Why touchless AP is a myth—and what really works for accounts payable automation
Vendors have sold finance leaders on a compelling promise: touchless AP—a fully automated process where AI handles invoices completely hands-free, from receipt to payment, with zero human intervention. The pitch is simple. The appeal is obvious. And for most organizations, it’s fundamentally false.
Spend enough time in the accounts payable world and the touchless AP narrative breaks down quickly. The truth is that touchless AP works beautifully in marketing decks and for a narrow slice of trivial scenarios—one-line invoices, verified vendors, perfect PO matches. But the moment real complexity enters the picture, the system fails. And most organizations’ actual AP work is complexity.
The real problem isn’t that touchless AP is impossible. It’s that chasing it as a goal actually makes your AP operation worse. You end up with systems that require extensive upfront configuration, break when your business changes, and shift the manual work rather than eliminate it. Finance teams still spend enormous time maintaining rules, correcting AI mistakes, and handling exceptions.
There’s a better vision—one where AI works alongside your team as an intelligent partner. This post walks through what’s really happening in touchless AP automation today, shows how competitors are marketing fantasy while delivering limitations, and introduces a fundamentally different approach.
The myth of “touchless AP”
What does “touchless” actually mean?
When vendors in accounts payable talk about touchless AP, they’re describing a simplistic scenario. An invoice arrives. It gets coded. It gets routed to the right approver. It gets approved. Payment is made. All without a single person in your finance team needing to lift a finger.
In theory, it sounds perfect.
In reality, it’s a fundamental misunderstanding of what accounts payable actually is.
Why your real AP process can’t be touchless
Let me paint you a picture of what actually happens in a mid-market or enterprise company’s AP operation. It’s messy. It’s layered. It requires judgment.
- The verification problem: When an invoice line item arrives, someone has to verify that the goods or services were actually delivered. This isn’t a data entry task – it requires someone to check your receiving documentation or project records and confirm that what’s on the invoice matches what was actually received.
- The allocation challenge: An expense might need to be split across three different departments, two projects, and four different GL accounts. The invoice itself doesn’t tell you that – your business logic does. Someone has to make that decision based on context that only exists in their head.
- The exception handling: Missing PO numbers, pricing/quantity discrepancies, unrecognized vendors, changed banking information – these scenarios emerge constantly in any real business. Handling them requires reasoning, not just rule application.
- The complexity multiplier: Now imagine invoices with multiple line items, multiple approvers, subsidiary entities, different currencies, and compliance requirements. The moment you introduce real-world complexity, the notion that an invoice can flow from start to finish untouched becomes absurd.
The gap between simple and complex
Here’s the uncomfortable truth about “touchless” automation: it works smoothly for trivial scenarios.
If you have a one-line invoice for $500 from a verified vendor that perfectly matches an existing purchase order, then yes, you can auto-approve it. That should be automated. That’s not interesting or valuable – but it’s definitely automatable.
The problem is that in most mid-market companies, those trivial cases represent maybe 20-30% of your invoice volume. The other 70-80%? That’s where the real work lives. That’s where touchless AP breaks down.
The hidden maintenance burden nobody talks about
When vendors market “touchless AP,” they’re often hiding something crucial: the tremendous amount of manual configuration that happens behind the scenes.
To achieve what looks like 100% touchless automation on day one, these systems require extensive rules-based setup. Your team spends weeks or months building if/then statements: “if vendor code = X and amount < $Y, route to approver Z.” It feels like teaching the system your processes.
But here’s what happens: your organization changes. You hire a new approver. You open a second office. Your vendor consolidation initiative merges three vendors into one. Your GL structure gets reorganized. Suddenly, every one of those hard-coded rules needs updating. Someone – likely your AP manager – has to go in and reprogram the system.
That’s not zero touch. That’s just outsourcing the touches to the backend configuration layer, and then forgetting about it until something breaks.
From rules-based automation to agentic AI: the evolution
Yesterday’s automation: the OCR and rules era
I want to be fair about where we came from. The previous generation of AP automation was genuinely innovative for its time.
These systems could read invoice data using OCR technology. They could apply pre-set rules to route and code invoices. They were significantly faster than purely manual processing. And for straightforward scenarios, they worked reasonably well.
But they were also fundamentally brittle. They couldn’t adapt. They couldn’t learn. They couldn’t handle anything outside the bounds of their programmed rules. As soon as something new emerged – a new vendor format, a new GL structure, an exception scenario – the system either failed or required manual reprogramming.
What agentic AI actually is (and why it matters)
This is where the technology landscape is shifting, and it’s worth understanding what’s really new here.
Agentic AI is fundamentally different from the rule-based systems that came before it. An agentic AI system isn’t passively applying pre-set rules. It’s actively observing patterns. It’s making judgments based on context. It’s taking actions autonomously, the way a smart employee would.
In the context of AP, think of it this way: a traditional system is a bureaucrat following a flowchart. An agentic AI is an experienced colleague who knows your business, understands the patterns, and can make smart decisions even when something new comes up.
The key distinction is continuous learning. A traditional system stops learning after initial setup. An agentic system keeps getting smarter with every invoice it processes.
Why vendors are suddenly talking about “agents”
Recently, companies like Ramp have started popularizing the language of “AP agents” in their marketing. This is a direct response to the limitations of the older rule-based approach. However, agents are multiple and don’t speak to the full context of the problems in AP.
The industry was recognizing a fundamental problem: without a more adaptive form of AI, automation hits a hard ceiling. Vendors understood they needed to evolve or become irrelevant. So they started using the language of agentic AI.
But here’s where marketing gets ahead of reality.
How competitors are overpromising: Ramp and Bill
Ramp’s “touchless AP” approach and where it falls short
Ramp is an expense software that offers corporate credit cards with some added AP capabilities, claiming fully touchless invoice processing powered by multiple AI agents.
What they claim:
- 2x faster approvals thanks to AI recommendations
- AI can auto-approve invoices with a 90%+ acceptance rate by approvers
- Over 60 fraud signals to catch issues
- Ability to handle multi-line invoices “touchlessly”
What’s actually happening:
When I look past the marketing, here’s what I find: Ramp’s automation works well for simple cases. One-line invoices. One approver. Tied to a Ramp card transaction. That’s their sweet spot.
But the moment complexity enters the picture – multiple approvers, multiple GL allocations, non-standard scenarios – the system struggles. Much of what Ramp markets as AI is actually just rule-based automation that the customer has to pre-configure.
The “2x faster approvals”? That only happens because you’ve manually set up static approval routes in advance. The AI isn’t learning who should approve what – you’re hard-coding it, and Ramp is just executing your rules.
The real limitations:
That 90% acceptance rate on AI-driven approvals? It sounds good until you do the math. In a 1,000-invoice scenario, 100 invoices still need manual correction or rerouting. Your finance team could spend dozens of hours fixing those errors, erasing the time saved on the others.
Ramp doesn’t adapt when your organization changes. If you reorganize your approvers or update your approval policies, Ramp won’t naturally adjust. You have to go in and update the rules manually.
Their fraud “signals” are mostly generic data anomaly checks with little evidence of actually preventing fraud before payment happens. It’s more reactive than preventative.
And despite claims of multi-line support, Ramp can only auto-code line items if a PO is attached. For non-PO multi-line invoices – which are common in service scenarios – Ramp’s AI essentially gives up.
What’s happening behind the curtain:
Ramp’s documentation shows administrators setting up if/then rules to “teach” the system. This isn’t magic – it’s the customer doing heavy lifting upfront. Ramp’s AP automation is more static than agentic. It’s useful for routine tasks in controlled scenarios, but the agents don’t have full context and can’t handle the messy middle ground where real judgment is required.
Bill’s AI agents: the promise versus reality
Bill recently announced a new suite of “intelligent AI agents” with claims of “touchless transactions” across AP, expense management, and vendor onboarding.
What they claim:
- Automatically extract invoice data
- Reconcile receipts to expense reports
- Collect and validate W-9s from vendors
- Accelerate new user onboarding
- Trained on a trillion dollars of transactions and a billion documents
What the product actually does:
The actual capabilities are more limited than the marketing suggests. In AP specifically, Bill’s AI agents focus on capturing a few key fields when you upload an invoice: vendor name, invoice number, date, and amount.
That’s useful for data entry, but it stops far short of full invoice processing. The system doesn’t automatically code invoices to the correct GL accounts or departments. It doesn’t handle multi-line allocations. Most coding still requires an AP staffer because the AI isn’t context-aware enough to choose the right accounts or dimensions.
The W-9 collection fantasy:
Bill’s marketed “autonomous W-9 collection agent” is actually just email tools and response tracking. You email vendors requesting documents, and Bill tracks whether they’ve responded. That’s helpful, but it’s not an AI calling vendors on your behalf or autonomously approving forms. It’s a feature update dressed up in agentic language.
The onboarding story:
The “agentic-powered onboarding” is essentially a guided setup process for new users in their expense module. It’s not an AI that learns your company’s policies. It’s a useful workflow, but not intelligent in the way that matters for AP.
The learning question:
The claim about being trained on vast data is conventional, so what matters is whether the AI learns your business. Current evidence suggests Bill’s AP AI doesn’t adapt meaningfully to each customer’s unique patterns. If Bill’s AI miscodes an invoice and you correct it, does it get smarter next time? The documentation suggests the AI’s “learning” is mostly generic model improvement, not customer-specific adaptation.
Scope reality check:
Bill can match expense receipts to transactions if amounts and dates align, saving employees time in expense reporting. But in actual AP approvals and validations, there’s no autonomous agent deciding whether an invoice is okay to pay. A human still needs to review and approve. Bill might flag obvious duplicates or missing fields, but it’s not conversing with approvers or chasing exceptions.
The pattern both reveal:
Both Ramp and Bill illustrate something crucial about the current market: AI features that sound end-to-end in marketing language but are narrow in reality. These solutions handle pieces of the AP puzzle – data capture, simple routing – but don’t genuinely adapt to the multi-layered workflows that real companies operate.
Finance teams still end up doing enormous amounts of manual work: maintaining rules, correcting AI mistakes, managing exceptions. The efficiency gains plateau quickly because the AI isn’t continuously learning the organization’s unique processes.
What’s missing from touchless AP solutions
When I look at the limitations of these companies, a few things become clear:
- They can’t handle real complexity. AP agents work in controlled scenarios with simple structures. But mid-market and enterprise companies have multiple entities, multiple currencies, complex approval chains, and non-standard exceptions. When the AI hits something outside its training, it breaks down.
- They don’t truly learn. They might improve their generic model over time, but they don’t learn your business. They don’t understand your vendor relationships, your approval patterns, your exception handling philosophy, your GL structure. They’re generic assistants, and do not have full context of the accounts.
- They create hidden work. By requiring extensive upfront configuration and rules maintenance, they actually shift the burden rather than eliminate it. You’re not saving time – you’re just moving where the time gets spent.
- They miss the point of AP. Accounts payable isn’t about achieving some theoretical maximum of automation. It’s about freeing your team to do work that actually matters – analyzing spend, building vendor relationships, managing cash flow – instead of being buried in data entry and exception handling.
Billy: a different vision of agentic AI
This is where I want to introduce you to a fundamentally different approach.
Meet Billy: your AI employee, not an agent
Stampli’s Billy isn’t a simple feature. It’s much more fundamental: an AI team member embedded in your AP process.
Think of it this way: Stampli structures your AP process in the platform, and Billy operates it. Billy isn’t a generic assistant designed to work with any company. It’s your company’s dedicated AP expert, already equipped with 83 million hours of AP experience and ready to start contributing from day one with context on how you operate.
There’s something meaningful in that distinction. You’re not adopting a tool. You’re hiring a colleague.
Why Billy excels at complexity
Here’s where Billy’s design philosophy becomes clear: it was built for the real world, not for the simplified fantasy that most automation vendors are selling.
Multi-line invoices: Whether an invoice has one line or twenty, whether it relates to zero POs or multiple POs, Billy can interpret and code it appropriately. Billy understands multi-line invoices even without purchase orders backing them, having context, vendor history, and business logic to make smart decisions.
Complex approval hierarchies: Billy recognizes when an invoice needs department manager approval and project lead approval and routes it accordingly. It understands that different types of invoices need different approval chains. Complexity doesn’t break Billy, it’s where Billy demonstrates its value.
Real-world exceptions: Billy handles scenarios that most automation fails on – unusual vendors, non-standard invoices, allocation decisions across multiple GL codes. These are the situations where most AI systems throw their hands up.
The fundamental difference is that Billy isn’t trying to achieve 100% touchless automation. It’s optimizing for what matters: eliminating the grunt work while preserving the strategy for the finance team.
Learning within your organization
Billy learns from every interaction. As it processes invoices, it observes the corrections and decisions your team makes. It learns from your ERP data – vendor masters, chart of accounts, historical transactions. It learns from user behavior – how particular vendors are consistently coded, which approvers typically handle certain invoice types, what exceptions your team considers normal.
Over time, Billy’s suggestions and actions become increasingly aligned with your organization’s unique patterns. When something changes, Billy adapts naturally – no manual rule updates.
This adaptability translates directly into real-world results for finance teams:
Colin Madden, Controller at The Pines at Davidson, shares “Billy has helped our efficiency. Billy’s suggestions, especially the more that we’ve used the system, are almost always right. It saves time without having to search for the GL accounts or the other different codings that we use. To see what we’ve used in the past and see the suggestions and just click them – that’s been a great time saving.”
How Billy approaches approvals differently
Here’s another distinction that matters: Billy’s approval engine is dynamic, not static.
Traditional systems say: “Invoice type X always goes to Approver Y.” Billy says: “This invoice from ABC Supplies should go to Jane in Ops because she’s approved the last five invoices from that vendor, and typically Jane approves these before they go to Mark in Finance.”
Billy provides context and reasoning with its recommendations. Approvers understand not just who to send it to, but why. This builds confidence. And because Billy is learning your patterns, its recommendations get increasingly accurate over time.
This reduces the time AP staff spend figuring out routing. It also means approvers trust the system – they’re not sending it blindly to someone random, they’re sending it to someone who historically handles this exact scenario.
Proactive risk management built in
Billy isn’t just processing data – it’s actively preserving your AP process.
- Duplicate detection: As invoices enter Stampli, Billy and the system automatically check against all invoices in your ERP and Stampli to identify potential duplicates — before they ever reach approval.
- Vendor verification: Billy’s vendor management capabilities ensure vendor banking and profile updates are reviewed and verified by authorized users, protecting against accidental or fraudulent changes.
- Compliance monitoring: Billy lets you block payments to non-compliant vendors, set document expiration alerts, and enforce submission of required forms before an invoice moves forward.
Unlike tools that flag risks after payments are sent, Billy is designed to catch issues early, ensuring accuracy and compliance from the start.
How humans and AI actually work together in Billy’s design
Here’s something crucial: Stampli and Billy are designed so that AI and humans work in tandem, not in replacement.
Billy automates the heavy lifting – data entry, coding, routing, verifying routine details – which frees your team to focus on exceptions and strategic work. But your team is always in control: nothing gets paid without appropriate approval.
Billy provides a full audit trail of every action and recommendation, so you can always see why it suggested something. When Billy isn’t confident about something, it flags it for human review rather than guessing.
Billy’s philosophy is simple: AI should eliminate drudgery, not decision-making. The goal is efficiency without sacrificing accuracy or governance.
The Stampli difference: complexity and AI working together
Here’s something that distinguishes Stampli fundamentally from most of its competitors.
Many solutions force you to choose: either you get sophisticated processes with weak or no AI, or you get flashy AI that only works if you simplify everything. It’s a false binary.
Stampli is built to handle both. You get a robust, flexible AP platform that can support multiple entities, multi-currency operations, layered approvals, complex allocation rules – all the complexity that mid-market to enterprise companies actually need. And layered on top of that, you get Billy as an employee who adapts to that complexity rather than being limited by it.
This matters because it means you don’t have to compromise. You don’t have to strip down your controls to enable automation. You don’t have to sacrifice policy nuances to make AI work. Billy adapts to your process – you don’t have to bend your process to fit the AI.
That’s a crucial difference when you’re trying to maintain governance while pursuing efficiency.
Rethinking AP transformation: beyond the touchless myth
The real goal of AP automation
I want to circle back to something fundamental here. The goal of AP automation isn’t to achieve some mythical state of zero human involvement. That’s not the point.
The real goal is to maximize automation where it makes sense and minimize human effort on low-value tasks. The ROI isn’t in achieving some percentage of “touchless” invoices – it’s in freeing your team from repetitive processes so they can do work that matters, while maintaining control of their AP process.
That’s the crucial distinction.
What this means for your AP team
The end goal of all this technology is to empower your finance team. The future isn’t zero-touch AP – it’s better-touch AP.
It’s an AP process where your team intervenes only where they uniquely add value. Reviewing an odd exception. Building vendor relationships. Analyzing spend trends. Making strategic decisions about vendor management or cash flow. They handle the judgment calls.
The repetitive grind – data entry, routine coding decisions, basic matching – goes to Billy.
When your AP team is freed from that work, something remarkable happens. The role transforms from data processors to process managers and strategists. The team gets to think more strategically about vendor relationships, spend optimization, and risk management.
That’s not job elimination – that’s job elevation. Learn more about Billy here.
FAQ: common questions about touchless AP
1. What is touchless AP / touchless invoice processing?
Touchless AP refers to a fully automated process in accounts payable with no human involvement. This concept sounds ideal but is impractical in real-world finance operations, due to AP workflows that require human verification, allocation, and judgment. Billy, the AI employee operating within Stampli, serves as the best option for automating mundane tasks, freeing up strategic decisions for the finance team.
2. How does AI in accounts payable actually work?
AI in accounts payable learns from data patterns to automate invoice capture, coding, routing and other processes. AI agents have entered the market to automate simple tasks and workflows, however they still require maintenance and break within workflow complexity. An AI employee, like Billy, can understand context — learning how your team codes, approves, and handles exceptions — then acting autonomously on similar future invoices. This means Stampli’s AI doesn’t just process invoices faster; it gets smarter and more aligned with your organization over time.
3. What is agentic AI in accounts payable, and how is it different from standard automation?
Agentic AI acts like an experienced colleague rather than a programmed rule engine. Instead of relying on static if/then logic, Billy within Stampli observes patterns, reasons through context, and takes proactive actions — coding, routing, or flagging issues intelligently. This continuous learning makes Billy adaptive, resilient, and capable of handling complexity that breaks other systems.
4. What are the benefits of implementing AI / agentic AI for accounts payable?
AI and agentic systems dramatically reduce manual work, improve accuracy, and accelerate invoice cycles. With Billy or Stampli, your AP team focuses on analysis and relationships instead of repetitive data entry. The result is faster processing, stronger compliance, and higher-value finance work — efficiency without sacrificing control.
5. How can an organization improve its touchless AP rate using AI / agentic AI?
The best way to improve touchless AP isn’t chasing 100% automation — it’s using adaptive AI that learns and adjusts continuously. Billy increases automation coverage by recognizing invoice patterns, adapting to new vendors, and routing approvals intelligently. Over time, Stampli helps teams achieve modernization, where human effort is applied only where it truly adds value.
