Your first AI automation: start with the process that eats someone's week
Start with one process, not with the technology. How to find the right one, measure it and decide whether it is worth automating, with a calculator that runs on your own numbers.
AI and automation5 min readSoventia Ltd
The best first AI automation is rarely the most impressive one. It is one specific process that takes up hours of someone's week, follows rules a person could explain, and ends with something being entered into a system. Find that process, measure it, automate the routine part, keep a person on the exceptions, and measure it again.
What makes a good first candidate
Look for a process that ticks most of these:
- It happens often. Daily or weekly, not twice a year.
- The rules are mostly clear. Someone can explain how they do it, even if there are exceptions.
- The inputs are digital. Emails, PDFs, forms, spreadsheets, tickets.
- It ends in a system. An entry in the accounting software, a record in the CRM, a reply sent.
- Mistakes can be caught. A person can check the result before it matters.
Typical examples are supplier invoices, incoming enquiries that need sorting, weekly reports pulled together from several tools, and documents that have to be read and summarised before someone decides what to do with them.
What to leave for later
Some processes look tempting and make poor first projects:
- Rare or one-off work. There is not enough repetition to pay back the build.
- Pure judgement. If two experienced people would disagree on the answer, a model will not settle it.
- Decisions about people. Recruitment, promotion and similar uses are high-risk under the EU AI Act and come with their own obligations. They can be done, but not as a first experiment.
- No system of record. If the result lives in someone's head or on paper, connect that first.
Measure it before you touch it
For two weeks, write down how often the process runs, how long each run takes, and how often something goes wrong or needs a second look. It sounds tedious and it is the most useful thing in the whole project, because it turns "this takes ages" into a number you can compare against.
Then run your numbers through the calculator below. It shows the time you could get back and, if you have a quote for the build, how long the build would take to pay for itself.
Be conservative. Exceptions stay with people.
Salary plus overheads.
Optional, from a quote.
With your numbers
Hours spent on it now, per year
460
Hours you could get back
230
Value of that time, per year
€9,200
This leaves out running costs. Model usage and hosting are billed monthly, and we work them out before you commit.
What the automation actually does
Most useful automations have the same five parts: something triggers them, a model reads the input, the system acts through a defined interface, a person checks what matters, and the result is measured.
Read the transcriptHide the transcript
Anatomy of an AI automation, using supplier invoices as the example.
Trigger: something happens that someone currently handles by hand. Here, an invoice lands in a shared inbox.
Read: a language model reads the document and pulls out the fields that matter, such as the supplier, the amount, the due date and a suggested cost centre, running where you decided it should.
Act: the automation drafts the entry in your accounting system through a defined, permissioned interface such as an MCP server or an API. The entry is drafted, not posted.
Check: a person approves. Anything unusual, such as a new supplier or an amount out of range, goes to a human first and is never posted automatically.
Measure: time per invoice, before and after, measured on your own data. If it did not pay back, you will know.
The check step is the one people try to skip, and the one that makes the rest safe. A draft that waits for one click of approval is still most of the time saved, and it means a wrong reading of an unusual invoice never reaches your books.
Build small, then widen
- One process, end to end. Not three half-finished ones.
- Approval on by default. Loosen it only where the numbers show it is safe.
- Logs from day one. Every input, every output, every approval, kept in your own account.
- Measure again after a month, on the same terms as before. If it did not pay back, you will know, and you can stop or change it.
Three things to settle early
- Running costs. Model usage is billed per use and can grow with volume. Work it out at the volume you expect, not the volume of the pilot.
- Where the data goes. If the process touches personal data or client documents, decide where the model runs before you build. Our guide to where your AI should run covers the options.
- Who owns it. The automation, its accounts and its logs should sit in your name, so it keeps working and stays yours whoever maintains it.