Implementing AI in your business: an SME roadmap

For an SME, implementing AI mainly means choosing one specific task that costs time today, honestly checking the data behind it and then testing a small pilot inside the real workflow. Only once the pilot holds up in daily business does it move into regular operation, with clear ownership, training and support. Companies that start by picking a tool instead of looking at the process usually end up with a demo that nobody uses for long.

Why do AI projects in SMEs often stall at the experimenting stage?

Many AI initiatives in small and mid-sized companies start with the question “What can we do with AI?” rather than “Where are we losing time today?”. The result is a chat tool that a few employees use to polish texts, while the actual workflows remain unchanged. That rarely produces a benefit anyone in the business notices.

The second common reason is the state of the data. A language model can only work with what can be found and is unambiguous. If quotes, bills of materials or maintenance logs exist in several versions spread across folders, mailboxes and individual PCs, even a good model will not give reliable answers. The problem then is not the AI but the filing system in front of it.

How do you find the right first AI use case?

The right first AI use case is a recurring task with clear inputs and a result that a person can check quickly. A good starting point is a walk through the business: where do people retype, search, sort, or answer the same question over and over? Such spots are better candidates than big visions, because the difference before and after can be observed directly. Which use cases hold up in everyday SME work, with their prerequisites and limits, we describe in a separate article.

• The task occurs regularly, not just a few times a year.
• The inputs are digital or can be digitised with reasonable effort, such as receipts, emails or forms.
• Mistakes can be spotted and corrected before they cause harm.
• Someone in the business can judge the result professionally and owns the process.
• The task is tied to a system that can be integrated, rather than creating yet another isolated tool.

Less suitable as a starting point are tasks where a mistake immediately reaches the outside world, such as quotes sent automatically without review, or tasks for which the business has no clear rule. If people already disagree about what the right result is, an AI cannot make that decision for them.

What data does an AI pilot actually need?

An AI pilot does not need all of the company's data, only the documents required for the chosen task – in a form where it is clear which version is valid. For a knowledge search across manuals, that means current manuals in one place, with old versions marked or removed. For reading incoming invoices, it means one inbound channel through which receipts reliably arrive.

This groundwork is unspectacular, but it carries the project. Taking stock clarifies where documents are stored, who maintains them and which information so far only exists in people's heads. Even if no AI tool is introduced in the end, the business is better organised afterwards than before.

What does a realistic AI roadmap for an SME look like?

A realistic AI roadmap for an SME follows the same order as any good IT project: understand first, then decide, then test small, then operate properly. None of the steps is elaborate. If you keep to the order, you end up with a tool used in daily work rather than a presentation in a drawer.

1. 01Look at the workflow: walk through the chosen task on site together with the people who do the work today – using real receipts, emails or documents.
2. 02Check the data: where are the necessary documents, which version is valid, who maintains them, and may this data be sent to an external service?
3. 03Define the goal: describe in advance what a usable result is and who reviews it, so the pilot can be measured against something.
4. 04Choose the operating model: cloud service, hosting in the EU or running it on your own premises – depending on how sensitive the data is.
5. 05Pilot in the real workflow: a small team uses the tool for a limited period in daily business, while people continue to review the results.
6. 06Evaluate and decide: where was the result usable and where not, and was that down to the tool or to the underlying data?
7. 07Regular operation: integration with existing systems, permissions, logging, training and named responsibility for operation and further development.

How do you run an AI pilot without disrupting operations?

An AI pilot does not disrupt operations if it runs alongside the existing workflow and the AI initially only makes suggestions. Accounting keeps entering invoices itself but compares its entries with what the tool extracted. Sales receives a draft quote but only sends what it has checked. This shows how often the AI gets it right without any mistake reaching customers.

The pilot should take place in the tools people already use. An additional program that has to be opened separately tends to be forgotten soon. When the result appears where the work happens – in the mailbox, the ERP system or the ticketing system – it is more likely to be used.

Who in the business is responsible for the AI?

AI in a business needs two clearly named responsibilities: a business one and a technical one. Business responsibility lies with the person who knows the workflow and can judge whether the results are correct. Technical responsibility covers operation, access rights, updates and incidents, and can sit with an IT partner.

• Who decides which data the AI is allowed to see?
• Who spot-checks whether the results still fit when documents or workflows change?
• Who can be contacted when the tool is unavailable or delivers obviously wrong results?
• Who documents which AI services are used in the business and for what?

How should you train employees to work with AI?

Employees learn to work with AI best on their own use case, not in a generic presentation. Good training shows what the tool does reliably, where it goes wrong and how to recognise mistakes. That explicitly includes the fact that language models phrase wrong information just as convincingly as correct information, and that responsibility for the result stays with the person.

Clear rules for using public AI services belong here too: which data may be entered, which services are approved, and whom to ask when in doubt. Such rules fit on a single page and protect a lot when everyone knows them. Which data may go into which kind of AI service is covered in our article on AI and the GDPR.

How does NDVDL take over AI implementation in your business?

NDVDL starts AI projects with an appointment at your business, not with a tool demo. We let the employees who do the task today show it to us, and we look at where the documents are stored and which systems are involved. You then receive a written proposal with a tightly scoped first use case. If we think the project is premature, we say so openly and recommend organising the filing first.

We then implement the pilot and connect the solution to your existing systems instead of building another island. In regular operation we take care of access rights, updates, monitoring and maintenance. Where the data requires it, we also run AI applications on your own premises or with a provider operating a data centre in the EU. You have one fixed contact person for technology and workflow.