Let's start with what nobody tells you
Every week I talk to business owners who have heard they "need to get AI in" and have no idea where to start. Some have already tried: they paid for a tool, used it for three weeks, and today nobody opens it. Others are afraid a chatbot will say something outrageous to an important customer.
Both reactions are reasonable. AI in the right place saves hours and brings in sales. AI in the wrong place creates work, errors and a monthly invoice nobody remembers the reason for.
The difference is not the technology. It is the judgment with which it is applied. And judgment can be learned. Let's go.
Rule number one: AI does not fix a broken process
If your quotes currently take four days because each one passes through three people who don't talk to each other, AI won't solve that. It will make a process that was already bad run faster.
Before automating anything, draw the process on a sheet of paper. Literally: who does what, in what order, with which tool. In most of the companies I work with, this exercise alone removes a third of the steps. Without any new technology.
Then, and only then, look at where AI can take repetitive work away.
Where to start: the three areas that almost always work
After rolling out automations in very different companies, there are three areas where AI almost always performs well with little risk:
1. Answering repeated questions. Opening hours, prices, availability, "do you ship to…?", "how does … work?". If your team answers the same thing twenty times a day on WhatsApp or email, an assistant trained on your information can handle 60 to 80 percent of those and pass you the rest. This is what I build in the AI process automation service.
2. Sorting and preparing information. Reading an email and deciding whether it's an order, a complaint or an invoice. Extracting data from a PDF and putting it in your system. Summarizing a call and creating the follow-up task. AI does this well and never gets tired.
3. Drafting first versions. Proposals, replies to customers, product descriptions, internal reports. AI writes the draft; a person reviews it in two minutes instead of writing it in twenty.
What they have in common: high-volume tasks, clear rules, where a mistake is easy to spot and cheap to fix.
Where not to put AI (yet)
This is where companies get wrecked, so I'll be very clear.
Money decisions without supervision. AI proposing a discount, fine. AI applying it on an invoice by itself, no. AI suggesting which supplier to pay first, fine. AI executing the payment, no.
Delicate conversations. A serious complaint, a negotiation with a large customer, a dismissal. AI can prepare the ground, but a person has the conversation.
Anything affecting health, safety or legal compliance. If your business is a clinic, an accounting firm or a transport company, there are decisions that carry legal responsibility. AI assists; it does not decide.
Processes that change every week. Automation works when the process is stable. If you're still defining how you work, wait.
A rule I use with all my clients: AI proposes, a person approves. Once you've spent three months seeing it propose well, you give it more autonomy. Never the other way around.
The mistake that costs the most money: starting with the tool
"We've bought ChatGPT for the whole company." I've heard it many times. And it almost always ends the same way: five people actually use it, the rest don't know what for, and nobody has measured anything.
The right order is the opposite:
- Identify one concrete task that eats hours and annoys the team.
- Calculate what it costs today (hours per week times the cost of that person).
- Design what it would look like with AI, including who supervises it.
- Choose the tool that fits that design.
- Test for two weeks with real data.
- Measure. If it saves what you said, expand. If not, adjust or drop it.
Notice that the tool shows up in step four. Not step one.
What it costs and what it returns
Real figures from what I see in companies with 5 to 50 people:
- A WhatsApp assistant that qualifies inquiries and books appointments: €2,500 to €5,000 to implement, plus €30 to €150 a month in usage. Typical saving: 40 to 80 hours a month of one customer service person.
- A workflow that reads supplier invoices, extracts the data and loads it into accounting: €3,000 to €6,000. Saving: 20 to 40 hours a month of admin, and far fewer errors.
- A system that automatically follows up sent quotes and alerts the salesperson when the customer opens the document: €2,000 to €4,000. Effect: 10 to 25 percent more closed quotes, because follow-up stops depending on each person's memory.
With those numbers, most projects pay for themselves between the second and fifth month. If someone proposes something that doesn't pay back in under a year, ask why.
What about your data
Mandatory and very healthy question. Three short answers:
Yes, AI needs to see your data to work with it. But not all of it. A customer service assistant needs your catalog and your terms, not your accounting.
No, serious providers do not train their models on your data if you use their business plans or their APIs. This has to be configured correctly and put in writing.
Yes, you have GDPR obligations (and similar ones in the US) if you process people's data. You need to document what is sent, to whom and for what purpose. It's not complicated, but it has to be done. I include it in every project because it's what avoids problems later.
The team: the part almost everyone forgets
Technology gets implemented in weeks. Getting people to actually use it takes months. And that's where it's decided whether the investment was worth anything.
Three things that work:
- Start with whoever is keen. Every company has one or two curious people. Let them test first. Their enthusiasm spreads better than any order from management.
- Train with real cases from the company, not generic demos. "Look, this email you got yesterday, this is how the system handles it." That connects. A 40-slide presentation on artificial intelligence does not.
- Make clear what will NOT change. The main fear isn't the tool, it's losing the job. If automation is going to free hours for people to do higher-value work, say so from day one and keep your word.
A 90-day plan you can copy
This is the framework I use in the monthly digital partnership, and it works just as well if you do it on your own:
Weeks 1 and 2: diagnosis. Map the five processes that eat the most time. Calculate the cost of each.
Week 3: pick one. The one that hurts most and carries the least risk. Usually answering inquiries or some admin task.
Weeks 4 to 8: build and test. With real data, with human supervision, with metrics from day one.
Weeks 9 to 12: measure and decide. Did it save what you expected? Expand to a second process. It didn't? Adjust before spending another euro.
At the end of 90 days you have something most of your competitors don't: an automation that works, a team that uses it and real data to decide the next one.
In short
AI in your company is not a gamble; it's a management decision like any other. It's made with data, tested at small scale, measured and expanded if it works.
If you'd like me to review those five processes with you and tell you where I'd start, book a call. Fifteen minutes, no commitment, and you leave with a clear idea of what makes sense in your case and what doesn't.
