Start with the problem, not the tool
The question is not “what can we do with AI”, but “what makes us lose the most hours”. Orders copied by hand from email into the ERP, customer questions that are always the same, reports someone puts together every Monday, documents that have to be read just to pull out a few pieces of data.
Pick one. Just one, that happens often and where it is easy to check whether the result is right or wrong.
Where AI tends to work well
- Reading documents: pulling the data out of PDF orders, delivery notes, supplier invoices or forms.
- Preparing replies: email drafts based on the company’s history and terms, which a person reviews and sends.
- Sorting and routing: issues, emails or requests that need to reach the right person.
- Asking in plain language: questions about sales, stock or customers without having to build a report.
- Summarizing: an opportunity, a case file or a meeting, with the next step.
Where it is not worth it
- When the work rarely happens: a task done twice a year does not pay off.
- When a mistake is very costly and nobody reviews the result.
- When the data does not exist or sits where it cannot be accessed.
- When the problem is the process: automating a mess only makes it go faster.
- When what is missing is two programs talking to each other: that is an integration, not AI.
Data and confidentiality
Before trying anything with customer or company data, check where it goes. The free versions of many assistants may use what you type to train their models; business versions, under contract, usually do not.
Work with a provider that puts this in writing, give the AI only the data it needs for each task, and make it inherit the permissions of whoever uses it: it should not see anything that person cannot see. If there is personal data, the legal notice and the record of processing activities have to be adjusted to match. We are not lawyers: if you have a legal question, check it with yours.
A small, measured trial
- Before you start, write down how it is done today: how long it takes, how many errors there are, who is involved.
- Build a first version for the chosen process, with real data and a person who reviews every result.
- Let it run for a few weeks alongside the usual way of working.
- Compare it with what you wrote down. If it saves time, you expand it; if not, you stop it.