MIXEDArtificial intelligence for SMEs: 7 concrete uses (with real costs)
Almost every week a business owner asks us: "can AI help my company?". The honest answer is: it depends on what you do by hand today. Here are the seven uses we see working in companies of 10 to 100 people, with ballpark costs and timelines, and which one to start with.
First rule: AI solves a process, not "the company"
Projects that fail start with "we want to introduce artificial intelligence". Projects that work start with a much more boring sentence: "every day Maria spends two hours copying delivery note data into the management system". AI is useful when there is work that is repetitive, follows recognisable rules and has enough volume. If you do something three times a month, it isn't worth automating.
The 7 uses that really work
1. Answering incoming requests (emails, quotes, tickets)
An assistant reads requests coming in by email or from the website, classifies them (quote, support, supplier, spam), prepares a draft reply using your price list or FAQs, and suggests it to the person who has to send it. That person checks and clicks. Time saved: 1 to 3 hours a day for whoever manages the inbox. Cost: a project of a few weeks, a few thousand euros, plus a few tens of euros a month in model usage.
2. Extracting data from documents (invoices, delivery notes, orders, contracts)
PDFs and scans become rows in the management system without retyping anything: number, date, line items, amounts, references. It also works with documents from different suppliers, each in their own format. It's the use with the most immediate return in administration and logistics. Cost: similar to the previous one; most of the work is connecting it to your management system.
3. Searching company information (manuals, procedures, contracts, history)
"What warranty did we give customer X in 2023?" "How do you configure machine Y?" An assistant that reads your documents and answers citing the source. It reduces interruptions to expert colleagues and speeds up onboarding for new hires. Beware: quality depends on the quality of your documents; if procedures aren't written down, AI won't invent them.
4. Forecasting demand and purchasing
With 2-3 years of sales history, a model estimates how much to order and when, by item and season. It reduces both stock-outs and dead stock. Requirement: clean historical data (see our article on data strategy). Cost: higher than the previous ones, because the data needs work before the model.
5. Quality and safety control with a camera
A camera or smartphone that recognises defects on a part, checks that PPE is being worn, counts packages on a pallet. Mature technology, but it has to be trained on your cases: it needs real photos, lots of them, and a calibration period.
6. Assistant for the sales team and customer service
Before a visit, a salesperson asks "summarise customer Z's history: orders, complaints, contracts" and gets a profile in ten seconds. In support, a technician photographs a fault and receives the relevant procedures. It requires the data to be accessible (CRM, management system, tickets).
7. Reporting and management control in natural language
"What's the margin per customer over the last 6 months?" — and the answer arrives as a table and chart, without waiting for someone to prepare the spreadsheet. It's the end point of a good data foundation, not the starting point.
From dozens of Excel files to a single system. Several companies that run construction sites kept attendance, deadlines, vehicles and costs in separate Excel sheets, updated by hand by the office and site managers. With V-Site attendance comes from workers' QR clock-ins, deadlines generate their own alerts and costs update from data entered on site.
The result clients describe is always the same: hours of work recovered every week by the admin office, site managers and the workers themselves, who no longer fill in sheets. And finally reliable data on which to apply forecasts and AI assistants as a next step.
Cost and timelines: ballpark figures
| Use | Time to first result | Initial investment | Recurring cost |
|---|---|---|---|
| Replies to incoming requests | 2-4 weeks | Low | Model usage: tens of €/month |
| Data extraction from documents | 3-6 weeks | Low-medium | Tens-hundreds of €/month |
| Search across company documents | 3-6 weeks | Medium | Tens-hundreds of €/month |
| Demand/purchasing forecasts | 2-4 months | Medium-high (data work) | Hosting and maintenance |
| Visual quality control | 2-4 months | Medium-high (photo collection, calibration) | Hardware and maintenance |
| Sales/support assistant | 1-2 months | Medium (data integration) | Tens-hundreds of €/month |
| Natural-language reports | After the data foundation | Medium | Tens of €/month |
The ranges depend on how many systems need connecting and on data quality. In our article on how much an AI project costs we go into detail on what pushes the price up.
Where to start
Pick the process that currently takes up the most hours, has clear rules and whose data is already digital (emails, PDFs, management system). It's almost always use 1 or 2. Run a one-month pilot on that process, measure the hours saved, and only then expand.
If instead your data is on paper or in misaligned Excel sheets, the first step isn't AI: it's getting your data in order. We cover this in Automating business processes: where to start.
Frequently asked questions
Does AI make sense for a company with 15 employees?
Yes, if applied to a specific, repetitive process. No, as a generic project to "introduce AI".
How much does it cost to get started?
A first targeted project can be delivered in a few weeks for a few thousand euros, plus a modest monthly model usage cost.
Which use should you start with?
The one that takes the most repetitive hours with clear rules: usually incoming requests or data entry from documents.
Want to know which of these uses makes sense for your company?
In 30 minutes we analyse one of your processes and tell you whether, how and at what cost it can be automated. No commitment.