Business process automation: when you need AI and when you don't

Key Takeaways

• Business process automation means having a system carry out steps of a process that a person does today. If those steps follow fixed rules, you don't need AI. You need it when information that changes has to be read or interpreted.
• RPA can only follow the processes someone defined, while AI recognizes patterns in unstructured data (IBM). That's why they complement each other more than they compete.
• EY observed in its practice that 30% to 50% of initial RPA projects fail, and made clear it isn't because of the technology (EY, Get ready for robots, 2016).
• Process fragmentation was the top barrier to scaling automation in four consecutive Deloitte surveys (Automation with intelligence, 2022). Automating a messy process only makes it messy faster.
• Currently demonstrated technology could, in theory, automate about 57% of work hours in the United States, according to the McKinsey Global Institute (November 2025), which makes clear it's not a forecast of job losses.

Many companies come to AI asking about the wrong process. They want an agent to answer quote requests when the real problem is that inventory lives in three different Excel files. They want a model to read invoices when almost all their suppliers already send e-invoices in a format any simple integration can read.

AI is one way to automate, not the only one, and often not the cheapest. This guide explains how to tell when a process is solved with rules, when with RPA, when with AI, and when with a mix, and in what order to do it. It's part of our guide to artificial intelligence for businesses.

What is business process automation?

Business process automation is using software so that a set of steps in a business process happens without a person doing them by hand: moving data between systems, generating documents, sending notifications, validating information, or approving according to rules. The goal isn't to remove people from the process, but to take them out of the steps where they add no judgment.

There are four ways to automate, and they differ in the kind of work they can do:

Gartner uses the term hyperautomation for the disciplined approach of combining these technologies to rapidly identify, vet, and automate as many business and IT processes as possible. The word sounds big, but the idea is simple: use the simplest tool that solves each step.

What is RPA and how is it different from AI?

RPA (robotic process automation) is software that mimics a person's actions at a computer, like copying data, filling in forms, or moving files, following exact instructions. IBM puts it this way in its definition of RPA: RPA bots can follow only the processes defined by a user, while AI bots use machine learning to recognize patterns in data, particularly unstructured data.

In business terms: RPA is very good at doing exactly what you told it, and very bad when something changes. AI is good when information arrives differently each time, and it needs review because it can be wrong with great confidence.

That's why the useful question isn't "RPA or AI?" but which step of the process needs which. RPA vendors themselves, like UiPath, now present their robots as the layer that executes what an AI agent decides.

When is automation without AI enough?

Automation without AI is enough when information arrives structured (fields, codes, fixed formats) and the decision can be written as a rule. In those cases AI adds cost and error risk without adding anything.

Three signs you don't need AI:

• The data is already in a system. If the information lives in the ERP, the CRM, or a database, moving and matching it is integration work.
• The decision has a rule. "If the customer has receivables more than 60 days overdue, block the order" doesn't need a language model.
• The format doesn't change. E-invoices, a flat file from the bank, or a web form always arrive the same way.

When do you need AI to automate?

You need AI when a step in the process requires reading, interpreting, or drafting from information that changes every time. That's where rules and RPA break.

Most real processes mix several rows. A purchase order may arrive as a PDF (AI to read it), be matched against inventory (rules), be loaded into an old system with no connection (RPA), and notify the sales rep (rules). Automating well means assigning each step the simplest tool that solves it. If the process has steps where the path truly changes, the guide on what an AI agent is explains when one is justified.

Why do automation projects fail?

They fail because they automate a process nobody cleaned up first. Almost a decade ago, in Get ready for robots, EY wrote that it had seen as many as 30% to 50% of initial RPA projects fail, and that this wasn't a reflection of the technology but of planning. The same cause shows up today with AI.

Deloitte confirms it from another angle. In its Automation with intelligence 2022 survey, process fragmentation appeared as the top barrier to scaling automation for the fourth year in a row, and the payback period for organizations still piloting rose to 22 months, from 16 in 2020.

In the companies we work with, we see two mistakes that explain almost every failure. The first is automating the process as it is, with its unnecessary steps and duplicate approvals. The second is automating what's visible instead of what's costly: the report that annoys management instead of the data entry that eats hours every week in the admin team.

How to decide what to automate and with what, step by step

This is the order we follow with companies. It works the same whether the answer ends up being AI, RPA, or a simple integration.

1. Map the process as it works today

Draw the process with the people who do it, not with whoever designed it. Write down each step, who does it, what information they receive, which system it comes from, and how long it takes. The exceptions they tell you about matter as much as the normal path. If you've never done it, start with how to build your company's process map.

2. Eliminate and simplify before automating

Ask whether each step is needed. Double approvals, reports nobody reads, and data typed into two systems often disappear at this stage, and that alone is a saving with no technology.

3. Classify each step by the type of information it handles

Separate the steps with structured data and a clear rule from the ones that require reading, interpreting, or drafting. This classification decides the technology, not the vendor's catalog.

4. Assign each step the simplest tool that solves it

Use rules and integrations where they're enough, RPA where a system has no other way to connect, and AI only in the steps that require reading or writing. If you already pay for an ERP or an office suite, check first which automations it includes.

5. Keep a person on the exceptions

Define which cases leave the automated flow and who handles them. An automated process with no owner for exceptions ends up with errors nobody sees until the customer complains.

6. Measure hours and errors against the baseline

Compare the automated process with the measurement from step 1: hours per week, errors per month, response time. If the number hasn't changed after 60 days, revisit step 2 before blaming the tool.

Real case: automation prototypes in three days in Ciudad Juárez

In August 2026 we took our BuildInside method to Innova Lab V3.0 in Ciudad Juárez, with about forty people from eighteen companies: cement producers, maquilas, metalworking shops, retailers, and a bakery chain. None had a technology department or a software budget.

The workshop began with step 1 of this guide. Before opening any tool, each company completed a sentence with figures: when this happens, today we do that, it takes us this long and costs us this much. That sentence forced them to choose processes that really cost money, not the ones that looked best in a presentation.

What came out of it replaced manual work. A metalworking plant built a purchase order analyzer for the traceability it had been tracking by hand. A cement company built an analyzer that assigns each order to the distribution point closest to the customer. A tire distributor built an inventory analyzer that replaced hours of work in Excel. They were built by the people who live the problem, with generative AI as the tool.

Read how the workshop in Ciudad Juárez went.

Frequently asked questions about business process automation

Which processes can be automated in a company?

You can automate processes that repeat often and whose steps can be described, such as invoicing, reconciliations, order entry, periodic reports, request classification, and receivables follow-up. The best candidates are the ones that consume many hours of the team's time and have a result that's easy to check.

What's the difference between RPA and AI-powered automation?

RPA repeats exact on-screen actions following fixed instructions, while AI-powered automation adds steps that can read, classify, or draft from information that changes. RPA fails when the format changes. AI handles variation, but it needs review of doubtful cases.

Is RPA obsolete now that there's AI?

No. RPA is still useful for connecting old systems that have no other way to integrate, and many vendors now use it as the layer that executes what an AI system decides. What has changed is that it no longer makes sense to use RPA to read documents with variable formats, because AI does it better.

How much does it cost to automate a process?

It depends on the technology each step needs. Rule-based automations inside tools you already pay for may cost nothing more than the time to set them up. RPA and AI add licenses or usage costs, and all of them require the team's time to map the process and test. That's why it's worth first calculating what the process costs today and setting how much you're willing to invest to improve it.

Does business process automation eliminate jobs?

Automation changes tasks before it eliminates jobs. The McKinsey Global Institute estimated in 2025 that available technology could, in theory, automate about 57% of work hours in the United States, and made clear that the figure is not a forecast of job losses. In mid-sized companies, the most common outcome is that the team stops typing data and moves on to reviewing, handling exceptions, and talking to customers.

Automate the process that costs, with the simplest tool

The rule that sums up this guide is short: first clean up the process, then assign each step the simplest tool that solves it, and use AI only where something has to be read or written. Companies that follow that order spend less and see results sooner.

If you want to map a process in your company and walk away with a clear plan of what to automate and with what, at Suricata Labs we do it with your team and build the first solution with them.

Explore our AI services | Schedule a conversation

Read also: How to implement artificial intelligence in a business