Blog
Finance automation is shifting from scripted RPA to agentic platforms. Discover how AI agents tackle complex AP exceptions, improve finance workflows, and transform controls and auditing.
The renewal call started the way these calls usually do. The account manager thanked the AP team for three years of partnership, shared a slide with the number of invoices their bots had processed, and then clicked to the next slide. The logo was the same. The product name was not. The company that sold them "robotic process automation" in 2023 now described itself as an "agentic automation platform," and the renewal quote had a new line item for agents.
The AP manager had one question, and it was a fair one. What would actually be different on Monday morning?
Here is the invoice she had in mind. A packaging supplier bills $48,312.60 against a purchase order for $46,900.00. The difference is a freight surcharge of $1,412.60 that nobody on the buying side remembers agreeing to. Under the current setup, the bot reads the invoice, tries the three-way match, fails, and drops the invoice into an exception queue. A person then opens the PDF, pulls up the PO, digs out the supplier contract, emails the buyer, waits two days, and finally decides whether to pay, short-pay, or dispute. The bot saved about ninety seconds of typing. The person spent forty minutes on the part that mattered.
The new pitch says an agent would handle that forty minutes. It would read the invoice, notice the surcharge, check the contract for a freight clause, see that freight above $1,000 requires buyer sign-off, draft the query to the buyer, and hold the invoice with a clear note about why. That is a very different promise. Finance teams deserve a clear way to judge whether it holds up.
This piece is about that shift. Not which vendor is winning it, but what it means for the people who own the invoices, the close calendar, and the audit.
Finance automation has moved in waves, and each wave solved the problem the last one left behind.
The first wave was scripting. RPA bots copied what a person did on screen. Open the email, download the attachment, type these fields into the ERP, click save. It worked well for tasks that looked identical every time. It broke the moment a screen changed, a field moved, or a supplier sent a slightly different layout. Many finance teams still remember the week an ERP upgrade quietly broke forty bots at once.
The second wave was document capture. OCR and template-based extraction pulled data off invoices, receipts, and statements so the bots had something structured to type. This helped, but templates had their own weakness. Every new supplier layout needed a new template, and every template needed someone to maintain it. Teams with thousands of suppliers ended up with a template library that nobody wanted to own.
The third wave is the one arriving now. Large language models made it cheap to read messy documents, reason about what they say, and decide on a next step. Automation vendors saw that the hardest and most expensive part of finance work was never the typing. It was the judgment in between. So they started rebuilding around agents, software that is given a goal and a set of permitted tools rather than a fixed recipe.
The market data shows how fast this is moving. Gartner predicts that a third of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. One of the largest RPA vendors told investors in May 2026 that agentic AI was its fastest-growing area, with 16 of its top 20 deals that quarter including AI and agentic components. Another major automation vendor launched a pre-built finance package that same month with more than 55 agents covering quote-to-cash, procure-to-pay, record-to-report, treasury, and tax. ERP suites are embedding agents directly into their finance modules. Accounts payable point tools are adding agent layers on top of their existing workflows.
Buyers are pulling in the same direction. In a Gartner survey of 314 organizations, 75% of CFOs said they were raising technology budgets for 2026, and nearly half planned increases of 10% or more. Finance leaders ranked AI agents as their second-highest future investment priority, right behind generative AI.
So the rebrand on that renewal slide is not a one-off. The whole category is moving. The question for finance is what the move changes in practice.
The simplest way to see the difference is to follow the same invoice through both kinds of system.
A scripted bot works like a recipe card. Step one, step two, step three. If step three does not match what the card expects, the bot stops and hands the whole thing to a human. It has no idea why the match failed. It only knows that it did.
An agent works more like a capable junior analyst who has been given a job description and a set of keys. The job description says "get this invoice ready to pay, or explain clearly why it cannot be paid." The keys are the tools the agent is allowed to use, such as reading the ERP, looking up the vendor master, opening the contract repository, and sending a message to a buyer. The agent decides which keys to use and in what order, based on what it finds.
Back to the $48,312.60 invoice. The scripted bot stops at the match failure. The agent keeps going. It reads the freight line, recognizes it as a surcharge rather than a pricing error, pulls the supplier agreement, finds the freight clause, checks the approval policy, and sees that anything over $1,000 in unplanned freight needs buyer approval. It holds the invoice, sends the buyer a two-line summary with the contract clause attached, and logs every step. When the buyer approves the next morning, the agent releases the invoice for payment.
Nobody typed anything. More to the point, nobody spent forty minutes reconstructing context that the system could have gathered on its own.
Three things change when a tool makes this jump.
First, variation stops being fatal. A bot fails when a supplier changes invoice layouts. An agent that actually understands documents reads the new layout the way a person would. For finance teams dealing with hundreds or thousands of suppliers, this is the difference between automation that covers 60% of volume and automation that covers most of it.
Second, the exception queue becomes the main event. Under scripted automation, exceptions were the leftover. Under agents, handling exceptions is the whole point. The routine invoices were already cheap. The value sits in the messy ones.
Third, the software starts making decisions, not just moving data. That is the part finance teams should think hardest about, because decisions come with accountability. A bot that types the wrong number is a data quality problem. An agent that approves the wrong payment is a controls problem.
Vendors are pushing agents into many departments, but finance keeps showing up first in launch announcements. There are good reasons for that.
Finance runs on documents. Invoices, credit notes, bank statements, remittance advices, W-9s, supplier contracts, expense receipts, and lease agreements all arrive in different formats from different senders. That variety is exactly what broke the older tools, and exactly what agents are built to handle.
Finance also runs on explicit rules. Tolerance thresholds, approval matrices, payment terms, and segregation of duties are written down. An agent needs clear boundaries to operate safely, and few departments have boundaries as well documented as a finance function preparing for audit.
The outcomes are measurable, too. Cost per invoice, days sales outstanding, touchless processing rate, days to close, and duplicate payment rate are numbers every controller already tracks. When a vendor claims an agent will help, a finance team can check the claim against a baseline it already has.
And finance has patience for proof. Gartner surveyed 160 senior finance leaders earlier this year and found that the lower-complexity use cases, including data extraction, AP and AR automation, and report creation, take roughly 9 to 10 months to show value. Forecasting and insight generation take longer. That timeline is a useful reality check. Even the "easy" agent use cases are not overnight wins, and finance leaders who plan for that are far less likely to abandon a project halfway through.
The rebrand on the vendor side creates real changes on the buyer side. Some are obvious. Several are easy to miss until a contract is already signed.
For years, finance teams bought automation by the bot. You paid for a number of licenses, each license ran scripts, and costs grew roughly in line with how many processes you automated. Agent platforms are moving to different pricing units. Some price per agent, some per task completed, some per document, and some on consumption of underlying AI capacity.
None of these is wrong. Each one behaves differently when volume changes. A per-document price that looks cheap at 8,000 invoices a month can look very different in the quarter your company acquires a business and volume jumps to 20,000. A consumption model can be efficient for steady work and surprising during year-end close, when agents are running reconciliations around the clock. Before any renewal, ask the vendor to model your costs at current volume, at double volume, and during your busiest week of the year.
When agents take over routine judgment, the work left for people changes shape. AP clerks spend less time keying and chasing, and more time reviewing agent decisions, handling the truly unusual cases, and managing supplier relationships. AR analysts spend less time matching remittances to open items and more time on disputed balances and credit risk.
This is good news for most teams, but it needs planning. Gartner's most recent finance research found that low AI literacy is now the biggest barrier to getting value from AI in finance, ahead of talent acquisition. The people who will supervise agents need to understand what the agent is allowed to do, how to read its reasoning, and when to overrule it. That is a skill, and it takes practice to build.
This is the change auditors will care about most. In a scripted world, controls sat around the automation. A bot keyed data, and a human approved the payment. The control point was obvious.
When an agent can read, decide, and act, the control points need to live inside the agent's design. Which approval limits apply to it? Whose authority is it acting under when it releases a payment? Can the same agent that creates a vendor record also pay that vendor? If it cannot, what stops it? When it makes a decision, can someone reconstruct exactly what it read, what rule it applied, and why?
Finance teams with SOX obligations should treat each agent like a new employee with a defined role, a defined approval limit, and a defined set of systems it can touch. If a vendor cannot show you how those boundaries are set and enforced, the agent is not ready for your payables.
Agents make decisions based on what they read. If the reading is wrong, every decision after it inherits the mistake. An agent that confidently misreads a remittance advice will confidently apply cash to the wrong invoice.
This is why the unglamorous work of document understanding has become more important in the agent era, not less. A strong reasoning layer sitting on top of weak extraction is a fast way to make wrong decisions at scale. When you evaluate an agent platform, spend as much time on how it reads your real documents as on how it reasons about them. Bring your worst samples. The faxed invoice with the coffee stain. The German supplier's credit note. The 14-page bank statement with continuation lines. That is where the difference shows.
Every category shift attracts relabeling, and this one is no exception. Gartner has a name for it, "agent washing," and estimates that only about 130 of the thousands of vendors claiming agentic capabilities actually have them. Many products marketed as agents are existing assistants, chatbots, or RPA workflows with a new name.
Gartner also predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, mostly because of rising costs, unclear business value, or weak risk controls. For finance, that is not a reason to wait. It is a reason to test carefully. A tool that still follows a fixed script is not worse because it is scripted. It is simply not what the new price tag says it is.
The most useful thing a finance team can do right now is walk into every vendor conversation with the same set of questions. They work whether the vendor started in RPA, ERP, document capture, or AP software, and they keep the conversation on your processes rather than the vendor's roadmap.
This is the single best question to ask. A good agent knows the limits of its confidence and escalates with context. A weak one either stops cold or guesses. Ask for a live example of an uncertain case and look at what the human receives. Is it a vague "please review," or a clear note explaining what the agent found, what it could not confirm, and what it recommends?
Pick one invoice from the demo and ask to see everything. What did the agent read? What data did it pull from which system? Which rule did it apply? Who approved the final action, and when? If reconstructing a single decision takes a support ticket, your auditors will not be happy.
Find out whether you can limit an agent to specific systems, specific actions, and specific dollar thresholds. Ask whether those limits sit in a configuration your finance team controls, or in code only the vendor can change.
Hand over a supplier invoice from a vendor you onboarded last month. Watch what happens. This tells you more about real-world accuracy than any benchmark on a slide.
Get the pricing model modeled against your real numbers, including peak periods and growth scenarios. Ask what counts as a billable unit and what happens to cost when an agent retries a failed task.
Most finance teams have years of bots, rules, templates, and ERP configurations. Ask whether the new platform runs alongside them, replaces them, or requires migration. A shift that forces you to rebuild every working automation before you see any agent value is a very expensive shift.
The finance teams getting real value from agents are not the ones who bought the biggest platform. They are the ones who picked a narrow starting point, measured it honestly, and widened the agent's authority step by step.
Start where documents dominate and the rules are already clear. AP exception handling is a natural first choice, because every team knows how many invoices land in the queue and how long they sit there. Cash application is another, especially for companies whose customers send remittance details in emails, PDFs, and bank portal exports that never look the same twice. Vendor onboarding works well too, since it involves W-9s, bank letters, and certificates of insurance that all need reading and checking before anyone gets paid.
Run the agent in shadow mode first. Let it process real documents and make recommendations without acting on them, then compare its calls to what your team actually did. After a few weeks, you will know its accuracy on your documents, not a demo set.
Then expand authority in small, measurable steps. A reasonable first rule might let the agent release invoices under $5,000 when the three-way match falls within a 2% tolerance and the supplier has a clean history. Anything outside those limits goes to a person, with the agent's full reasoning attached. As accuracy holds, raise the threshold. If it slips, lower it. Treat the agent's approval limit like any new team member's, earned over time and reviewed regularly.
Measure what matters to the business rather than what flatters the tool. Touchless rate is useful, but it can hide problems. Also track how long exceptions take to resolve, how often people overrule the agent, how many payment errors reach suppliers, and how much time your team gets back each week. Those numbers tell you whether the agent is actually carrying weight.
Artificio was built around AI agents from the start rather than retrofitting them onto OCR templates or screen scripts. That design choice came from the same observation that is now driving the whole market. The expensive part of finance work lives in the documents and the decisions around them, not in the keystrokes.
Our focus is the layer where every agent decision begins, which is reading the document correctly and understanding what it means. Artificio agents classify incoming documents, extract the data, check it against business rules and ERP records, and route exceptions to people with the context already gathered. They work alongside the systems finance teams already rely on, including SAP and other ERPs, existing RPA bots, and approval workflows, so teams can add agent capability without tearing out what already works.
Every decision is logged, every threshold is configurable by the finance team, and every escalation explains itself. That is how we think agents should behave in a function that answers to auditors.
Back to the renewal call. The AP manager does not need to decide whether "agentic" is the right word for what her vendor sells. She needs to know what happens to the $48,312.60 invoice on Monday morning.
If the answer is that the invoice still lands in a queue with no explanation, the product has a new name and the same behavior. If the answer is that the invoice gets held with the contract clause attached, the buyer gets a clear two-line question, and the whole path is logged for audit, something real has changed.
That is the test every finance team can apply, regardless of which vendor is on the call. Bring a real document. Ask what the agent does when it is unsure. Ask to see the trail. The shift from automation to agents is happening across the market, and it is a genuinely useful one for finance. The teams that benefit most will be the ones who judge it by what happens to their invoices, not by what appears on the slide.
Head IT Operations - Co Founder of Artificio
See it in your SAP environment
Bring us a document, a process, or a bottleneck. We'll show how Artificio captures, validates, and posts into SAP — then scale from there.
Security & compliance
ISO 27001:2013 certified, SOC 2 Type 2 compliant, GDPR and HIPAA ready. Every agent action is logged, auditable, and runs in isolated environments.
