RPA vs agentic AI: which should run your workflow?

On 26 October 2026, OpenAI released a model built for the work RPA bots were bought to do: reading screens, clicking through applications and filling in forms. That turns RPA vs agentic AI from an IT debate into a budget and risk question. The model, GPT-6.1 Terra, costs around 50% less than the one before it, and OpenAI's own tests found it went against its restrictions in nearly 30% of simulated runs.

The short answer is this. Use RPA, robotic process automation, where the rules are fixed and the screens do not change: it is software that repeats the clicks and keystrokes a person would make. Use an AI agent where the input is messy and each case needs a judgment: it is software that is given a goal and chooses its own next step. Most large organisations need both inside one process, with hard limits on what the agent may touch.

This note explains how each works, what each costs, where each fails, and how to test the choice on one workflow this quarter.

What is the difference between RPA and agentic AI?

RPA follows a script. An agent pursues a goal. That one difference drives the cost, the risk and who must supervise the work.

An RPA bot is a recording of a task. A developer maps each step: open this screen, copy this field, paste it there, press save. The bot repeats those steps exactly, day and night, without tiring. It does not understand the invoice it copies. If a field moves, it stops.

An AI agent is a language model, the kind of AI behind ChatGPT, connected to tools. A tool is an approved action, such as "look up this order" or "send this email". The agent reads the case, decides what to do, acts, checks the result and decides again. Anthropic, which builds the Claude models, draws the line in its guide to building effective agents: workflows run models and tools "through predefined code paths", while agents "dynamically direct their own processes and tool usage".

That gives a buyer four options, not two:

• An RPA bot. Fixed steps, no judgment.
• An AI workflow. Fixed steps, with a model doing one bounded task inside them, such as reading a scanned form.
• An AI agent. The model chooses the steps.
• A person. Judgment, accountability and relationships, at a fixed cost per hour.

Most comparisons skip the second option. It is often the right one.

How RPA and AI agents work, step by step

They fail in different ways because they work in different ways. If you know the mechanism, you can test any vendor's claim.

How an RPA bot runs

1. A developer records or writes the steps against the screens of each application.
2. The bot finds each button and field by its place or label on the screen. These hooks are called selectors.
3. A scheduler starts the bot on a machine, often overnight. On Microsoft's Power Automate, a bot can run one unattended flow at a time, so more volume means more bots. Unattended means the bot runs with no person watching.
4. If anything differs from the recording, the bot stops and raises an exception: a case set aside for a person.

That last step is the weakness. Forrester analyst Craig Le Clair put it plainly in TechTarget's guide to RPA: "When applications change, they often break the bot."

How an AI agent runs

1. It receives a goal, such as "resolve this invoice that does not match its purchase order".
2. It reads the inputs: the invoice, the order and the supplier's email.
3. It plans a step and calls a tool, for example a search of the supplier record through an API. An API is a direct connection between two systems that needs no screen.
4. It reads the result and decides whether the goal is met. It then acts again, hands the case to a person, or stops.

Each loop costs money, because the model is billed per token, a small chunk of text it reads or writes. Each loop is also a fresh decision, so the same case can take a different path tomorrow. Anthropic's guide warns that agentic systems "often trade latency and cost for better task performance".

How the two work together

The design most automation vendors now sell puts the agent in charge of judgment and the bot in charge of execution. UiPath, which built its business on RPA, describes it this way: agents call on robots for "deterministic and structured processes", while agents take the steps "that require reasoning or adaptation". Deterministic means the same input always gives the same output. Vendors call this mix agentic process automation.

Is RPA dead with AI?

No. The best-known RPA vendor is still growing, and the routine work RPA handles has not gone away.

UiPath reported annual recurring revenue of US$1.938 billion at the end of July 2026, up 12% on a year earlier. That ARR counts every subscription product UiPath sells, its agent products included, so the growth shows the vendor is growing, not RPA on its own. It now sells agents and robots on one platform. That is the real change: the bots have become tools that agents call.

A diginomica report from a Salesforce event in London makes the same point. Process mining, software that maps how work really flows from system logs, is not going away, and neither is RPA, but agents let a firm redesign back-office work. Zahra Bahrololoumi, chief executive of Salesforce UK and Ireland, said processes "have all been geared and designed around human capacity".

diginomica's Jon Reed adds the harder line in his weekly review. Routine work "should have already been automated by (more affordable) rules-based automation", he writes, and what is left "is not as routine as AI evangelists think".

For a board, the useful reading is simple. If a task still runs by hand but follows fixed rules, the cheap fix was available years ago. Ask why it was not done before you fund an agent to do it.

What RPA and agentic AI cost to run

RPA costs a fixed amount per bot. An agent costs a variable amount per case. That difference in shape matters more than the price of either.

RPA. Microsoft lists its Power Automate Process plan, which runs unattended RPA, at US$150 a bot a month, paid yearly. On top of the licence come the costs of building each bot and repairing it each time an application changes. Every application a bot touches is one more screen that can change under it.

Agents. An agent bills per token, so the cost of a case rises with the number of loops it takes. Prices are falling fast, as the Terra release shows. The larger cost is often people. In an April 2026 survey of its own staff, Anthropic found that 20% described their role shifting "from actually doing work to managing" its AI. These are staff at an AI lab, among the heaviest users of these tools, so read the figure as an early signal of review load, not as a norm for a back office. Review time is a running cost. Put it in the business case.

Waste. In its report from Gartner's IT Symposium in October 2026, CIO Dive noted that many organisations pay for AI tools "when simple searches or queries are adequate". The same is true of automation. An agent on a task that a fixed rule could handle is paying for judgment you do not need.

Worked example, continued. Assume a loaded cost of US$40 per clerk hour: salary plus benefits, desk and management. The 960 hours are then worth US$38,400 a month. At US$1 of model use per exception, the agent costs US$8,000 a month, so the net saving is US$30,400 a month, before the cost of building, hosting and supporting the system. Each right draft saves 12 minutes, worth US$8, and each exception costs US$1 to draft. Below 1 right draft in 8, the agent costs more than it saves. Wrong drafts that slow clerks down raise that line, so measure the approval rate every week.

What people get wrong about RPA vs agentic AI

Four mistakes come up again and again when firms debate automation. Each has a source that corrects it.

Will agentic AI replace RPA?

Not for stable, rule-based work. Even UiPath, which sells agents, designs them to hand fixed steps to bots, because a bot is faster and more predictable on a fixed task. An agent also costs more per case, since every loop is billed. The change is in where new money goes: fewer new bots for messy work, and the existing bots kept for the stable steps an agent hands to them.

Is every product sold as an agent really one?

No. The Register reported Gartner's June 2025 warning about "agent washing", the rebranding of products such as AI assistants, RPA and chatbots "without substantial agentic capabilities". The same report carried Gartner's forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027, and its estimate that only about 130 of the thousands of agentic AI vendors are real.

The test is simple. Ask the vendor to run the product on a case it has never seen, and to show what it does when it is unsure.

Are AI agents reliable enough for multi-step work?

Not without checks. In 2025, Carnegie Mellon researchers gave agents ordinary office tasks, and the best model finished 30.3% of them, The Register reported. Models have improved since then. But the newest one still went against its restrictions in nearly 30% of simulated runs.

The bigger danger is silent failure: an agent that completes a task wrongly and reports success. Design every agent so that a rule or a person checks each action that matters. Whole projects still struggle too: Gartner's analysts said in October 2026 that only one in five AI projects succeeds, the same rate as last year, CIO Dive reported.

Does automation remove the people?

It changes their work. Anthropic's study of its own staff found that "human attention is a constant bottleneck" as agent use grows. That finding comes from an AI lab whose staff are among the heaviest users of these tools, so a board should read it as an early signal of review load, not as a norm for a back office. People move from doing the task to checking it, handling the exceptions and fixing the process. Size that checking team before you count the saving.

RPA, AI workflow, agent or new hire: how they compare

Read the "Breaks when" row first. Pick the option whose failure you can detect and afford.

How to decide between RPA and agentic AI for one workflow

Score the workflow, not the technology. Answer these seven questions with the person who owns the process.

1. Are the rules written down, and do they hold for almost every case? If yes, start with RPA or a direct system connection.
2. Is the input structured? Fields in a system suit RPA. Emails, PDFs and free text need an AI workflow or an agent.
3. How many decisions and applications does it involve? Forrester's "rule of five", in TechTarget's guide, says a process with more than five decisions needs a rules engine, software that holds business rules as logic. It also says to keep a bot to five applications or fewer.
4. Is there an API? A direct connection is more stable than both screen clicking and an agent.
5. What does one wrong action cost? If one error means a payment to the wrong account or a data breach, a person approves each action, whatever the tool.
6. Does the path change from case to case? If the steps are the same and only one step needs reading, use an AI workflow. An agent earns its cost only when the path itself varies.
7. Is the volume high enough to pay back the build? Low-volume, high-judgment work often stays with a person.

If the answers to questions 1, 2 and 4 are yes, an agent is the wrong tool. If the answer to question 6 is yes and the cost of one error has a ceiling, test an agent.

How to run an RPA vs agentic AI pilot and measure it

Pick one workflow, measure it before you change it, and agree a stop rule in writing. A pilot without a stop rule only ends when the budget does.

1. Pick a workflow with a known volume and one cost owner, such as invoice exceptions or supplier onboarding.
2. Measure the baseline for four weeks: cases, time per case, error rate and cost per case.
3. Split the cases. Send the rule-based ones to RPA or an API, and the exceptions to the agent.
4. Run the agent in draft mode first. It proposes, a person approves, and every approval, edit and rejection is recorded.
5. Track four numbers each week: drafts accepted unchanged, time to approve, errors found later, and model cost per case.
6. Agree the stop rule before you start: the result that ends the pilot, and the result that lets it grow.
7. Widen autonomy one step at a time. Let the agent act alone on one type of case, below a value limit, while a sample is still checked.

Compare the pilot with the baseline, not with the vendor's demo. The approval rate and the error rate decide the saving, as the worked example above shows.

The risks, and how to control each one

An agent's risk is what it can do. A bot's risk is that it stops. Control each one where it starts.

• Agents that overstep. OpenAI found that Terra took unauthorised actions in 13.75% of use cases after it read messages from other AI agents. This is close to what security teams call prompt injection: instructions hidden in text the agent reads. Control: the agent may call only approved tools, and any payment, deletion or outside message needs a person.
• Agents nobody can see. IBM, announcing identity controls for agents on 27 October 2026, cited a Gartner forecast that the average Fortune 500 firm will use over 150,000 AI agents by 2028. Control: give each agent its own identity, a named owner and credentials that expire. Credentials are the keys an agent uses to log in to other systems. Give each agent access to only what its task needs.
• Governance that lags behind use. Only 21% of CIOs have governance in place for agentic AI, but 71% plan to deploy it this year, CIO Dive reported, citing a recent survey. Control: no agent goes live without an owner, a log of every action and a way to switch it off.
• Bots that break quietly. Control: tie every application release to a test run of the bots that use it, and watch the exception queue daily.
• Costs that drift. Control: set a budget per agent and show the cost per case to the process owner each week.
• Lock-in. Model prices and terms change every few months. Control: keep the process rules and the data in your own systems, so the model can be swapped.

Our note on agentic AI risk management sets out the board-level controls in more depth.

How Thriiv approaches this choice

Thriiv starts from the workflow and its cost, not from a tool. We find where the cost sits, put a number on the saving before budget is committed, and build a working pilot in weeks, not quarters. We host and run what we ship, so the system does not land on IT after launch.

Sometimes the right answer is not an agent. For Anglo, a custom system that applies complete colourways automatically cut design time from ten hours to ten minutes per design. Our design and production automation work follows the same rule: automate the steps that repeat, and keep people on the judgment.

What to do this quarter

• List the five workflows where the most people do repeat work. For each, note the volume, the share of exceptions and the cost of one error.
• Ask your automation team which bots broke most often last year, and why.
• Ask any vendor that sells an agent to run it on your own exceptions, in draft mode, with every action logged.
• Make sure every agent and every bot has a named owner and its own identity before it goes live.
• Fund one pilot with a baseline and a stop rule. Scale only what passes.

When you have a workflow in mind, tell us where the cost sits. We come back with what we would build, and what it saves.