A guide

AP automation cuts clerks, but only half of them. Here is which half, and why the other half still has a job.

The X thread you arrived from is technically right. AP automation companies are cutting AP clerks. They are not cutting all of them. There are two AP clerk jobs, and one of them survives every SaaS AP product on the market for a structural reason worth naming: the systems that job touches do not have the kind of API a SaaS AP vendor can wire up.

M
Matthew Diakonov
7 min

Direct answer, verified 2026-05-07

AP automation cuts the digital-invoice approval clerk. Bill.com, Stampli, Tipalti, Ramp Bill Pay, AvidXchange, Coupa, and Vic.ai all collapse the email-PDF-OCR-approve-API loop, and they replace the role that does it. They do not cut the desktop AP clerk who types approved invoices into SAP Business One, Oracle EBS, Jack Henry, Fiserv, FIS, NetSuite Desktop, QuickBooks Desktop, or a mainframe terminal, because those systems do not publish the API surface the SaaS AP vendor needs. That second clerk is cut by accessibility-API agents like Mediar, which drive the desktop ERP through Microsoft UI Automation (the same surface a screen reader uses) at $0.75 per minute of runtime. Confirmed against AvidXchange's August 2025 AP Pros Face Growing Layoff Concerns survey of 500 AP professionals.

The two AP clerk jobs.

When an industry analyst writes “AP automation is cutting AP clerks,” the unit they are measuring is rolled-up headcount. That rollup hides the split. The actual AP team in any company over a hundred employees is doing two distinct workflows that share a name and very little else.

The first workflow ends at the GL. A vendor sends a PDF, a clerk codes it, an approver clicks, and a record is pushed by API into NetSuite, QuickBooks Online, Sage Intacct, or Dynamics 365. This is the workflow Bill.com was founded to automate. Every modern SaaS AP product, plus the AvidXchange wave that came before, plus the new AI-native crop (Vic.ai, AppZen, Glean AI), is built around this shape. The clerk who runs it is being cut.

The second workflow ends at a desktop ERP that has no usable API. A food-and-beverage chain runs SAP Business One, the thick Windows client. A regional bank runs Jack Henry SilverLake. A community hospital runs Epic billing. A manufacturer runs Oracle EBS R12 with custom Oracle Forms screens. The approved invoice, the assigned claim, the new account, the patient intake form, all need to land in one of those systems. That landing is a human typing fields, twelve to forty of them per record. The SaaS AP wave does not reach this workflow because the SaaS AP wave is API-out and the desktop ERP is API-absent.

Two workflows that share a job title and very little else

Email arrives with a PDF invoice. The clerk drops it into Bill.com, Stampli, Tipalti, Ramp Bill Pay, or Coupa. The product reads the line items, suggests a GL code, routes for approval, and at the end pushes a record through a published API into NetSuite, QuickBooks Online, Sage Intacct, or Dynamics 365 Business Central. The whole loop happens in the browser. The clerk is the bottleneck, and the SaaS product replaces them.

  • API-driven, lives entirely in the browser
  • Targets cloud / modern accounting (NetSuite, QBO, Intacct)
  • Bill.com, Stampli, Tipalti, AvidXchange, Coupa, Vic.ai, AppZen
  • Industry-documented 60 to 80 percent headcount reduction on this loop

The two flows, side by side.

Top to bottom, here is what each flow looks like at the keystroke level. The shape is what every honest AP team will recognize. The first flow is what AvidXchange and Bill.com cut. The second flow is what Mediar cuts.

Tier 1 — what SaaS AP products cut

1

Vendor email

2

PDF invoice

3

OCR + GL coding

4

Approver in app

5

API push to GL

Tier 2 — what they cannot reach (and Mediar does)

1

Approved invoice

2

Open SAP / Oracle / Jack Henry

3

Type 12 fields

4

Save, screenshot

5

Email A/P inbox

The structural reason the second flow does not collapse the same way is on its third step. “Open SAP / Oracle / Jack Henry” is a Win32 process the SaaS AP vendor cannot reach over an API. It can only be reached by a different agent surface: Microsoft UI Automation, the same accessibility tree that screen readers consume. That is the surface Mediar drives, and it is the same surface the open-source Terminator SDK exposes for teams that want to write their own agent against it.

The numbers, when the desktop clerk is the one being cut.

These are workflow-specific, not aggregate marketing claims. Each one names the customer category, the workflow, the before number, and the after number. They appear verbatim in the public llms.txt brief on this site, and they are the kind of math an AP team can take to a CFO without flinching.

0%cost cut vs UiPath, F&B SAP B1 chain
$0K/yrinsurance claims, 30 min to 2 min
$0K/yrhealthcare patient intake savings
$0/minruntime billing, no per-seat
$750K/yr

Claims intake at one mid-market carrier went from 30 minutes per claim to 2 minutes. That is not a press-release number, that is the AP-team headcount math the CFO showed the board.

Mediar customer note, mid-market insurance carrier

The honest counterargument.

A reader on the X thread will push back: aren't Stampli and Vic.ai and AppZen using LLMs now? Doesn't that change the surface? It changes the extraction, not the surface. A modern LLM-driven AP product reads PDFs better than the OCR-and-rules pipelines from 2018, which is real progress. The bottleneck it removes is the clerk who codes the invoice. It does not remove the clerk who types the approved invoice into SAP B1 or Jack Henry, because no LLM can write to a Windows desktop ERP without an agent surface beneath it. That surface is what accessibility-API agents are.

A second pushback: doesn't Power Automate Desktop already cover this? Sometimes, narrowly. Power Automate Desktop ships with a UI Automation primitive, and on a Microsoft-native stack (Dynamics 365 Business Central, Excel, Outlook) it works. On SAP B1, Oracle Forms, Jack Henry, Fiserv, FIS, or a Citrix-rendered Epic Hyperspace, the recorded selectors break the moment a control's accessible name shifts, and the maintenance load is what made enterprise RPA centers of excellence expensive in the first place. The point Mediar is making with the self-healing layer is exactly this: when an accessible name or layout shifts, the agent reasons about which control it should be talking to instead of throwing a NotFound and stopping the bot.

A third pushback, which we hear from CFOs: aren't we just buying another vendor? Yes, but the unit you are buying is different. SaaS AP is per-seat or per-invoice. Desktop RPA (UiPath, Automation Anywhere, Blue Prism) is per-bot per-year. Mediar is per minute of runtime, which means an unused workflow costs nothing, and a workflow that runs once a quarter does not need its own license. That is the unit-of-work change behind the 70 percent cost cut at the F&B chain on SAP B1.

We moved an LG-customer F&B chain from UiPath to Mediar; their CFO told the board they're now saving 70% on costs. The cut wasn't on the digital-invoice clerk. That clerk had already been replaced. The cut was on the desktop SAP B1 clerk, which is the role UiPath had been licensed to handle and was billing per-bot for.
M
Mediar field note
On the workflow shape behind the 70 percent number

What this means if you are running an AP team in 2026.

Two practical reads. First, the AP-clerk-layoff narrative is real, but the rollup hides which clerks are being cut and which are not. If your team is split across a SaaS AP product on the front end and a desktop ERP on the back end (which most mid-market and enterprise finance teams are), the headcount math has two phases, run by two different vendors, on two different surfaces.

Second, the boundary tells you what to compare in your next AP RFP. Comparing Mediar to Bill.com or Stampli is a category mistake, the way comparing Snowflake to Tableau is. They live on different surfaces and cut different roles. If your AP RFP is comparing them on the same line item, the RFP is wrong, and the typical outcome is that the SaaS AP product wins the visible part of the workflow and the desktop ERP work stays manual until someone notices six months later.

The Mediar deployment we run for a finance team is twenty minutes on a call, then a workflow recording session on the customer's own machine, then the agent runs against that recording. The platform is SOC 2 Type II certified and HIPAA compliant, deploys on-prem or in a customer-owned cloud, and bills the desktop posting at $0.75 per minute of runtime, which is the unit-of-work change UiPath did not ship.

If your desktop AP clerk job is still manual, the cut is on a different vendor surface.

Twenty minutes is enough to walk a workflow you currently run by hand into SAP, Oracle, or Jack Henry, and tell you whether an accessibility-API agent collapses it or whether the API-driven SaaS wave already covers it.

Frequently asked questions

Which AP clerks does AP automation actually cut?

Two AP clerk jobs exist, and the answer is different for each. The first is the digital-invoice approval clerk: receives PDF invoices in email, drops them into Bill.com or Stampli or Tipalti or Ramp Bill Pay or Coupa, codes the GL line, routes for approval, and pushes the result via API into NetSuite, QuickBooks Online, or Sage Intacct. That clerk is being cut. Every API-driven AP automation product on the market cuts that role, with documented headcount reductions in the 60 to 80 percent range. The second is the desktop AP clerk: takes the already-approved invoice and types it, line by line, into SAP Business One, Oracle EBS, JD Edwards, Jack Henry, Fiserv Premier, FIS IBS, NetSuite Desktop, QuickBooks Desktop, or a mainframe terminal. That clerk is still there, because those systems do not publish the kind of API surface a SaaS AP vendor can integrate against. Cutting that second clerk requires a different kind of agent: one that drives the GUI through Microsoft UI Automation rather than through a REST endpoint.

Why can't Bill.com or Stampli or Tipalti or Ramp Bill Pay reach into SAP B1 or Oracle EBS or Jack Henry?

Because the SaaS AP architecture is API-out, and the legacy desktop architecture is API-absent. Bill.com integrates with NetSuite, QuickBooks Online, Xero, Sage Intacct, and Microsoft Dynamics 365 Business Central; these are cloud or modern accounting products with REST or SOAP surfaces. SAP Business One is a Windows desktop ERP whose primary interface is a thick client (the SAP B1 client) and whose published APIs (the DI Server, Service Layer, B1if) require a server-side install and a specific licensing tier that smaller B1 customers usually do not buy. Oracle EBS R12 has Oracle Integration Cloud and SOA Suite, but most line workers are using Oracle Forms screens that the SaaS AP vendor never integrates with. Jack Henry SilverLake, Fiserv Premier, and FIS IBS are core banking systems that ship with vendor-controlled integration partners and run-book paperwork, not a tenant-installable webhook. The SaaS AP product cannot ship a connector for a system whose vendor either charges five figures for the integration license or does not publish an integration surface at all.

Is the desktop AP clerk job actually still a job in 2026?

Yes, and the headcount is bigger than the AP-automation marketing suggests. The Bureau of Labor Statistics still tracks tens of thousands of accounts payable and bookkeeping clerks whose primary tool is a Windows desktop ERP, not a SaaS AP product. The AvidXchange 2025 'AP Pros Face Growing Layoff Concerns' survey found that the AP professionals who feel most exposed to layoffs are exactly the ones whose day is dominated by data entry into a desktop ERP, because their work is the most repetitive AND the most invisible to the SaaS automation wave. They survive because the SaaS wave physically cannot reach them. They get cut when an accessibility-API agent does, which is a different vendor category from the one their CFO is comparing in the AP RFP.

Why does an accessibility-API agent reach work that an API integration cannot?

Because the accessibility tree is universal where APIs are not. Every Windows GUI app, including the ones built in 1998, exposes its controls through Microsoft UI Automation, the same surface that screen readers consume. An accessibility-API agent reads the SAP B1 invoice form the way a screen reader does: 'edit field, label Vendor Code, value empty.' It writes the way a person does: focus the field, type the code, press Tab. It does not need a REST endpoint, an OAuth client, or a vendor-issued integration license. It does not need pixel matching, which is what makes vision-based RPA brittle when fonts or themes change. That is the exact difference between Mediar's surface and the surface that Bill.com and Stampli and Tipalti reach. Different surface, different boundary, different clerk job cut.

What does Mediar's pricing look like next to a UiPath quote for the same desktop work?

Mediar bills $0.75 per minute of agent runtime, plus a $10,000 turn-key program fee that converts to credits with a bonus, so it is effectively prepaid usage rather than a separate license. There is no per-seat fee, no per-bot fee, and no developer-certification fee. UiPath, Automation Anywhere, and Blue Prism quote on per-bot licensing, often in the $5,000 to $15,000 per unattended bot per year range, plus implementation services that for an enterprise SAP or Oracle project commonly land at $200,000 to $500,000 in the first year. We moved an LG-customer F&B chain from UiPath to Mediar; their CFO told the board they're now saving 70% on costs. That number is workflow-specific, not a generic claim, and it shows up in the section above.

Is this only useful for very large enterprises, or can a regional bank or a mid-market carrier use it?

Mid-market is where the math is cleanest. A regional bank on Jack Henry or Fiserv with an 8-week new-account onboarding window has the same workflow shape as a top-50 bank, but does not have the budget for a UiPath Center of Excellence. The same is true for a mid-market insurance carrier handling claims intake at 30 minutes per claim, or a community healthcare network on Epic or Cerner running patient intake by hand. Documented Mediar deployments at this size: bank onboarding from 8 weeks to 2 weeks, claims intake from 30 minutes to 2 minutes ($750K/year for one carrier), patient intake at $210K/year of recovered clerk time. The platform is SOC 2 Type II certified and HIPAA compliant, deployable on-prem or in a cloud environment the customer already owns.

What about the AI-Overview claim that 'AI is replacing the AP clerk'? Is that the same thing as Mediar?

No, and the conflation is what makes most coverage of this topic unhelpful. The AI-Overview headline is usually about LLM-driven document understanding (read a PDF, return JSON). That is genuinely better than the OCR-then-rules pipeline that AP teams have been running for fifteen years, and it is what Bill.com, Stampli, AppZen, and Vic.ai are shipping. It is also a different problem from getting structured data into a desktop ERP that does not have an API. LLM extraction plus a SaaS AP product cuts the digital approval clerk. LLM extraction plus an accessibility-API agent cuts the desktop posting clerk. The first half of the AP team and the second half are getting cut by different products built on different surfaces; the AI-Overview headline is reporting on the first.

Is the open-source Terminator SDK the same thing as the Mediar product, or something separate?

Terminator is the open-source Windows-automation SDK Mediar publishes at github.com/mediar-ai/terminator. It is an MIT-licensed library for driving Windows UI Automation from Rust, Python, or Node, and it is the same primitive layer Mediar uses internally. Teams that want to extend Mediar with a custom workflow, or write their own agent against the same accessibility-API surface, work in Terminator. The Mediar product on top adds the workflow recorder, the no-code web app at app.mediar.ai/web, the SOC 2 / HIPAA control plane, the audit logs, the validation rules, the self-healing layer that recovers when a control's accessible name changes, and the runtime billing. The Terminator SDK is the open part; Mediar is the productized layer for teams that want the workflow recorded once and the runtime managed.