AI in SME operations: where to start
Short answer
Start where work passes between systems: orders that arrive by email and get retyped by hand, documents someone sorts manually. That is where AI gives hours back to a small company, with a person approving every entry. Prices, purchase orders and legally required invoicing come later, and the result is measured in hours returned per week.
Start with one question: where does the team lose hours copying information from one place to another? In a small or mid-sized company, the AI that pays off almost always starts there, where work passes between systems, long before any customer-facing chatbot.
Where AI pays off first
AI pays off first where work passes between systems and someone retypes by hand what already arrived in writing. Two patterns repeat in companies with a store, a warehouse or a factory:
- Orders arrive by email or as a PDF, and someone types them into the management system again. AI reads them and prepares the entry; a person confirms it.
- Supplier invoices, delivery notes and complaints arrive mixed together, and someone sorts them. AI sorts first; the person handles the exceptions.
This is how AI entered Plurifil, a Portuguese company: today it handles the orders and delivery notes that come in, and nothing is recorded until a person approves it.
What not to automate first
Approval always stays with a person, and so do prices, purchase orders to suppliers, legal correspondence and anything that touches certified invoicing (the invoicing software the law requires). Automating approval is the most expensive mistake a small company can make in its first year. A mistake in those areas costs more than the hours saved, and the team’s trust in the system takes a long time to recover.
How to measure whether it paid off
It paid off if the saving shows up as hours returned to the team each week. Before automating, count: how many times a day the task happens, how many minutes it takes, who does it. After automating, count again. If the numbers do not add up, the automation was a toy. That is why the work starts by mapping the process as it runs today, from the email coming in to the confirmed entry, before proposing anything.
The first choice
The first choice is, of the points mapped, the one that happens most times a day, and it is rarely the most impressive one. Close the scope in writing and deliver it whole before opening the second. Half a dozen automations left half-done give nobody an hour back, and one finished automation gives hours back every week.
Questions
Should we start with a chatbot for customers?
No. The chatbot is the most visible automation and rarely the one that returns the most hours. It pays more to start where work passes between systems, such as the order that arrives by email and gets typed in again, and look at customer service later.
Can AI approve orders on its own?
It can propose; approval stays with a person. At Plurifil, AI reads incoming orders and delivery notes and prepares the entry, and a person confirms each record. Automating approval in the first year is the most expensive mistake, because a wrong record costs more than the hours saved.
How do we know the automation paid off?
Count before and after: how many times a day the task happens, how many minutes it takes and who does it. If the difference does not show up as hours returned to the team each week, it did not pay off. Draypec maps the current process before proposing anything.