What to automate first in your company — and what to leave alone

Everyone in the company can feel how much is still done by hand: someone retypes orders, someone builds the report every Friday, someone answers the same emails over and over. When the question comes up of what to automate first in your company, the winner is usually whatever is most visible, or an idea someone saw at a conference. Both tend to be the wrong choice.

The short answer: automate first the process that happens often, eats a lot of hours, costs something when it goes wrong, and whose inputs arrive in a similar shape. In a typical small or mid-sized business that is almost always retyping documents — orders, invoices, delivery notes — or answering repeat questions. Not strategic decisions, and not “an AI assistant for everyone”.

Business process automation does not have to start with a big project. Below is a way to find the right candidate in your own company within a week, how to work out the payback, and what to steer clear of first time round.

What to automate: four questions for every process

Give each candidate 1 to 3 points per question. Multiply the points rather than adding them — a process that fails badly on one count should not go first, however well it scores elsewhere.

• Volume — how often does it happen? 1 = a few times a month, 2 = several times a day, 3 = dozens of times a day.
• Time — how long does one case take? 1 = under two minutes, 2 = two to ten minutes, 3 = more than ten minutes.
• Cost of errors — what does a mistake cost? 1 = nothing, it gets fixed in passing, 2 = someone has to track it down, 3 = it leads to a complaint, a wrong invoice or an unhappy customer.
• Input regularity — what does the input look like? 1 = every case is different and handled by judgement, 2 = always the same data, but in different formats, 3 = the same data in the same form.

One thing to watch on the last question: if you score a 3 — the input is always identical — automation will be cheap, and you probably will not need AI for it. AI earns its keep mainly at a 2: the data is always the same, but every customer or supplier sends it differently.

A model example of what the result might look like at a distribution company with twenty people:

The score is not science. It ranks the candidates; the actual decision comes with the payback estimate.

A one-week log of manual work

Without a log you decide on gut feeling, and gut feeling is misleading. People remember the tasks that annoy them, not the ones that actually take up most of their time. A five-minute chore repeated forty times a day simply does not stick in the memory.

How to do it:

1. For one normal working week, everyone who does admin work notes down their repeat tasks — as they go, not from memory on Friday. A shared spreadsheet is enough.
2. Five columns per task: what I do, how many times, how long it takes, where the input comes from (email, PDF, phone, paper) and where I enter it (ERP, spreadsheet, email).
3. A sixth column: what happens if I get it wrong. This tells you more about the cost of errors than any statistic.
4. Month-end tasks (closing, invoicing, reports) go on a separate list so they are not lost.
5. Ask the people who do the work, not just management. The best automation candidates sit at a desk in purchasing or in the warehouse, not in the meeting room.

After a week you have a list with hours attached. A common side effect is discovering that some steps should not be automated at all, just dropped — a check done by two people because nobody is sure who owns it, for instance.

If you want a quick first read on the state of your processes and data, try our short AI readiness check.

How to estimate the payback

The sum is simple and does not need dressing up:

hours per month × labour cost × share the automation handles + errors avoided − monthly running costs

A model calculation for the order retyping from the table above. We assume a hundred hours a month (three people, roughly an hour and a half a day each) and a fully loaded labour cost of €16 an hour:

• Manual work: 100 h × €16 = €1,600 a month.
• The automation handles about 70% of cases with no human involvement: roughly €1,120 saved.
• Fewer errors and complaints: a modelled €200 a month.
• Operations and monitoring: −€390.
• Net benefit: roughly €930 a month.

With an investment of €10,000–18,000 (analysis, pilot and production rollout connected to the ERP), that pays back in about 11 to 19 months.

Now the important part: the same solution at thirty hours a month never pays back. The running costs swallow the entire saving. Much of the cost of a bespoke solution is fixed, which is why volume is the first criterion, not the last. If a process does not account for at least several dozen hours a month, reach for a simpler tool (see below). We break down the price ranges in more detail in how much an AI solution costs.

What is a bad fit for a first project

A first project has one job: to prove that automation works in your company and to teach you how to run this kind of project. So avoid anything with a high risk of ending up in a drawer:

• Rare processes. A task done once a quarter will struggle to pay back, and by the time it comes round again something will have changed and the automation will break.
• Politically sensitive processes. Staff appraisals, candidate screening, bonus decisions. People see such a project as a threat, not as help — and under the EU AI Act, recruitment and employee evaluation count as high-risk uses anyway.
• Processes with no data. If something is done from one person’s head and there is no history or sample of cases, there is nothing to measure the automation against. The same goes for a catalogue where one product has three codes — tidy up first, automate second.
• Processes that are about to change. If you are replacing the ERP in six months, wait. Otherwise you build the integration twice.
• Processes nobody owns. If nobody is responsible for the outcome, nobody will approve exceptions or say whether it works.

When you do not need AI: a spreadsheet or Make will do

Process automation in a small business often needs no AI at all. An honest list of situations where AI is needlessly expensive:

• The input is always the same — a shop export, a web form, a bank statement. Rules, a macro or a scenario in a tool like Make or Zapier will do it for a fraction of the price.
• The report is built from the same tables every time. Power Query in Excel or a dashboard in something like Looker Studio will handle it without AI.
• Your system already does it. Automatic payment matching, order import, scheduled exports — features you have paid for in the ERP but nobody has switched on. Worth checking before you buy anything.
• The problem is the process, not the tool. An unnecessary step, once automated, is still unnecessary.

AI makes sense where something has to be understood: an email written in free text, PDFs from hundreds of suppliers, a customer question whose answer is spread across three documents.

How to get started with automation in three steps

1. A one-week log of manual work, ranked by the four questions.
2. A payback estimate for the top two or three candidates — using real hours, not management’s guess.
3. A pilot on a small sample of real data with a clear success number, such as “70% of invoices with no human involvement, at a lower error rate than today”. Only once the pilot hits that number do you go to production.

You can see what such solutions look like in practice in our solutions by problem — for example order automation, automated invoice processing or reporting automation. More examples by department are in 15 business processes worth automating.

If you would rather have someone from outside do the log and the numbers, that is exactly what our AI audit is for: in one to two weeks you get a ranked list of processes with estimated savings and effort — including the ones we recommend not doing.

Summary

• Choose your first process by volume, time, cost of errors and input regularity, not by what is most visible.
• Decide from a one-week log, not a hunch.
• When you work out the payback, remember the running costs — at low volumes they eat the whole saving.
• Rare and sensitive processes, processes without data and processes about to change are poor first projects.
• If the input is always the same, a spreadsheet, Make or a feature of your existing system is enough.

What could be automated in your company? Tell us what you do by hand today and how many hours it takes — you do not need to know whether you need AI. We are happy to go through it with you, and if it does not pay off, we will tell you.