AI automation means giving routine work to software that can read, decide within your rules, and act. Start with work that repeats often, follows a known pattern and is not too complex.
Every team has work that nobody should do by hand: the weekly export, the follow-up email, the report that is copied from three systems. The hard question for a mid-sized company is not whether to automate it. It is where to start. This guide gives you a way to choose.
It is written for leaders of companies with a few hundred to a few thousand people. You do not need a technical background.
Key takeaways
- AI automation is automation that can handle text, judgement within rules, and exceptions. Classic automation only follows fixed steps.
- Start with work that repeats often, follows a known pattern and is not too complex. A review of research on rule-based automation lists these among its criteria.[2]
- In the EU, 34.70% of companies that use AI use it for marketing or sales, and 31.05% for administration.[3]
- Fix a broken process before you automate it. Automation makes a bad process faster, not better.
What is AI automation?
AI automation is the use of AI to carry out work that used to need a person at each step. The International Society of Automation defines automation in general as “the creation and application of technology to monitor and control the production and delivery of products and services”.[1] AI adds one thing: the software can now read and write language, so it can handle work that was too messy for fixed rules.
For example, take a reminder for an unpaid invoice. Classic automation can send the same reminder to everyone once an invoice is overdue. AI automation can read the customer’s last three emails, see that a complaint is open, and draft a different message for a person to approve.
AI automation vs classic automation
Classic automation has existed for decades. Its best-known office form is robotic process automation, or RPA. IEEE Standard 2755-2017, as quoted in a review from Ulm University, describes RPA as software “that uses business rules and predefined activity choreography” to carry out tasks across systems.[2] The important words are “predefined” and “rules”. Someone has to write every step in advance.
| Classic automation | AI automation | |
|---|---|---|
| How it works | Follows steps that someone wrote down | Reads the situation and chooses the next step within your rules |
| What it can read | Fields and numbers in a fixed format | Emails, documents, notes and tables |
| When something unusual happens | It stops and waits for a person | It can handle many exceptions, and asks a person for the rest |
| Best for | The same steps, every time | The same goal, with different inputs each time |
| What you must check | That the steps are still right | The result, because AI can be wrong |
The two work well together. Use fixed steps where the work never changes. Use AI where a person used to read, decide and write. And keep one honest limit in mind: AI is still probabilistic, which means it does not give the same answer every time. So a person checks the result.
Where to start with AI automation: three tests
Which process should go first? Researchers at Ulm University reviewed 63 publications on RPA and collected the criteria used to select a process.[2] Their review is about RPA, not AI agents. We think three of the criteria carry over:
- It repeats. The most frequently mentioned criterion is “repetitiveness”: a high number of cases.
- It follows a pattern. The process should be “standardised” and “only require limited exception handling”.
- It is not too complex. In the words of the review, “the lower the complexity, the better the process is suited”.
We turn those tests into a simple picture. Every use case gets two scores: how much it is worth, and how hard it is to do. Then you start in the corner where the value is high and the effort is low.
Where the value usually is
You do not have to guess which departments to look at first. The data points in the same direction.
- Marketing, sales and administration are common uses. Among EU enterprises that use AI, 34.70% use it for marketing or sales and 31.05% for business administration or management.[3]
- Reported gains exist, and most are small. In the Stanford AI Index 2025, 49% of respondents whose organisations use AI in service operations report cost savings, and 71% of those using it in marketing and sales report revenue gains. The same report adds that most of them report savings below 10%, and that the most common revenue gain is below 5%.[4]
- Office work is the most exposed. The International Labour Organization finds that clerical occupations have the highest exposure to generative AI. Exposure means that AI could do part of the tasks in a job. The paper finds that the most likely effect is that jobs change.[5]
That is why we often start in go-to-market: sales, marketing and customer success. The work repeats every week, it runs across the CRM (the customer database), email and the calendar, and the result is easy to see.
A worked example: sixteen manual steps, one loop
Here is an illustration from sales. From the first signal, which is a sign of interest, to a logged outcome, a team can work through sixteen steps by hand. They fall into four groups:
- Find. Source the account, enrich the contacts (add missing details), verify the emails, score the fit, pull the signals.
- Prepare. Match the playbook (the plan for that type of customer), write the copy, personalise it, schedule the send.
- Run. Send and track, classify the reply, book the meeting.
- Learn. Update the CRM, alert the sales rep, log the outcome, feed the result back into the first step.
Each step is small. Together they fill the week. With AI automation, one agent runs the loop, with your rules at every step. An agent is an AI that carries out a task in several steps, inside your tools. A person steps in where it matters: the reply, the meeting, the relationship. See this loop and other examples on our page about what agents do.
This is annoying work that nobody should do. That is the work AI should take. Your people keep the work only people can do.
If the difference between a chat and an agent is not yet clear, our guide to agentic AI vs generative AI shows it with one task done both ways.
What not to automate first
- A broken process. In 1990, Michael Hammer wrote an article in Harvard Business Review with the title “Reengineering Work: Don’t Automate, Obliterate”.[6] The title says it. If a process has steps that nobody needs, remove them before you automate.
- Work that is rare. A task that happens twice a year will not pay for the effort.
- Decisions that belong to people. Hiring, pricing a difficult deal, a hard conversation with a customer. AI can prepare these. A person decides.
- Anything without clear rules for data. Decide first which data may go into AI and what it may do. Then build.
How to find your own starting point
You can do a first pass yourself in an afternoon. List the tasks each team repeats every week. Mark the ones that follow a pattern. Estimate the hours. The top of that list is your shortlist.
If you want the numbers checked and a working prototype at the end, that is what PromptPartner’s AI assessment does in four weeks. Here is the full method.
Frequently asked questions
What is AI automation with an example?
AI automation is the use of AI to carry out work that used to need a person at each step. For example: an agent reads a customer’s recent emails, drafts a payment reminder that fits the situation, and waits for a person to approve it.
What is the definition of automation?
The International Society of Automation defines it as the creation and application of technology to monitor and control the production and delivery of products and services.
What is the difference between AI and automation?
Automation follows steps that someone wrote down. AI can read language and choose the next step within rules. AI automation combines the two: fixed steps where the work never changes, and AI where a person used to read, decide and write.
Which process should we automate first?
One that repeats often, follows a known pattern and is not too complex. Score each candidate on its value and its difficulty, and start where the value is high and the difficulty is low.
What are the four main types of artificial intelligence?
The list most people mean comes from an essay by the researcher Arend Hintze in 2016: reactive machines, limited memory, theory of mind and self-awareness. It is one researcher’s way to sort AI, not an official standard, and it does not help much with business decisions.[7]
Does AI automation replace jobs?
The International Labour Organization finds that, globally, one in four workers is in an occupation with some exposure to generative AI, and that the most likely effect is that jobs change. Our aim: agents take the admin, and people keep the work only people can do.
Next step
If you have a list of manual work and do not know which item to take first, book a 30-minute call.
No pitch. You’ll leave with at least one recommendation you can use — whether or not we ever work together.
Sources
- [1]International Society of Automation, “What is automation?”— retrieved 5 October 2026
[2] Wewerka and Reichert, Ulm University, “Robotic Process Automation: A Systematic Literature Review,” 2020 — a review of 63 publications; it quotes IEEE Standard 2755-2017
[3] Eurostat, “Use of artificial intelligence in enterprises” — data extracted December 2025
[4] Stanford HAI, AI Index Report 2025, Economy chapter — based on a McKinsey survey
[5] International Labour Organization, Working Paper 140, May 2025 — exposure of occupations to generative AI
[6] Michael Hammer, “Reengineering Work: Don’t Automate, Obliterate,” Harvard Business Review, July–August 1990 — Harvard Business Review article
[7] Arend Hintze, “Understanding the four types of AI,” The Conversation, 14 November 2016 — one researcher’s essay
