In short
- Start with a repetitive process with many cases and clear rules.
- If the rules can be written down, plain automation is enough - it is cheaper and predictable.
- AI makes sense where data is unstructured: emails, documents with varying layouts, descriptions.
- With AI, keep a person checking the result, at least at the start.
- A pilot on one process with a measurable result - before rolling out wider.
Automation or AI?
| Rule-based automation | AI | |
|---|---|---|
| When | The rules can be written down ("if amount > X, send to the manager") | Data is unstructured, there are too many rules or they change |
| Examples | Routing documents, notifications, reports, syncing data | Reading data from documents with varying layouts, classifying emails, summaries |
| Result | Always the same for the same data | Very good, but needs checking |
| Cost | Usually lower | Higher: testing, quality control, model usage costs |
How to pick the first process
- Repetitive and frequent - hundreds of cases a month, not a handful.
- Measurable - you know how long it takes today and what it costs.
- With a clear owner on the company side who knows the exceptions.
- No disaster when something goes wrong - do not start with the process where a mistake costs the most.
How to run a pilot
- Measure the current process: time, number of errors, cost.
- Build the solution for one process and one user group.
- With AI, keep a person approving the result and count how often they have to correct it.
- After a few weeks, compare the results with the starting point.
- Decide: expand, improve or stop.
Data and security
With AI you need to answer: where does the data go, can the model provider use it, how are personal data and trade secrets protected. Make these decisions before the pilot, not after. In many cases the data sent to a model can be limited to the necessary minimum.