Score on 5 Factors: RPA vs AI Automation for Leaders
TL;DR
RPA repeats fixed steps on stable, structured work. AI reads messy input and makes a judgement call.
Score a process from 1 to 5 on input structure, judgement, exceptions and rate of change. 4 to 8 points is RPA, 9 to 14 is hybrid, 15 to 20 is AI.
Most real processes land in the hybrid middle. AI reads and decides, and fixed code writes to your systems.
RPA is cheap per run and expensive to maintain. AI is close to the reverse, so cost it over three years.
Pilot with a test set of 50 to 100 real past cases and one named owner before you scale.
Pick RPA for steady, predictable work. RPA means software bots that repeat the steps a person defines. Pick AI for messy work that needs judgement. Most real processes need both, working together. That hybrid setup is now the standard way to build automation that lasts. Below is a five-factor score to pick the right one, and a checklist for the pilot.
RPA vs AI automation: what RPA actually is
RPA stands for robotic process automation. Think of it as a very fast, very obedient new hire.
As IBM puts it, RPA bots can follow only the processes a person has defined. Click here, type this, copy that. It never learns and it never guesses. It repeats the same steps, every time.
That's why it's good at boring, stable jobs.
Where RPA wins:
- Data entry between two systems
- Reconciling numbers in spreadsheets
- Moving data in and out of old systems with no API
- Any task where the steps never change
For the sales-team version of the first job on that list, see Njin's guide to automated CRM data entry.
RPA is cheap to run once it's built. If your inputs stay the same, the returns come fast.
Here's the catch. If someone redesigns the screen the bot clicks on, the bot breaks. Fixing bots turns into a job nobody budgeted for. Ownership falls through the cracks, and that's a big reason RPA projects stall before they scale.
Pro tip: Before you build a bot, ask: "Will this screen look the same in 12 months?" If you're not sure, don't build it yet.
What is AI automation and when does it win?
AI automation doesn't follow fixed steps. It reads, weighs things up and makes a call.
UiPath's own comparison puts AI on processes that rely on unstructured data: documents, images, emails. It uses machine learning and language models to work out what something means, then decides what to do.
Where AI automation wins:
- Reading contracts or invoices with no fixed layout
- Sorting support tickets by urgency and topic
- Summarising long documents
- Spotting patterns people miss
The upside is flexibility. When the rules change, you update the instructions instead of rebuilding a bot from scratch.
The downside is cost and risk. Every AI decision has a running cost. And AI can be confidently wrong. That's called hallucination: it states something false as if it were true. Codelevate's guide cites Gartner's forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027, mostly for cost and scoping reasons. Build a test set of past cases and check accuracy before you trust it with real work.
The two were never rivals. Gartner named hyperautomation the number one trend in its top 10 strategic technology trends for 2020, and UiPath describes RPA as its foundation. RPA plus AI has been the plan for years.
How do you score a process to choose the right tech?
Don't guess. Score it. It takes the guesswork out and stops you buying tech you don't need.
This scoring comes from Codelevate's guide. Rate your process from 1 to 5 on four factors, where 1 is steady and predictable and 5 is messy and changeable:
- Input structure. Does every case arrive in the same shape, or is it different each time?
- Judgement per case. Does a person need to think, or just follow steps?
- Exception rate. How often does something odd happen?
- Rate of change. How often do the rules or systems change?
Add the four numbers.
- 4 to 8: go RPA. Stable, rule-based, little judgement. A bot handles it.
- 9 to 14: go hybrid. Some judgement, but the action itself is simple.
- 15 to 20: go AI-led. High judgement, high variation. This needs a brain, not just hands.
The fifth factor: volume
Volume doesn't change which tool can do the job. It changes which one is cheapest. A stable process that runs thousands of times a day favours RPA, because it's cheap per run. A messy process that runs a few times a week can carry AI's running cost without much pain.
A worked example. Matching invoices against a fixed set of suppliers usually scores low, so RPA wins. Reading supplier emails to work out what they want, then updating a record, usually lands in the hybrid band. Triaging free-text customer complaints scores high, so AI does the reading.
Codelevate's finding is that most real processes land in the hybrid middle, not at either end.
Pro tip: Cost it over three years, not one. RPA upkeep creeps up. AI running costs grow with volume. Do the maths before you commit budget.
The hybrid pattern that holds up in real use
Here's the setup that holds up once it's live, not just in a demo.
AI reads and decides. A separate block of fixed code writes the result to your systems.
Think of a kitchen. The AI is the chef, working out what to cook from a messy order. The fixed code is the line cook, plating the same dish the same way every time. You don't want the chef reinventing the plating every night.
This split matters. The judgement part stays separate from the part you can't undo.
Controls that make it safe:
- Set a confidence threshold. Below it, a person reviews before anything happens.
- Log every input and every reason the AI gave for its decision.
- Watch for drift, when accuracy slips quietly over time.
- Keep irreversible actions in fixed code. Payments, deletions and contract changes are never left to AI judgement alone.
Build a test set of real past cases. Measure accuracy against it before go-live, and keep measuring after. Our guide to human-in-the-loop workflows covers where the human checkpoint should sit.
Pro tip: Never let AI write straight to your system of record without a human checkpoint for low-confidence cases. That one rule prevents most of the expensive mistakes.
What do RPA and AI cost to run?
RPA costs more up front. You pay for licences and setup.
Running costs stay low until your systems change. Then upkeep climbs fast.
AI flips this. Setup is often cheaper, but every case you process has a running cost, and that adds up with volume.
A quick way to think about it over three years:
- High volume, stable process: RPA usually wins on total cost.
- Lower volume, messy inputs: AI's running cost is worth it for the accuracy.
- Most processes: hybrid. AI handles the small, hard part and RPA or code handles the big, easy part.
Timelines differ too. Enterprise RPA rollouts often take months because of testing and change control. AI pilots move faster, weeks rather than months, but need proper governance before you trust them at scale.
Phase it. Pilot small, prove the numbers, then expand.
How do you run a pilot without breaking things?
Follow this order. Don't skip steps.
- Map the process. Write down every step exactly as it happens today.
- Score it. Use the five factors above. Be honest.
- Pick a pilot with clear numbers. Set a target for accuracy, time saved and cost per case. Write it down before you start.
- Build a test set first. Pull 50 to 100 real past cases. Set your confidence threshold from them, not from a guess.
- Name an owner. One person watches it, gets the alert when it breaks, and fixes it. No owner, no future.
- Match the tool to the job. If an API exists, use integration code. If it's an old screen with no API, use RPA. Add AI only for the messy bits. Here's more on how AI fits with the tools you already run.
Pro tip: Start with the smallest process that still proves the model works. Big first projects are where automation budgets go to die.
Where Claude Code fits in a hybrid build
The hybrid pattern maps straight onto how we build.
The reading and deciding layer is an AI employee, built with Claude Code. It's trained on how the founder makes the call: which supplier emails matter, which complaints need a person, which exceptions are fine. That judgement is the founder's IP. Encoding it is what lets the work run without the founder in every loop.
The writing layer stays deterministic. Where a system has an API, that's plain integration code, and Claude Code writes that too. Where it's an old screen with no API, RPA still has a job.
That's the difference between orchestration and automation. Automation runs the steps. Orchestration runs the judgement calls the founder used to make, and hands the steps to the right tool.
Our 90-day program works through this as Explore, Map, Transform. The scoring above is the kind of work the Map phase is for. Transform ends in a working prototype, not a slide deck. For a look at what those builds involve, see custom AI delivery systems built with Claude Code.
My take: the biggest win I built had no AI in it
The highest-ROI system I've built for a client had no AI in it at all.
It was staff onboarding for an NDIS agency. They were always hiring support workers. Someone vetted each prospect and sent a form. Then someone keyed the details into an employment contract by hand. Once it was signed, a finance person added the new hire to Xero, an HR person added them to the HR system, and a tech person set them up in Office 365.
Three people, three platforms, all by hand. Typos in names everywhere. Some people ended up with two email accounts.
The fix was a plain automation. Mark someone as a hire and the form goes out. The form fills the contract. The signed contract fills all three platforms. Hours of work and errors, down to minutes.
Force AI into that flow and it would have wanted to rewrite the contract until it wasn't legal any more. That's the hybrid rule from above, learnt the practical way. Irreversible writes belong in fixed code.
The same shape turned up at a media company. Two full-time staff spent their year building custom invoices by hand, about $200k a year in people-time, because every client needed one. A partner in my agency chipped away at it for six months and built around 200 pieces of unique logic. There wasn't much AI in it. It was process automation. Both people moved onto revenue work, and nobody lost their job.
Work that looks too custom to automate is usually a process nobody has mapped yet.
So shop for tools last. Score the process first, and prove it by hand before you automate anything.
James Killick
Want a second opinion on your automation plan?
Most businesses don't have an automation problem. They have a "we never scored the process" problem.
The AI Orchestrators is the alternative to guessing your way through an automation budget. Instead of buying tools first and hoping, you get a map of your real processes, a working prototype built on them, and a scaling roadmap that shows where RPA, AI or hybrid fits.
It's built for founder-led consulting and education businesses doing $1M or more, with proven methods that need to run without the founder in every decision. Curious about the terms first? Check the orchestration glossary.
If that's you, start here: take the assessment. It shows how monetisable your IP is, and whether a build like this is your next step.
Sources
Frequently Asked Questions
James Killick
Founder
The AI Orchestrator. 10+ years building digital products and 200+ apps shipped, now helping $1M+ educators and consultants turn their IP into AI-powered delivery systems.
James Killick founded and runs The AI Orchestrators.
More from James Killick