The debate about RPA vs agentic automation is no longer theoretical. The largest RPA vendors have rebuilt their platforms around AI agents, Microsoft now lets agents operate desktop and web applications directly, and many businesses are asking whether their portfolio of brittle bots should be retired.
The short answer: some should, many should not, and most should be wrapped rather than replaced. Rule-based bots are still the cheapest, most predictable way to run a stable, high-volume, well-defined task. Agents earn their place where the work needs judgement, unstructured input or recovery from the unexpected.
This guide explains what has actually shipped, what Gartner calls "agent washing", and a practical framework for deciding what to do with each bot you run today.
What RPA was good at, and where it hurt
Robotic process automation mimics a person clicking through screens: log in, copy a value, paste it elsewhere, press save. It became popular because it automated legacy systems that had no APIs, without touching their code.
It worked well for:
- High-volume, rule-based tasks with stable screens, such as posting bank entries or downloading reports.
- Bridging two systems where no integration existed.
- Tasks with structured inputs, such as spreadsheets in a fixed format.
The costs showed up later. Bots broke whenever a screen changed, a pop-up appeared or a field moved. Maintenance teams grew. Exceptions, the cases that did not fit the rules, still landed on people. Many organisations found that a large share of RPA effort went into keeping existing bots alive rather than automating new work.
What "agentic automation" actually means
An AI agent is a system that uses a language model to decide what to do next, calls tools (APIs, databases, other bots), observes the result and continues until the goal is met or it hands off to a person. The key difference from RPA is that the steps are not fully scripted in advance.
What has shipped, with dates:
- UiPath launched its agentic automation platform with Maestro, an orchestration layer for agents, robots and people, on 30 April 2025, and expanded it with governance and testing tools in September 2025. It also added two-way integration with Microsoft Copilot Studio agents.
- Microsoft added computer use to Copilot Studio, letting an agent operate websites and desktop apps through the screen, much like a bot, but driven by a model that can adapt to layout changes.
- Open protocols such as the Model Context Protocol (MCP) have made it easier to expose business systems as tools any agent can call, which reduces the need for screen scraping in the first place.
The direction is clear: RPA platforms are becoming orchestration platforms where deterministic bots, AI agents and human approvals are steps in one process.
Separating real agents from agent washing
In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls. It also warned of "agent washing", vendors rebranding chatbots, assistants and RPA as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI offered genuine agentic capability.
Ask these questions of any product or proposal:
- Does it decide the next step itself, or follow a fixed script with a chat window on top?
- Which tools can it call, and who controls that list?
- What happens when a tool fails or returns something unexpected?
- Can you see a full trace of every decision and action?
- How is it evaluated before release, and how are regressions caught?
- What is the cost per completed task, not per seat?
RPA vs agentic automation compared
| Factor | Rule-based RPA | AI agents |
|---|---|---|
| Best input | Structured, consistent | Unstructured or variable |
| Behaviour | Deterministic | Probabilistic |
| Change in UI | Breaks | Often adapts |
| Cost per run | Very low | Higher (model calls) |
| Speed | Fast | Slower per step |
| Audit | Simple logs | Needs decision traces |
| Exceptions | Escalates everything | Can resolve some itself |
| Risk | Does the wrong thing consistently | Can do something unexpected |
Neither column wins outright. The design question is which steps of a process belong in which column.
A framework: keep, wrap, upgrade or replace
Take your bot inventory and place each one in one of four buckets.
- Keep: stable screens, structured input, low exception rate, cheap to run. Leave it alone. Replacing a working bot with an agent adds cost and risk for no gain.
- Wrap: the bot works but exceptions pile up on people. Put an agent in front to read emails or documents and prepare clean input, or behind it to triage failures. The bot still does the deterministic part.
- Upgrade to an API: the target system now has an API, or a connector exists. Replace screen-scraping with a direct integration. This is often the biggest reliability win and has nothing to do with AI.
- Replace with an agent: the task is mostly judgement, the input is unstructured, or the screens change constantly. Rebuild it as an agent with clear tools, guardrails and a human approval step for anything irreversible.
Where agents genuinely add value
Good candidates share three traits: messy input, a bounded set of actions, and a clear definition of done.
- Email and inbox triage: reading supplier, customer or TPA emails, classifying them, extracting data and creating the right ticket or record.
- Exception handling: investigating why an invoice did not match a PO and proposing a resolution for approval.
- Document-driven processes: onboarding, claims and KYC, where inputs vary widely. See our guide to intelligent document processing.
- Multi-system lookups: answering "where is this order?" by checking the ERP, the courier portal and the CRM.
- Legacy screens that change often: where a computer-use agent can tolerate layout drift better than a recorded bot.
For a deeper look at scoping a single agent, see AI agents for small business.
Guardrails for production agents
An agent that can click through your ERP can also click the wrong thing. Treat it like a new employee with system access:
- Least privilege: a dedicated service account with only the permissions the task needs.
- Allow-listed tools: the agent can only call actions you have defined, with validated parameters.
- Human approval for irreversible steps: payments, deletions, external emails and refunds.
- Rate and spend limits: caps on actions and model calls per hour.
- Full tracing: every prompt, tool call and result logged, searchable and retained per your policy.
- Evaluation sets: a fixed set of past cases the agent must pass before each release.
- Kill switch: one control to pause the agent and route work back to people.
These controls are what separate projects that survive from the 40% Gartner expects to be cancelled.
Migration plan for an existing RPA estate
- Inventory: list every bot, its process, volume, run cost, failure rate and maintenance hours.
- Mine the process: use logs or process mining to see real paths and exception rates. Our guide to process mining explains how.
- Bucket each bot: keep, wrap, upgrade to API, or replace.
- Pick two pilots: one wrap, one replace, in processes with measurable outcomes.
- Build the orchestration layer: a single place where bots, agents, APIs and human tasks are sequenced and logged. This can be a commercial platform or an open tool like n8n.
- Run in shadow mode: the agent proposes, people act, and you compare.
- Promote gradually: move to auto-execute for low-risk cases first.
Costs and what drives them
Indicative drivers rather than fixed prices:
- Licences: RPA platforms typically charge per bot or per run; agent features often add consumption-based charges.
- Model usage: cost per task depends on the number of steps, model size and how much context each step carries.
- Integration work: replacing screen scraping with APIs is usually the biggest one-off cost and the biggest long-term saving.
- Evaluation and monitoring: budget engineering time for test sets and tracing; it is not optional.
A useful comparison metric is cost per successfully completed case, including the human time spent on exceptions, measured before and after.
Common mistakes to avoid
- Replacing stable bots first: it feels modern but delivers little. Start where maintenance pain and exception volume are highest.
- Letting the agent touch production on day one: shadow mode exists for a reason; use it until error rates are known.
- No owner for the process: an agent needs a business owner who reviews its traces and decides on edge cases, just like a team lead.
- Measuring activity, not outcomes: "messages handled" or "steps executed" says nothing about whether the case was resolved correctly.
- Ignoring the people who did the work: the staff who handled exceptions know the edge cases better than any document. Involve them in designing the evaluation set.
How to start
- Export your bot list with maintenance hours and failure rates.
- Identify the five bots that cost the most to keep alive.
- Check whether each target system now has an API or connector.
- Choose one exception-heavy process and design an agent to triage exceptions, with human approval.
- Define success metrics up front: completion rate, error rate, cost per case, time to resolution.
- Run for 30 to 60 days, then decide whether to scale.
Algon builds this kind of hybrid automation, combining deterministic workflows, APIs and bounded agents. Our AI agent development and workflow automation pages describe the approach, and our application modernization team handles the API work that often makes bots unnecessary.
Conclusion
RPA is not dead, and agents are not magic. The useful question is not "RPA or agents?" but "which step of this process needs determinism, and which needs judgement?" Keep the bots that work, replace screen scraping with APIs wherever you can, and use agents for the messy, exception-heavy work that bots always pushed back to people. Measure everything, and insist on traces and tests before anything goes live.
Frequently asked questions
What is the difference between RPA and agentic automation?
RPA follows a fixed script to click through applications and works best on stable, structured tasks. Agentic automation uses an AI model to decide the next step, call tools and handle variable input, which suits exception-heavy or unstructured work.
Will AI agents replace RPA?
Not entirely. Rule-based bots remain cheaper and more predictable for stable, high-volume tasks. The trend is toward platforms that orchestrate bots, agents, APIs and human approvals together, with each used where it fits.
What is agent washing?
Agent washing is Gartner's term for vendors rebranding existing chatbots, assistants or RPA tools as AI agents without genuine agentic capabilities such as autonomous planning, tool use and recovery from errors.
Should I replace my UiPath or Power Automate bots?
Only where it pays off. Keep stable bots, replace screen scraping with APIs where available, wrap exception-heavy bots with an agent for triage, and rebuild only those processes that are mostly judgement or have constantly changing screens.
Why do agentic AI projects fail?
Gartner cites rising costs, unclear business value and inadequate risk controls. In practice, projects also fail without evaluation sets, decision tracing, limited permissions and human approval for irreversible actions.
How do I measure whether an agent is better than a bot?
Compare cost per successfully completed case, including human time on exceptions, along with error rate, completion rate and time to resolution, measured on the same process before and after.
