Which AI projects pay off for SMEs? An honest assessment
6 min read
Short answer
AI projects pay off for small and mid-sized companies mainly where many similar texts, documents or enquiries come in: extracting data from receipts, pre-sorting emails, drafting standard replies, making internal knowledge searchable. Custom models, projects without a measurable goal and applications where every error is costly but nobody checks rarely pay off.
What does this mean for your company? Book an intro call
“We want to do something with AI, but what exactly?” I hear this question from managing directors almost every week. The good news: the AI projects that pay off in small and mid-sized companies are usually unspectacular. They take routine work off your team that ties up many hours today, and free up time for the work that made your company strong.
What do AI projects that pay off have in common?
From my projects I know four traits that almost always come together:
- The task comes up often. Hundreds of receipts, enquiries or reports a month, not five.
- It involves language or documents. Classic automation needs fixed rules and clean data. AI can also handle free text, changing layouts and incomplete information.
- Errors are noticeable and correctable. A person checks the result before it has consequences.
- The benefit is measurable. You know beforehand how many minutes a case takes today.
Which AI applications often pay off?
- Extracting data from documents: invoices, delivery notes, forms or application documents. The AI recognises the relevant fields even when every sender uses a different layout.
- Pre-sorting emails and enquiries: recognising the topic, forwarding to the right person, extracting the customer number and preparing a draft reply.
- Drafts for recurring texts: quotes, standard replies or reports. Your team reviews and adds to them instead of starting from scratch.
- Making internal knowledge searchable: manuals, contracts, product data and test specifications become searchable, with a source for every answer.
Which AI projects often disappoint?
- Training a custom AI model: expensive and almost never necessary. The reasons are in Do we need our own AI model?
- AI without a concrete process: a general AI strategy without a first use case produces slides, not savings.
- Full automation with high error risk: prices, commitments or legal assessments without human review are a risk, not progress.
- A chatbot for its own sake: if hardly any enquiries come in through your website, it will not pay off. More in Is an AI chatbot on your website worth it?
How do you check an AI project before it starts?
The basic calculation is the same as for any automation: cases per month, minutes per case, cost of an hour of work. Is process automation worth it? shows how. With AI, two questions are added: how good does the result need to be, and what does an error cost?
An example: 300 enquiries a month, 8 minutes each. The AI pre-sorts and writes drafts, which halves the handling time. That is 240 hours a year, or 12,000 euros at 50 euros per hour. After 2,400 euros of ongoing costs, 9,600 euros remain. If the implementation costs 10,000 euros, it pays for itself after a little over a year.
Start small, then expand
Start with one use case, build a prototype with real data and measure the benefit after a few weeks. Why this is cheaper than a big concept up front is explained in Prototype instead of specification.
Julius Launhardt works with small and mid-sized companies to find out which AI applications genuinely pay off for them, and builds them himself. More on the AI consulting page.
Frequently asked questions
From what company size does AI pay off?
Size matters less than the volume of similar cases. Even a company with twenty employees can benefit if many similar documents or enquiries arrive every day.
How quickly should an AI project pay for itself?
For manageable applications, within twelve to twenty-four months. If a project only pays off after three years or more, it is usually too large or its benefit too uncertain.
Do I need my own training data?
For most applications, existing documents, emails and a few good examples are enough. Custom training is rarely needed. It matters more that the data is accessible and reasonably up to date.
What happens when the AI makes mistakes?
You should expect it. Good solutions are built so that a person checks results before they have consequences, and errors can be corrected in a few clicks.