Automating processes with AI: where it really pays off for SMEs

Most business owners who want to do something with AI know there is time to be won. What they do not know is where. Every supplier shows a demo with a chat window, and afterwards you still do not know which piece of work could be different tomorrow.

Below is the concrete answer: processes at SMEs that genuinely lend themselves to AI, with what happens in each, what your people keep doing, and when you are better off not starting. Plus the processes you leave alone.

What “automating processes with AI” means in practice

Putting a chatbot on your website is not process automation. A chatbot answers questions. A process is work that comes in, passes through a few pairs of hands and ends up somewhere in a system. Automating means some of those actions disappear.

In practice such an automation consists of three layers. The order matters, because most projects come unstuck on the third layer while the demo only showed the first.

• Understanding. Turning something messy that arrives — an email, a PDF, a photo of a scribbled note, a voice message — into fields your system understands. This is where AI adds something fixed rules cannot.
• Deciding. Working out what should happen with that case, following rules you have written down. Often plain if-then logic, sometimes a judgement call.
• Pushing it through. Actually getting the result into your software: a line in Exact or AFAS, a draft in Moneybird, a task in the planning. Without this layer you have a demo and nothing else.

AI is good at the messy front end of a process. You only capture the gain once the back end lands in your own system.

Almost everything below works the same way: the machine makes a proposal, a person approves it. That sounds like half a job, but it is the reason these things keep running. Looking and approving takes seconds, typing it in yourself takes minutes.

Sorting incoming email and routing it

The classic case: an installation company with a shared info mailbox where breakdown reports, quote requests, supplier invoices and newsletters all arrive together. A system reads each email, works out the type, extracts which customer and which address it concerns, and puts the message in the right queue. Urgent breakdowns get their own signal.

What people keep doing: judging whether a breakdown is genuinely urgent, and picking up the cases the system flags as uncertain. What you never let happen automatically: sending a reply to the customer without anyone having seen it.

When it does not work: if nobody can explain what the categories should be. Sorting only works if you know what you are sorting into. And if your customers mainly phone, you are solving the wrong problem here.

Producing quotes from a standard base

Many quotes are largely the same as the previous one. At a painting or installation company they differ in address, quantity, choice of materials and a few lines of text. A system reads the request, pulls the customer details from your software, picks the standard text blocks and the right price lines, and prepares a draft in your own template.

What people keep doing: setting the price and sending the quote. The margin and your read on whether you actually want this job are in no system anywhere. The gain is that a quote can go out the same day.

When it does not work: if every job is unique and the calculation is the actual work. With detailed specifications you automate the text blocks around it at most. And if your price list only exists in the owner’s head, that is where you start.

Reading in packing slips, order confirmations and invoices

At a wholesaler, supplier documents arrive daily in ten different layouts, partly as PDFs and partly as photos. Someone retypes them. A system reads them out, matches the lines to your own product numbers, compares them with the order and puts the discrepancies at the top: wrong quantity, different price, missing line.

What people keep doing: assessing the discrepancies and calling the supplier. What disappears is the retyping of the lines that are simply correct. That is normally the bulk of it.

When it does not work: if your product data is messy. Three spellings of the same item, suppliers using their own codes without a translation table. A system cannot guess what you mean. Cleaning that up is dull manual work up front and often the biggest part of the project.

Job sheets from paper into the system

Engineers and site crews write on paper or send a message. On Friday afternoon someone types it all in. A system reads a photo of the job sheet or a voice message, recognises the job, the hours, the materials and the comment, and prepares a draft job sheet that the engineer checks on their phone.

What people keep doing: confirming it is correct, the same day. That is also the real gain: not the minutes saved in admin, but being able to invoice on Thursday instead of two weeks later.

When it does not work: with illegible handwriting or half-completed sheets. Recognising printed text goes well; pencil on a dirty sheet stays hard. Often a simple form on the phone is the cheaper solution, with no AI at all.

Stock alerts and purchase proposals

A system looks at your sales, your open orders and your lead times, and flags which items are heading towards zero before it goes wrong. It prepares a purchase proposal with the right supplier, the quantities and a short reason: this is selling faster than last month, this supplier is now slower to deliver.

What people keep doing: ordering. Always. A purchasing decision costs money and carries context that is not in your data, such as a customer starting a big job next month.

When it does not work: if your stock levels do not match what is actually on the shelves. Then you are forecasting on noise. It also disappoints where demand is project-driven with no pattern; there your buyer is simply better.

Summarising files and meeting notes

At an accounting firm or a consultancy, a lot of knowledge sits in long email threads and loose documents. Getting back into a file costs twenty minutes. A system produces a summary: what was agreed, what is still open, which documents are missing, when there was last contact.

What people keep doing: drawing the conclusion and giving the advice. The summary is groundwork, not a judgement. Always show which sources it came from, so someone can check it in two clicks.

When it does not work: when the content is legally or financially binding. A summary that leaves out a nuance reads well and is wrong. Use it to get up to speed faster, not as a replacement for reading.

Preparing the planning

At a haulage firm or an installation company, planning is a daily puzzle. A system can prepare the puzzle: bundling open orders by region, taking account of driving times, engineers’ certifications, available equipment and appointments already fixed. The draft plan names the bottlenecks.

What people keep doing: shuffling it. The planner knows this customer is difficult if you arrive later than ten o’clock, and that this particular engineer had better not go to that address. You will not capture all of that, and you do not want to.

When it does not work: if your planner does not trust it. This is the process where resistance most often kills the project, and rightly so, because it touches craftsmanship. So start with preparing and flagging, not with deciding.

Processes you should deliberately not automate

Just as important as the list above: what you leave alone. Not out of caution, but because it costs money without delivering anything.

• Work that happens a few times a month. Ten minutes of work, twice a month, will never repay an integration costing a few thousand euros.
• Processes nobody can explain. If five staff describe five ways of working, the process itself is the problem. Write it down first, automate second.
• Reports nobody reads. Producing them faster is waste with a technical coating. Scrapping them is the better move.
• Decisions about people: screening applicants, drafting appraisals, flagging absence. Legally loaded, and it costs you trust on the shop floor.
• The final contact with an angry customer. You can automate the preparation, not the conversation.
• Anything where money goes out without a person in the loop: ordering, paying, invoicing. Put those on proposal, with someone who presses approve.
• Processes you are going to replace within a year anyway. If you are moving to a different ERP package, do not build integrations on the old one.

How to choose where to start

You do not have to pick what delivers the most. You have to pick what proves fastest that this works, here, with your people.

1. Walk around for a week and note where people sigh. Do not ask how it goes, watch. What people tell you is the tidy version.
2. Pick three candidates you recognise from the list above, and immediately strike out anything that happens less than weekly.
3. For each candidate, count the actions and the waiting time. How often is the same piece of data retyped, and how long does it sit idle.
4. Look at how many systems have to join in. One or two is a good first project. Four is a good second project.
5. Choose something where a mistake is visible and fixable: a wrong label, a draft that is off. Not the payroll.
6. Measure one number before you start, then build the narrowest link that has value on its own. Let it run alongside the old way for two weeks.

With us, that usually starts with a half-day core session (four hours) on your own shop floor. We follow one process from start to finish and count the actions. That costs €596 excluding VAT and travel expenses and delivers a core report. Build work is charged per hour: €110 for straightforward work, €165 for complex development.

What disappoints about this

Two things, every time. The first is the exceptions. You hear that a process always goes the same way, and within two weeks six cases turn up that did not fit that story. The last ten percent is often more work than the first ninety.

The second is that saved hours are rarely saved payroll. The work shifts to things that had been left undone. Valuable, but it does not appear as a minus on your profit and loss. What is measurable in hard money: invoices going out earlier, and quotes you currently lose because they arrive too late.

When you need us and when you do not

If the process sits entirely within one package that already has the feature on board, get going without us. Many first steps are a matter of switching on a setting and spending an afternoon trying it out. The same applies if you mainly want to answer email faster; that needs a subscription, not an agency.

We are useful as soon as a process runs across several systems that do not talk to each other by themselves, as soon as custom software is needed that no tool provides, or as soon as you cannot choose between ten candidates. And if after the core session we think you are better off with a standard package, we say so.