# What to automate first: choosing processes by payback

URL: https://casterconsulting.hu/en/what-to-automate-first/
Published: 2026-10-15
Updated: 2026-10-15
Author: Caster Consulting

> Automate first what is frequent, rule-based, time-consuming and done today by someone copying data between systems. Don't start with the most impressive process; start with the one whose payback you can show a month later.

**Short answer: automate first what is frequent, rule-based, time-consuming and done today by someone copying data between systems.** Don't start with the most impressive process, or with the one that annoys the boss most. Start with the one whose payback you can show a month later. A successful first automation isn't about big savings; it's about the company learning how to do it.

In most SMEs the problem isn't a lack of things to automate. It's too many options, and picking the wrong one first. After a badly chosen first project the team learns that "automation doesn't work here", and the topic disappears from the agenda for years.

## What experience says

Some automation projects fail, and the reasons are surprisingly consistent. An early but still widely quoted EY analysis found that 30 to 50% of initial RPA projects (process automation with software robots) fail, and one of the main causes is choosing the wrong process. The same analysis suggests automating roughly 70% of a low-value process in the first round and leaving the remaining exceptions to people.

In Deloitte's 2019 survey, companies still piloting expected payback in 15 months on average, while those automating at scale reported payback in 9 months and a 27% cost reduction. In the same survey, companies with well-defined processes raised back-office capacity by 19% through automation, against 12% for the rest.

The two lessons are simple: **a well-chosen, well-documented process is half the success**, and **payback speeds up with experience.**

## The "long tail": why you don't have to automate everything

Wil van der Aalst, one of the founders of process mining, describes work with a "long tail" curve. 80% of cases come from 20% of case types. Those frequent, repetitive cases are worth automating with classic automation. In the middle of the curve, with rarer but still rule-based cases, software robots or AI come into play. The rare, one-off cases at the end of the tail should stay with people, because automating them costs more than it returns.

That's a liberating idea: **you don't need to automate 100% of a process.** If the frequent cases run on their own and people handle the exceptions, that's already a big result.

## A simple scoring method

Collect 10 to 15 candidate processes with your team and score each from 1 to 5 on five criteria:

1. **Frequency:** how often does it happen per week? (many times a day = 5, once a month = 1)
2. **Time:** how much work time does it take in total? (many hours a week = 5)
3. **Rules:** can it be described with clear rules? (fully = 5, lots of judgement = 1)
4. **Data availability:** are the inputs available digitally and in structured form? (in a system = 5, on paper = 1)
5. **Cost of errors:** how expensive is it when it goes wrong today? (very = 5)

Add up the scores and look at the top three. Then ask two control questions:

- **Can the current state be measured?** If you don't know how long it takes today, you can't show the improvement.
- **Does it have an owner?** Who knows the process and will be glad to see it automated?

The process that answers yes to both and scores high is a good first candidate. Keep the scored list; it becomes your roadmap for the next steps.

## Typical good first processes in an SME

- **Copying data between systems:** web shop orders into the business system, CRM data into invoicing, a spreadsheet into the ERP.
- **Capturing incoming documents:** extracting and recording data from delivery notes, orders and simpler invoices, with a human check.
- **Routine notifications:** deadline reminders, payment reminders, status updates to customers.
- **Building reports:** the weekly sales or stock report someone assembles from several sources today.
- **Monitoring:** tracking prices, stock levels, competitors or deadlines, and flagging deviations.

What they have in common: they're frequent, rule-based, and today a person acts as a "human interface" between two systems.

## Examples by department

| Area | Typical first automation | What it delivers |
|---|---|---|
| Sales | logging enquiries into the CRM, follow-up emails | no lost leads, faster replies |
| Finance | extracting incoming invoice data, matching it to purchase orders | less typing, fewer errors |
| Logistics | passing web shop orders to the warehouse system | no double entry, faster shipping |
| HR | leave approvals, onboarding checklists | a transparent process, fewer emails |
| Marketing | segmentation and campaign sending by rules | regular, personalised communication |
| Management | weekly report assembled automatically from several sources | up-to-date numbers without manual work |

The table isn't a mandatory list but a source of ideas. In your company, the scoring decides what comes first.

## Which tool?

Automation isn't a single technology. There are four main tool groups, and the process decides which one you need:

- **integration:** two systems talk to each other directly through an API. This is the most stable solution when both systems support it;
- **workflow tools and no-code platforms:** processes built from ready-made blocks with little or no code, including approvals and notifications;
- **software robots (RPA):** they click and copy in the user interface in place of a person. Good when there's no API, but more fragile, because they're sensitive to interface changes;
- **AI:** understanding text, documents or images, for example extracting data from an email. Often built into one of the three above as a step.

The most common mistake is choosing a tool first and then looking for a process to use it on. Do it the other way round: process first, tool second.

## What not to automate first

- **Rare processes,** however complicated and irritating. Four times a year isn't worth it.
- **Processes that aren't written down,** where everyone does it differently. Standardise first, then automate.
- **Processes full of judgement,** such as one-off quotes or difficult complaints.
- **Irreversible actions** before you trust the tool: payments, contracts.
- **Anything about to change:** if a new system is coming next month, wait for it.

## What we've seen

Earlier projects of our team show the pattern well.

- **Competitor price monitoring:** at a home furnishings retail chain, three employees monitored competitors' prices and promotions manually. We built an AI-based system on a no-code platform that tracks the prices, product data and promotions of 20 web shops and suggests prices. It was a frequent, rule-based, measurable task, so it went live in four months.
- **Replacing SAP Emarsys:** at an electronics retail chain operating in two countries, a large, complex SAP Emarsys setup was replaced in five months by an in-house marketing system we developed. Campaign management, segmentation and customer communication now run automatically and at lower cost.
- **Mobile audits:** at an international animal health company, paper audits were typed into spreadsheets afterwards. Replacing paper made that after-the-fact data entry unnecessary.

In all three cases the starting point was a visible, recurring burden, not an abstract "digitalisation strategy".

## How to involve your team

The people doing the work, not management, know the best candidates. They know which task they start every morning with a sigh, and where mistakes usually creep in. A one-hour session where everyone names their three most time-consuming routine tasks is worth more than a month-long consulting assessment.

How you communicate matters too. When people hear "we're automating", many hear "we're cutting jobs". Say up front what the team will do with the time freed up: customers, quality, development. If people see that automation takes away their boring task, not their job, they become its strongest supporters.

## Where companies stand

According to Eurostat, 19.95% of EU enterprises used AI in 2025, with a wide gap between small companies (17%) and large ones (55.03%), and lack of expertise as the main barrier, cited by 70.89% of companies that considered AI. In other words, smaller companies aren't held back mainly by budget but by know-how and by not knowing where to start. That gap is also an opportunity: if most of your competitors haven't started yet, well-chosen first automations still give a tangible edge. The scoring method above is exactly the kind of starting point that lack of expertise usually blocks.

## How to measure payback

1. **Measure the starting state** for a week or two: number of cases, time, errors.
2. **Work out the full cost:** development or setup, licences, running costs, and the time needed to handle exceptions.
3. **After go-live, measure again** with the same indicators, after one month and after three.
4. **Look beyond time:** fewer errors, faster turnaround and happier staff are results too.
5. **Write the result down briefly** and show it to the team. A well-documented first success is the best argument for the second automation.

## What to do now

1. **Collect 10 to 15 candidate processes** with the team in a one-hour session.
2. **Score them** on the five criteria.
3. **Pick one from the top three** that is measurable and has an owner.
4. **Write down the process,** including common exceptions.
5. **Automate the frequent cases** and leave exceptions to people.
6. **Measure after a month,** and only then move to the next process.

If you'd like us to tell you, after a short assessment, which automation is the best first step for your company, have a look at our [process automation](/en/process-automation/) service.

## FAQ

### Which process should we automate first?

The one that is frequent, rule-based, time-consuming, works with digitally available data, and today involves someone copying between systems. The current state must be measurable, and the process needs an owner.

### How long does automation take to pay back?

In Deloitte's survey, companies still piloting averaged 15 months, while experienced companies automating at scale reached 9 months. With a well-chosen, narrow automation in an SME, the effect often shows within the first months.

### Why do automation projects fail?

Most often because the wrong process is chosen: a rare one, an undocumented one, or one that needs a lot of judgement. EY's analysis found that 30 to 50% of initial RPA projects fail.

### Do we have to automate the whole process?

No. Automate the frequent cases and leave rare, one-off exceptions to people. Automating around 70% of a process in the first round is already a significant result.

### What's the difference between RPA and AI?

RPA (software robots) performs routine steps in the user interface according to rules, for example copying data between two systems. AI helps understand text or documents, for example extracting data from an email. The two often work together.

## Sources

- [van der Aalst, Bichler, Heinzl – Robotic Process Automation (Business & Information Systems Engineering)](https://www.vdaalst.com/publications/p980.pdf)
- [Deloitte Insights – Automation with intelligence](https://www.deloitte.com/content/dam/assets-shared/legacy/docs/insights/2022/DI_DR26-Automation-with-intelligence.pdf)
- [EY – Get ready for robots](https://www.eyfinancialservicesthoughtgallery.ie/wp-content/uploads/2016/11/ey-get-ready-for-robots.pdf)
- [Process Excellence Network – The role of process mining and RPA in helping organizations pick automation battles](https://www.processexcellencenetwork.com/process-mining/articles/the-role-of-process-mining-and-rpa-in-helping-organizations-pick-automation-battles/amp)
- [Eurostat – Use of artificial intelligence in enterprises](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises)