Automating small business processes: what, with what and when not

To automate a process in a small business, start with the work someone copies by hand from one place to another: orders that arrive by email and get typed into the management software, supplier catalogues entered by hand, a shared inbox where nobody knows what has been dealt with. Then pick the smallest tool that solves it. Sometimes that is a connector that already exists; sometimes it is software built for that process; you almost never need AI for all of it.

What follows comes from what we have built, with the real status of each case.

Which processes are worth automating

A process is a good candidate when it meets three conditions at once:

• It repeats. It happens every week, not twice a year.
• It follows steps that can be written down. If two people on the team would do it the same way, it can be automated. If it depends on the judgement of whoever does it, it can be assisted, not replaced.
• Today it is a copy. Someone reads a piece of data in one place and writes it in another. That is where hours go and errors creep in.

The ones that repeat most in the companies we work with are always similar: order entry, loading supplier catalogues or price lists, sorting the email that lands in a shared inbox, and moving data between the management software and the accounting software.

What you should not automate first is whatever still changes every month. Automating a process that is not clear only makes the mess faster.

When a generic tool is enough

Tools like Zapier or Make connect applications that already publish connections. They are enough when these conditions hold:

• Both ends are well-known applications that already offer that connection (a form, a spreadsheet, a cloud CRM).
• The data arrives already tidy: fixed fields, always in the same format.
• If something fails, nothing serious happens: a notification that does not arrive, a row that has to be redone.

Before looking elsewhere, check the program you already pay for. Many management and accounting programs come with imports, rules or alerts that nobody has switched on.

They stop working when the process asks for something they do not do well: reading a scanned PDF, understanding a spreadsheet each supplier lays out their own way, writing into an old management program that has no way to connect, or applying rules that only exist in someone's head. That is where the generic tool fills up with patches and ends up being a custom system, but with nobody who understands all of it.

When you need custom software

You need it when the input has no fixed shape, or when the process is what sets the company apart. Two of our cases, told with their real status.

An importer in Barcelona: catalogues that each arrive their own way

An importer that works with factories in Asia receives product information in spreadsheets that each supplier lays out their own way, in English or in Chinese. Entering them by hand into the management software was the heaviest job in the office, and that software no longer did what they needed.

We are building them custom management software, module by module, starting with the product catalogue. Importing those spreadsheets uses AI to recognise which column is which, and each import is tested against their suppliers' real catalogues. The accounting software they already use is not touched: it keeps doing the accounts, and the connection to it comes in a later module.

Status: contracted and in development. It is not in production yet and has no users.

A manufacturer: orders as PDFs, photos and handwritten

A manufacturer received its customers' orders by email in every format possible: PDF, Excel, Word, phone photos and handwritten sheets. Someone read them and typed them into the management software.

We built a system that reads those orders and puts each line into a file with the columns their management software can import. Nothing is exported until a person has checked it:

Order by email

Automatic reading

Lines to review

Is it right?

Export to management software

no

Correction by a person

It can run entirely on the company's own machines, without sending the documents to any outside service.

Status: built and technically validated with real orders, handwritten ones included. It is not in use.

In both cases automation does not remove the person: it removes the copying. Whoever knows the product still decides what is unclear.

When you do not need AI

AI is for reading what has no structure: a photo, handwriting, a spreadsheet that changes from one supplier to the next. For everything else, well-written rules are cheaper, faster and can be explained.

An accounting firm: email sorted without AI

An accounting firm had one department's email in a shared inbox, with no priorities and no way to know what had been dealt with. We built a board where each email comes in as a card and moves from pending to in progress to done.

The sorting does not use AI. Each category (new hires, leaves, payroll, garnishments, enquiries) has its keywords and its senders, and the one with the most matches wins. Whatever fits none goes to enquiries. Each category also has a response target, and an unanswered email rises in priority on its own as it gets closer to that target.

Incoming email

Rules by words and sender

Does it fit a category?

Pending column

no

Enquiries

If a rule sorts something wrongly, you can see why and fix it. With an AI model you would have to explain a decision nobody wrote down. In an accounting firm, which also handles sensitive data such as sick leave, that matters.

Status: built and tested. Waiting to be started at the firm.

The useful question before bringing in AI is this: if you were explaining the job to a new person, would you give them rules or ask them to read? If you would give them rules, you do not need it.

What you should not build custom

Custom does not mean everything custom. There are parts of a company where the right move is to use a product that already complies with the rules and connect to it.

Invoicing is the clearest example. Spain's Royal Decree 1007/2023 (BOE, 2023) requires invoicing software to guarantee, among other things, that records cannot be altered and can be traced, chained so their order can be checked. Royal Decree-law 15/2025 (BOE, 2025) postponed the obligation: companies that pay corporate income tax have until 1 January 2027, and everyone else until 1 July 2027.

That is why we do not build our own invoicing engines. Custom management software can prepare the data and hand it to the accounting or invoicing program that already complies, and issuing the invoice is that program's job. It is the same approach as with the importer in Barcelona: their accounting software stays where it is.

The same goes for accounting and payroll: they are regulated processes with mature products. What gets built custom is what goes around them.

Where to start

1. List the copies. For two weeks, note down every time someone moves a piece of data by hand from one place to another. Whatever shows up most is the first candidate.
2. Look at where the data comes from. If it arrives tidy, try what you already have and a generic tool first. If it arrives as PDFs, photos or formats that change, that is what needs a custom solution.
3. Decide who reviews. Every automated process needs a point where a person sees what is unclear. Design it in from the start, not when it fails.
4. Start with one module. One process solved and in use teaches more than a plan to automate the whole company.
5. Leave the regulated parts out. Invoicing, accounting and payroll, with software that already complies.

If you want to see the kind of projects we do, they are on services. If you are unsure whether to automate with AI or without it, or whether you can start with a single module, we answer that in the FAQ.