Agentic process automation is automation where an AI agent figures out the steps instead of following a rigid script. Old-school bots (called RPA, short for robotic process automation) do exactly what you program them to do, in exactly the order you tell them. An agentic system looks at the goal, reads the situation, and decides what to do next on its own.
That flexibility is the whole pitch. It’s also the whole risk.
What is agentic process automation
Think of it like hiring help. RPA is the new hire who follows a recipe taped to the wall. Every step is spelled out: open this screen, copy this field, paste it here. It works perfectly until someone rearranges the kitchen. Then everything breaks.
Agentic process automation is the experienced cook. You say “make dinner for four, there’s chicken and rice in the fridge.” They figure out the rest. If the rice is out, they adjust.
The term itself was coined in a 2023 research paper out of Tsinghua University. It’s since become the label for any automation that uses AI agents to handle work that changes, rather than bots that replay the same steps. If you’re curious about the broader split between agentic AI vs generative AI, I broke that down separately. The short version: generative AI creates things when you ask. Agentic AI takes a goal and works toward it.
There’s a catch, though. The word “agentic” has become a marketing free-for-all. Some vendors mean “we added an AI step to our existing bot.” Others mean “a fully autonomous agent that runs end-to-end without humans.” Those are very different things. I’ll get into how to tell them apart in a moment.
The range of types of AI agents is wide, too. Some are simple (one task, one tool). Others coordinate multiple sub-agents across systems. APA sits somewhere in the middle for most real-world use.
How agentic automation works (the short version)
Every agentic automation system follows roughly the same loop, whether it’s built by a solo developer or an enterprise platform team:
- Read the inputs (an email, a PDF, a database record, a Slack message)
- Reason about what needs to happen (is this a new vendor? does this match a purchase order?)
- Decide the next step (approve it, flag it, route it somewhere)
- Act (update the record, send the notification, trigger the next workflow)
- Learn from the result (did the human override? store that for next time)
Say you process invoices. An RPA bot opens the email, finds the PDF, reads three fields from specific coordinates on the page, and pastes them into your accounting system. If the vendor sends a slightly different PDF layout, the bot breaks. You fix it. It breaks again next month.
An agentic system reads the whole PDF, understands it’s an invoice, figures out the vendor and amount regardless of where those fields sit on the page, checks it against your purchase orders, and routes it for approval. If the amount doesn’t match, it flags the discrepancy and asks a human to review instead of crashing.
That’s the promise. And for straightforward, messy-input tasks like this, it genuinely works. For the full design guide on how to build agentic workflows, I wrote a separate deep-dive. This post is about whether you should use APA at all, not how to wire it up.
APA vs RPA: what actually changes
The comparison table you’d expect, with a few columns that matter more than the usual ones:
| RPA | APA | |
|---|---|---|
| How it works | Follows a fixed script step by step | AI agent reads the situation and decides |
| Handles change | Breaks when the process or screen changes | Adapts to new layouts, formats, inputs |
| Best for | High-volume, stable, structured tasks | Variable inputs, exceptions, messy data |
| Predictability | High (does exactly what you told it) | Lower (may interpret differently each time) |
| Failure mode | Stops and throws an error | May complete the task incorrectly |
| Maintenance | Constant (30-50% of bots need rework yearly) | Lower maintenance, higher monitoring |
| Cost model | Per-bot licensing, predictable | Per-task or usage-based, less predictable |
| Maturity | 15+ years, proven at scale | 2-3 years, still early |
RPA isn’t going away. It’s a $3.8 billion market and still growing. 74% of companies already run it in some form.
But the pain is real too. Anyone who’s managed RPA bots knows the maintenance grind. When the vendor updates a screen or a form layout changes, bots break. Forrester estimates that 50% of RPA projects stall at the point where process variability exceeds what a scripted bot can handle. That’s exactly where APA is supposed to take over.
My take: APA vs RPA isn’t an either/or. It’s more like the difference between a dishwasher and a cook. The dishwasher does one job perfectly, every time. The cook handles the weird stuff. You need both. The question is which processes need which.
If you want to see how this comparison plays out across different intelligent automation approaches, that piece maps the full spectrum.
Where agentic automation works (and where it doesn’t)
APA doesn’t work everywhere. Knowing where it breaks matters more than knowing where it shines.
The best data on what actually happens in production comes from a study of 306 practitioners who run AI agents in real workflows. The findings are honest and a little sobering:
- 68% of production agents need human help within 10 steps
- 47% need help within 5 steps
- 74% still rely on humans to evaluate whether the agent did the right thing
Agents in production are simpler and more human-supervised than the marketing suggests. That’s not a failure. That’s actually the right design. The question is: where does that design make sense?
Three questions to ask before choosing APA over RPA:
- Does the process change often or handle varied inputs? (If yes, APA handles variation better than a scripted bot.)
- Is a wrong decision recoverable? (If the agent makes a mistake, can someone catch it before real damage?)
- Can a human review the agent’s work before it’s final? (A checkpoint where someone signs off on what the agent did.)
If yes to all three, APA is worth trying. If no to any of them, keep the bot.
The Klarna story tells the whole thing
Klarna’s AI agent is the most-cited success story in this space. And it’s real. According to their SEC filing, the numbers are serious: $60 million per year in cost savings, AI doing the work of 853 full-time employees, cost per customer interaction down 40%.
Then came the second part. In May 2025, Klarna started rehiring human agents. Customer satisfaction had dropped on complex cases. Disputes, fraud claims, hardship situations. The AI handled simple questions beautifully. The hard ones? Not so much.
Both things are true at the same time. That’s the whole point. APA is brilliant for high-volume, straightforward work where a wrong answer doesn’t ruin anyone’s day. It’s not ready for the tricky stuff.
For more real agentic AI examples with honest assessments of what works, I keep a running list. And the principles behind building reliable agents explain why this pattern (narrow scope, human checkpoint) keeps winning.
This is the kind of scoping exercise I do with founders and teams: figuring out which processes are worth automating with agents and which ones you should leave alone. If you’re working through that decision, happy to talk it through.
Agent washing: how to tell real APA from a rebrand
The term “agent washing” was coined by Gartner to describe what happened in 2025: thousands of vendors slapped “agentic” on their existing products. Only about 130 of them actually build genuine agentic systems. The rest are selling the same RPA or chatbot with a new label.
Every major automation vendor did some version of this. UiPath launched “Agent Builder.” Automation Anywhere acquired an AI company and rebranded around APA. Microsoft added agentic features to Power Automate. But UiPath’s own financial filings say agentic products are “not expected to significantly impact top line” revenue in the current fiscal year. That tells you how early this really is.
Red flags when evaluating an “agentic” platform:
- It can only follow pre-built workflows (that’s still RPA with an AI label)
- There’s no persistent memory between sessions (real agents remember context)
- It needs you to manually handle every exception (the whole point of APA is handling exceptions)
- Key integrations are “coming soon” (means they haven’t built them yet)
The simplest test: ask the vendor to demo a process it’s never seen before. If it can only run the pre-built demo, it’s not agentic. It’s a script with better marketing.
My take: The rebranding isn’t a scandal. Vendors follow demand, and “agentic” is what buyers are asking for. But it means you need to be a sharper buyer. The label on the box tells you nothing. What the thing does when it encounters something new tells you everything.
For vetted agent tools that actually deliver, I maintain a list of the best AI agents by category. And the broader AI agent marketplace is useful if you want to browse what’s available, as long as you bring healthy skepticism.
The practical migration path
The vendor pitch is “replace your RPA with our APA platform.” The reality is different. Forrester predicts that most enterprises will keep running their existing bots through 2026. Fewer than 15% of firms will even activate agentic features in their automation tools this year.
That’s not because APA doesn’t work. It’s because ripping out working automation to install something new and unproven is a terrible idea.
The smarter path looks like this:
- Keep your stable bots running. If a bot works and rarely breaks, leave it alone. It doesn’t need to be “agentic.”
- Send new automation requests to agents. When someone asks for a new workflow, build it with an agent instead of scripting another bot.
- Retire your highest-maintenance bot first. Pick the one that breaks the most. Replace it with an agent. Measure whether it actually reduces your maintenance load.
- Expand from there. If it works, do the next one. If it doesn’t, you’ve only risked one process.
McKinsey surveyed nearly 2,000 companies across 105 countries. Only 23% are scaling agentic AI, and most of those are only doing it in one or two business functions. That’s the right speed. Go narrow, prove it works, then expand.
Gartner’s prediction is blunt: over 40% of agentic AI projects will be canceled by the end of 2027. The reasons: unclear business value, escalating costs, and not enough guardrails. The projects that survive are the ones that started small and proved themselves.
For a practical build guide when you’re ready, see how to build AI agents. And if you want to use low-code automation tools to get started without a dev team, that’s a solid starting point. The agentic AI frameworks guide sorts the major platforms by skill level.
How I can help
The hardest part of agentic automation isn’t the technology. It’s the scoping. Picking the wrong process to automate wastes months and money. Picking the right one saves both.
I work with founders and small teams on exactly this. We map your processes, figure out where an agent would actually help (and where a simple bot or even a spreadsheet is fine), and build the first one together. No 50-slide decks. No “AI transformation roadmap.” Just finding where the leverage is and getting it running. If that sounds useful, let’s talk.
FAQ
What is agentic process automation?
Agentic process automation (APA) is automation where an AI agent figures out the steps to complete a task, rather than following a pre-programmed script. Where traditional RPA bots do exactly what you tell them in a fixed sequence, an APA system reads the situation, reasons about what needs to happen, and decides the next step on its own. The term was coined in a 2023 academic paper and has since become the industry label for AI-agent-powered business automation.
What is the difference between RPA and agentic process automation?
RPA follows a rigid script: click here, copy this, paste there. If anything changes, it breaks. Agentic process automation uses an AI agent that reads the inputs, figures out what to do, and adapts when things are different. RPA is more predictable and cheaper to maintain when the process is stable. APA is better when the process changes often or involves messy, varied inputs. Most teams will end up running both.
Will RPA be replaced by AI?
Not any time soon. RPA isn’t dying. It’s being absorbed. The bots that work will keep running. The RPA market is still growing at 18% per year. What’s changing is that new automation work increasingly goes to AI agents instead of scripted bots. Forrester predicts that RPA, BPM, and intelligent automation tools are all converging into one category. Think of it as RPA getting smarter, not getting replaced.
What is agentic process automation in A360?
A360 is Automation Anywhere’s platform. Their version of APA combines traditional RPA bots with AI agents that can handle more complex, variable tasks. The platform includes what they call a “Process Reasoning Engine” that predicts workflow steps. Whether it qualifies as truly agentic (as opposed to an AI layer on existing bots) depends on the specific use case. Ask for a demo on a process the system hasn’t been pre-trained on.
How much does agentic process automation cost?
There’s no standard pricing yet. Traditional RPA runs $5,000 to $50,000 per bot per year, depending on the platform. APA pricing is all over the place because the category is new and vendors are still figuring out their models. Some charge per task, some per agent, some per outcome. Budget for experimentation (a pilot on one process) rather than a full deployment. Deloitte’s 2026 survey found that only 25% of companies have moved even 40% of their AI pilots into production, which tells you how early this market still is.