RPA, an AI Agent, or Keep a Person on It?
Traditional RPA follows fixed rules through fixed screens. An AI agent reads unstructured input and makes bounded decisions. Some processes are better left with people. The right answer depends on the shape of the process, not on which technology is fashionable. Eight questions characterise the process and tell you which of the three fits.
Answer for one specific process, as it runs today. If you are choosing between several, run the quiz once for each and compare the results.
AI Agent vs RPA Decision Quiz
Eight questions about one process, as it runs today.
A structured self-assessment for one process. It runs in your browser and nothing you answer is recorded or transmitted.
The Three Results
Each result describes a kind of process and the kind of automation, or non-automation, that suits it. The score places your process on a single scale from rigid and repetitive to variable and judgement-heavy.
Traditional RPA
Structured input that looks the same every time, rules that have not changed in years, few exceptions, no judgement. RPA scripts the exact clicks and keystrokes a person would make, so it is fast to build for this shape and brittle for any other. Screen-based legacy systems with no API are its natural home.
AI agent
Input that varies in layout or arrives as free text, rules that are mostly stable but need interpretation at the edges, a meaningful share of exceptions, decisions a competent person makes from guidelines. An AI agent reads the input, works out what it is, applies the rules and asks a person when it is unsure.
Stay with people
Rules that change case by case, exceptions that are the norm, judgement that carries legal, safety or financial accountability, or volume too low to justify any build. Automating this shape produces an expensive system that hands most items back to a person anyway. Fix the process first, or leave it.
How the Score Works
Every question scores 0 to 3 on the same axis: how much variability and judgement the process demands. The total is expressed as a percentage of the maximum of 24.
Score each answer 0 to 3
A 0 describes a rigid, repetitive, structured process. A 3 describes one that is variable, judgement-heavy or low volume. Every question uses the same direction.
Add them up
Eight questions, so the maximum is 24. The result is shown as a score out of 24 and as a percentage.
Read the band
Under 25 per cent suits traditional RPA. From 25 to 49 per cent suits an AI agent. 50 per cent and above should stay with people until the process is redesigned.
Check the outliers
A single 3 in an otherwise low score is worth a look on its own. Very low volume or irreversible consequences can rule out automation even when everything else fits.
The Six Characteristics That Decide It
The quiz scores these directly. Understanding why each one matters lets you sanity-check the result against your own knowledge of the process.
Input structure
RPA needs the same fields in the same place every time, because it reads positions, not meaning. A supplier who changes their invoice layout breaks an RPA bot and does not trouble an AI agent, which reads the document the way a person would. Free text emails, varying PDFs and photographs of forms are AI territory. Fixed-format files and consistent system screens are RPA territory.
- Same fields, same positions, every time: RPA
- Same information, different layouts: AI agent
- Free text that has to be understood: AI agent
- Information that has to be chased from a person: not yet automatable
Rule stability and exceptions
RPA encodes rules as fixed logic. Every rule change is a code change, and every exception is a branch that has to be built. If the rules are stable and exceptions are rare, that is cheap. If exceptions are a fifth of the volume, the branches never end. An AI agent handles bounded variation within guidelines and routes what it cannot resolve to a person, which is a better fit when the rule book has edges.
- Written rules unchanged for years, exceptions under one in twenty: RPA
- Stable rules with interpretation at the edges: AI agent
- Rules that change monthly: fix the process before automating
- Exceptions above half the volume: there is no rule to automate
Judgement and consequences
Neither technology should make decisions that carry legal, safety or significant financial accountability without a person approving them. RPA cannot make judgements at all. An AI agent can make the kind of bounded judgement a trained staff member makes from guidelines, such as coding an expense or classifying an enquiry, provided the cost of a wrong call is recoverable and a person reviews the uncertain ones.
- No judgement, rule decides everything: RPA
- Guideline-based judgement with recoverable errors: AI agent with review
- Expert judgement or irreversible outcomes: person decides, automation prepares
- Make the approval step explicit in the design, not an afterthought
Systems and volume
RPA exists because some systems have no API, and driving the screen is the only way in. Where APIs exist, an AI agent, or plain integration, is more reliable than screen automation. Volume decides whether either is worth building. A process run a handful of times a month rarely justifies a build, whatever its shape, and the honest answer is to leave it with a person and a good checklist.
- Legacy screens with no API: RPA is often the only option
- APIs available: integration or AI agent, not screen scraping
- No system at all, paper and phone: redesign before automating
- Fewer than a few dozen items a month: usually stays manual
Next Steps
AI Automation vs RPA
The longer comparison of the two approaches, with where each one breaks.
Read the comparison →What Is an AI Agent?
A plain explanation of what an agent does, how it decides, and what it hands to people.
Read the explainer →Workflow Automation Audit Checklist
Map the process properly before committing to either approach.
Map the process →Frequently Asked Questions
What is the difference between RPA and an AI agent?
RPA, robotic process automation, records and replays the exact actions a person takes on screen: open this window, click this field, type this value. It follows fixed rules and fails when the screen or the input changes. An AI agent uses a language model to read input, work out what it means, decide what to do within the rules it has been given, and act through system integrations. It tolerates variation in input and can make bounded decisions, and it hands uncertain cases to a person. RPA is deterministic and brittle. An agent is adaptable and needs review.
Why does a low score mean RPA rather than AI?
Because the scale measures how much variability and judgement the process demands, and RPA suits the least variable processes. A low score describes structured input, fixed rules, few exceptions and no judgement. That is exactly what RPA was built for, and it is usually cheaper to build and easier to verify for that shape. Using an AI agent on a process that needs none of its adaptability adds cost and non-determinism for no gain.
Can a process be split between the three?
Often, and it is usually the best design. A typical pattern is an AI agent reading and classifying incoming documents, a rules-based step doing the fixed transformation, and a person approving anything above a threshold or below a confidence level. Run the quiz on the whole process first. If it lands in the middle or high band, look at whether the judgement-heavy part can be separated and left with a person while the rest is automated.
My process scored as stay with people. Is that final?
No. It describes the process as it runs today. Most processes in that band got there because rules were never written down, exceptions were never counted, or information arrives by phone instead of through a system. Fixing those things, which is worth doing anyway, often moves the process into the AI agent band. The workflow automation audit checklist on this site is the tool for that work.
Does the quiz consider cost?
Only through volume. It tells you which approach fits the shape of the process, not whether the build is worth paying for. Once you know the approach, the automation payback period calculator on this site takes a setup cost, a monthly fee and the hours saved and tells you when it breaks even. A process can fit an AI agent perfectly and still not be worth automating at low volume.
What about workflow tools like Zapier or Power Automate?
They sit between the two. Trigger-and-action workflow tools connect systems through APIs and follow fixed logic, so they suit the RPA end of the scale for processes whose systems have APIs. They do not read unstructured input or make decisions on their own, though several can call an AI model as a step. If your process scores in the RPA band and every system involved has an API, a workflow tool is usually the cheaper and more reliable choice over screen-based RPA.
Sources and further reading
- Zapier (Zapier)
Got Your Result?
Tell us the process and the band it landed in. We will confirm or challenge the result, say which parts belong with a person, and quote the approach that fits. If the honest answer is to leave it manual, we will say so.