in the operational sense. The ability to make decisions and act on them without waiting for a human to click "approve."

The scenario was simple. An invoice came in from a supplier, a mid-size packaging company. The AI read the invoice, matched it against the purchase order, verified the receipt of goods, confirmed the supplier's bank details against our database of millions of supplier records, calculated that paying by virtual card would capture a 2.1% rebate, routed the payment, and sent a remittance notification to the supplier. Total elapsed time: four seconds. No human touched it. No human saw it until after it was done.

Four seconds. For a process that would have taken Sylvia twenty minutes on a good day, assuming the facilities manager actually responded to her email.

Now, I've been in payments long enough to know that a single transaction doesn't prove anything. One payment going through isn't a revolution; it's a demo. What made this different wasn't the speed or even the autonomy. It was what happened next.

The system processed another invoice. And another. And another. Forty-seven invoices in the first hour, each one following a slightly different path, different suppliers, different payment terms, different approval thresholds, different optimization strategies. The AI wasn't following a script. It was making decisions. It was choosing virtual card for one supplier because the rebate offset the convenience fee, ACH for another because the supplier didn't accept cards, and flagging a third for human review because the invoice amount was 340% higher than the trailing twelve-month average and the remit-to address had changed. That flag, by the way, caught what turned out to be a business email compromise attempt. A human reviewer confirmed the fraud in under a minute because the AI had already done the detective work, pulled the historical patterns, identified the anomalies, and presented the evidence in a clean summary.

I stood in the room watching this happen, and I had two simultaneous thoughts. The first was: This is it. This is what we've been building toward. The second was: Everything I thought I knew about the software business is wrong.

For my entire career, I've operated in the SaaS paradigm. Build software. Sell licenses. Help customers use the software to do their work more efficiently. The value proposition was productivity: our tools make your people faster. The pricing model reflected this: per seat, per user, per transaction. The customer did the work; we provided better instruments.

But what I watched that day wasn't a better instrument. It wasn't a tool that made Sylvia faster. It was a replacement for the work itself. The AI didn't help someone process an invoice, it processed the invoice. It didn't assist with payment optimization, it optimized the payment. It didn't flag something for a human to investigate, it investigated, reached a conclusion, and only escalated when its confidence dropped below a threshold we'd set.

This is not a subtle distinction. This is a phase change. And it has a name, though I didn't know it at the time.

Service as software.

The term was first articulated clearly by Foundation Capital in early 2024, and by the time I heard it, it hit me like a freight train because it described exactly what I'd been watching unfold in our own business. The idea is elegant and, once you see it, blindingly obvious: traditional SaaS sells access to a tool, and the customer does the work. Service as software sells the completed outcome, and the AI does the work.

Foundation Capital pegs the addressable market for service as software at $4.6 trillion. Sequoia calls it "an order of magnitude larger than cloud computing". Andreessen Horowitz says "software is eating labor."

I would frame it this way, because I think the VC framing, while directionally right, misses the visceral truth of what's happening:

For twenty-five years, SaaS sold you a gym membership. Service as software does your pushups for you.

The $500 billion global software market? That was the gym membership. Nice facility, good equipment, motivated staff. But you still had to show up and sweat. The $4.6 trillion services market? That's the actual exercise. That's the work itself. And now, for the first time in history, software can do the work, not just provide the tools for humans to do it.

This is the biggest economic shift since cloud computing. Possibly bigger. Cloud computing changed how software was delivered. Service as software changes what software delivers. It moves the entire value proposition from "a tool" to "the outcome." From selling shovels to selling holes in the ground.

I didn't see it coming. Not the full picture, anyway.

When I co-founded a payments company, the thesis was simpler. B2B payments were broken, too slow, too expensive, too manual, too vulnerable to fraud. We could fix that by building a payments-as-a-service platform that embedded directly into the procure-to-pay systems companies already used. We'd optimize payment methods, digitize checks, capture rebates, prevent fraud, and we'd do it as a service layer underneath the existing workflow. It was a good thesis. It worked. We grew.

But payments-as-a-service, as we originally conceived it, was still operating within the SaaS mental model. We were a better tool. A more efficient layer. A smarter pipe. The human was still in the loop, still matching invoices, still approving payments, still managing exceptions. We'd automated the payment execution, but the work of deciding what to pay, when to pay, and how to pay was still fundamentally a human process augmented by software.

What's happening now is categorically different. The AI agents we're building, and that our competitors are building, and that startups I've never heard of are building in garages right now, don't augment the human process. They perform the process. The human doesn't do the work faster with better tools. The human defines the rules, sets the thresholds, monitors the outcomes, and intervenes when something goes sideways. The human becomes the supervisor, not the operator.

This is the inversion. The entire relationship between software and labor is flipping, and it's flipping first in payments because payments has every characteristic that makes AI agency possible and profitable.

Payments has a unique combination of characteristics that make it the ideal first domain for AI agents. I'll lay out exactly why in Chapter 3. For now, the short version: it's the rare industry where the data is structured, the outcomes are measurable, the rules are explicit, and the cost of doing it wrong is both high and quantifiable.

If you were designing a domain for AI agents to conquer first, you'd design something that looks exactly like B2B payments.

And the infrastructure is being built to support exactly that. In 2025, Mastercard launched "Agent Pay", agentic tokens that enable AI agents to transact on behalf of users. Visa introduced "Intelligent Commerce" infrastructure for AI agent payments. PayPal, Stripe, and Coinbase all released agentic commerce toolkits. These aren't science experiments. These are the payment rails being constructed for a world where software doesn't just help humans make payments, software makes payments on its own.

The numbers tell you how fast this is moving, if you know where to look.

BILL, one of the largest AP automation platforms for small and mid-size businesses, launched AI agents in late 2025. Within months, AI-processed transactions increased 533%. Not 5%. Not 50%. Five hundred and thirty-three percent. Their system processed 1.3 billion documents and stopped 8 million fraud attempts. Tipalti, valued at $8.3 billion and processing over $28 billion in annual payments, deployed AI agents that autonomously execute routine finance tasks. AvidXchange launched AI-powered purchase order matching agents and rebranded their offering as "AP as a Service", language that would have been nonsensical three years ago. Ramp claims customers using their AI achieve 7x fewer clicks than legacy systems and close their books two days earlier.

Meanwhile, an entirely new class of companies is being born, companies that never had a legacy system to begin with. Natural raised $9.8 million to build payment infrastructure designed from the ground up for agent-executed transactions. Catena Labs raised $18 million for an AI-native financial institution using stablecoin-based payment rails for the agent economy. Sardine raised $70 million for AI agents handling KYC, AML, and fraud screening. These aren't retrofitting AI onto old architectures. They're building as if the human-in-the-loop never existed.

The AI agent economy, meaning the market for autonomous AI systems that perform work rather than just assist with it, stood at $3.66 billion in 2023. It's projected to reach roughly $140 billion by 2033. That's a roughly 44% compound annual growth rate. In venture capital, AI startups attracted over $100 billion in funding in 2024-2025, with AI accounting for roughly two-thirds of total fintech deal value according to PitchBook and CB Insights data.

I'm not sharing these numbers to impress you. I'm sharing them because when you line them up, they tell a story that most people in the payments industry haven't fully absorbed: this isn't a feature upgrade. This isn't the next version of AP automation. This is a new category of economic activity where software performs services that humans used to perform, priced on outcomes rather than access, and scaling at near-zero marginal cost.

The SaaS model charged you per seat because it assumed humans would be sitting in those seats doing the work. When the work itself is performed by AI, the seat is empty. And if the seat is empty, per-seat pricing doesn't just look outdated, it looks absurd. We're already seeing this play out: seat-based pricing dropped from 21% to 15% of B2B software companies in just twelve months. Intercom abandoned its $39-per-seat model for its Fin AI agent and switched to $0.99 per resolution. The result? Forty percent higher adoption, with one enterprise customer cutting support costs by 60%.

The math is inescapable. If AI makes one person as productive as five, your per-seat revenue just dropped 80% even as your customer's value increased 400%. This is the innovator's dilemma applied to pricing models, and it's going to break every SaaS business that doesn't adapt.

So that's the landscape. That's what I see from my particular vantage point, the operator's chair at a company that sits at the intersection of AI, payments, and the service-as-software shift. And the view is clarifying.

Exhilarating because the problem I set out to solve, the brokenness of B2B payments, is finally being solved, not incrementally but fundamentally. The technology exists to process an invoice in seconds, optimize the payment method, capture rebates, prevent fraud, and notify the supplier, all without a human being involved. That's not a distant prediction. That's current capability in controlled environments.

Terrifying because the implications extend far beyond payments, and I don't think most business leaders understand what's coming. If AI agents can process invoices, they can match purchase orders. If they can match purchase orders, they can manage procurement. If they can manage procurement, they can optimize supply chains. If they can optimize supply chains, well, you see where this goes. The boundaries between "accounts payable" and "accounts receivable" and "treasury" and "procurement" start to look like what they are: arbitrary lines drawn by humans who needed to divide complex work into manageable chunks. AI doesn't need those boundaries. AI doesn't care about your org chart.

I wrote this book because I believe we're at the beginning of the most significant restructuring of back-office operations in the history of modern business. Not because AI is magic, it isn't; it succeeds only about 58% of the time on single-step tasks and 35% on multi-step ones, and claims to the contrary are misleading. But because the trajectory is unmistakable, the investment is massive, and the early results are already good enough to redraw the economics of entire industries.

This book is for the CFO who knows something big is happening but can't separate the signal from the noise. It's for the finance leader who's been pitched seventeen AI solutions in the last quarter and doesn't know which ones are real. It's for the founder building in this space who needs a framework to explain what they're doing and why it matters. It's for anyone who processes a payment, receives a payment, or manages the people who do, which is pretty much everyone in business.

This book traces the arc from SaaS to service as software, explains why B2B payments is the proving ground, and delivers a practical roadmap for leading through the transition. I'll be honest about the failures along the way, including the companies that replaced humans too fast and had to hire them back.

But I'm going to start with something simpler. A question.

Remember Sylvia? The AP specialist with nineteen years of experience, staring at that invoice from the janitorial supply company in Memphis? She's still there. She'll be there tomorrow morning. She'll open the same application, key the same data, chase the same approvals, cut the same checks.

But for how long?

The invoice she's processing today may be one of the last her department ever touches. Not because she'll be fired, the story is more complicated and more interesting than that. But because the work itself is migrating from human hands to AI agents, one invoice at a time, one payment at a time, one decision at a time. And when enough of those decisions are being made by software instead of people, we won't just have a new technology. We'll have a new economy.

That economy has a name. We're calling it "service as software." And it starts, as so many revolutions do, with something no one thought was interesting enough to disrupt.

It starts with a payment.

The AP Crisis: By the Numbers

These are the baseline metrics that define the problem. I'll reference them throughout the book rather than repeating them.

Cost per invoice (manual processing): $12.88-$19.83

Processing time per invoice: 10-30 minutes

Annual hours spent on invoices (US): billions (estimated)

Companies still using paper checks for B2B payments: 75%

Companies targeted by payment fraud (2024): 79%

FTE efficiency gap (best vs. worst performers): 3.3 vs. 14.4 per $1B revenue

Global B2B payment volume: approximately $120 trillion/year

What I Mean by "Agentic"

You're going to see this word a lot in this book, so let me be specific about what it means, because precision matters here.

An AI agent is software that makes decisions and takes actions on its own. Not software that gives you a recommendation and waits for you to click "approve." Not software that drafts something for you to review. Software that actually does the work.

It doesn't recommend that you pay an invoice. It decides whether to pay it and executes the payment. It doesn't suggest a supplier might be risky. It adjusts payment terms and flags the relationship for human review. It doesn't identify potential fraud and send you an alert. It blocks the suspicious transaction and notifies the right people.

This is a different animal from what came before. RPA (Robotic Process Automation) follows rigid scripts; if the invoice format changes, the bot breaks. OCR (Optical Character Recognition) extracts data but has no idea what to do with it. Traditional workflow automation routes tasks but doesn't complete them. Chatbots answer questions but don't take actions. AI agents do all of these things and connect them into a continuous, autonomous workflow.

Now, the word "autonomous" makes people nervous, and it should. So let me be equally clear about where agents stop.

The AI agents I'm describing operate within boundaries set by humans. They have spending limits they cannot exceed. They have approval chains they cannot bypass. They have escalation rules that route the unusual, the ambiguous, and the high-stakes to a human who makes the call. An agent can decide how to pay a legitimate invoice, but it cannot decide whether the underlying purchase was appropriate. It can optimize payment timing for cash flow, but it cannot renegotiate a supplier contract.

Think of it like a guided system with clear boundaries. A human chooses the destination and sets the rules. The AI handles the routine operations. If conditions get too complex or unusual, it escalates and asks for help.

One note on terminology: throughout this book, I use "AI agent," "autonomous agent," "intelligent agent," and "agentic AI" interchangeably to mean software that acts, not just assists. The distinctions between these terms matter to researchers; they don't matter for this book's argument.

This distinction between "tool" and "actor" is the hinge of everything that follows. The entire SaaS industry was built on selling you better tools. What's happening now is that software is becoming the actor. That shift, from tool to worker, is what I mean by "agentic," and it's what makes the transformation in this book possible.