Automation has existed for decades: spreadsheet macros, mail sorting rules, scripts that move files around. What artificial intelligence adds is the ability to handle unstructured input, meaning freely written text, scanned documents, emails with all sorts of different requests. This greatly widens the range of what can be automated, but it also makes it more important to choose carefully what to entrust to the machine.
Traditional automation and AI automation
| Aspect | Traditional automation | AI automation |
|---|---|---|
| Input | Orderly, predictable data | Free text, images, assorted documents |
| Logic | Explicit "if... then" rules | Models that interpret and classify |
| Behaviour | Deterministic, always the same | Probabilistic, may vary |
| Errors | Rare but systematic | More varied, sometimes unpredictable |
The practical consequence is that an AI workflow must be designed from the outset with error handling in mind, not as an exception but as a normal part of the system.
How to choose the right processes
A good candidate for AI automation usually has these characteristics:
- high volume: it recurs dozens or hundreds of times a month;
- recognisable structure: even if the text is free-form, the task is always similar (classify, extract, summarise);
- recoverable errors: a mistake can be corrected without serious harm;
- clear quality criteria: you can say with certainty whether the result is right or wrong.
Conversely, it is wise to avoid or postpone processes where an error has heavy legal, financial or personal consequences and is not easy to notice in time.
Typical examples
- Sorting incoming mail. Emails to a general inbox are classified by topic (orders, complaints, invoices, job applications) and forwarded to the right department.
- Extracting data from documents. Invoice number, date, amount and VAT number are extracted from different suppliers' invoices to pre-fill the accounting entry.
- First-line support replies. For frequent questions, a draft based on internal documentation is prepared and an operator approves it.
- Normalising master data. Addresses written in different ways are standardised and likely duplicates flagged.
- Periodic reports. A written commentary explaining the main changes is generated from raw data.
Designing the workflow
A well-built automation follows a clear sequence.
- Input. Where the data comes from and in what format.
- AI processing. What the model must do, with precise instructions and ideally a fixed output format, such as a list of fields.
- Automatic checks. Simple tests on the result: is the amount a number? Is the date plausible? Is the category one of those expected?
- Confidence threshold. If the system is uncertain or a check fails, the case goes to a person.
- Action. Recording, forwarding, sending.
- Log. Every step is recorded so that what happened can be reconstructed.
Human oversight
There are different levels of human involvement, to be chosen according to risk:
- human in the loop: every result is approved before any action;
- human on the loop: the system acts on its own, but a person spot-checks and can intervene;
- human out of the loop: full automation, acceptable only for very low-risk tasks.
A prudent strategy is to start with a human in the loop and reduce oversight only when the data shows the system is reliable.
What to measure
Without measurement it is impossible to know whether the automation works. The most useful indicators are:
- correction rate: how often a person has to change the result;
- processing time per case, before and after;
- cases sent to manual review and the reasons why;
- errors that slipped through and were found later, for instance through a complaint.
Tracking these values over time also helps you notice when something changes: a new supplier with invoices in a different layout can make extraction quality collapse.
Security and data
An automated workflow often handles sensitive data and acts without anyone watching each step. Some precautions:
- give the automation only the permissions it strictly needs: if it has to read emails, it should not be able to delete them;
- treat incoming content as untrusted: an email may contain instructions written specifically to confuse the model;
- keep logs for an appropriate period and protect them;
- check where the data is processed and under what contractual guarantees.
The most common mistakes
- Automating a messy process: you only get faster mess.
- Removing human oversight straight away to maximise savings.
- Not planning what happens when the model cannot answer.
- Forgetting maintenance: models, formats and suppliers change.
In short
- AI extends automation to unstructured data but introduces probabilistic errors.
- Ideal candidates are high-volume, repetitive processes with recoverable errors.
- Design automatic checks, confidence thresholds and a log of every step.
- Start with a human in the loop and reduce oversight only on the basis of measured data.
- Limit permissions and treat every external input as potentially hostile.
