AI for business: practical guide for Dutch SMEs

What AI for business means (and what it doesn't)

AI for business isn't a magic button that reads your mind and fixes everything. It's software that can read text, extract structured data, draft replies, and make decisions based on rules you set. The practical difference for SMEs breaks down into three layers:

• Generic chatbots like ChatGPT or Google Gemini answer questions based on their training data. Useful for drafting emails or summarizing documents, but they don't know your product catalogue, pricing rules, or internal jargon unless you paste it in every time.
• Workflow automation with AI steps uses tools like n8n, Make, or Zapier to connect your apps (CRM, accounting, email) and drops an AI step in the middle. Example: when a customer email arrives, GPT reads it, categorizes the topic, drafts a reply, and posts it to your support inbox for approval.
• Custom AI agents are trained on your company knowledge: product specs, support history, compliance checklists. They handle multi-step tasks end-to-end, like reading an invoice PDF, extracting line items, checking your approval rules, and booking it into Exact Online without human input.

The ROI comes from picking the layer that fits your process. If you need a quick answer once a week, use ChatGPT. If you want to eliminate a repetitive task that happens fifty times a day, build a workflow or agent. The key is mapping the process first: who does what, where does the data live, what decision points exist. That's where our AI consultancy work starts.

Concrete AI use cases Dutch SMEs are using now

Here are four use cases we see delivering payback in under a month for Dutch SMEs. Each one names the tools involved and the back-office software it connects to.

Customer service: AI agents that read, tag, and draft email replies

Before: your support inbox gets twenty emails a day. Someone reads each one, figures out if it's a return request, a product question, or a complaint, then drafts a reply. That's two to three hours every day.

After: an n8n workflow watches your inbox. When a new email arrives, GPT reads it, tags it (return, question, complaint), checks your knowledge base for the answer, and drafts a reply in your tone of voice. Your team reviews and sends in under a minute. Time saved: fifteen hours per week for a typical webshop.

Tools: n8n or Make for the workflow, GPT or Claude for the reading and drafting, your existing email platform (Gmail, Outlook). No new software for your team to learn.

Bookkeeping: scan invoices and post them to Exact Online or Moneybird

Before: invoices arrive by email or post. Someone opens the PDF, reads the supplier name, amount, VAT rate, and invoice number, then types it into Exact Online or Moneybird. Five minutes per invoice, fifty invoices a week, that's over four hours.

After: a workflow grabs the PDF from your inbox, sends it to GPT Vision or a document AI model, extracts the fields, matches the supplier to your contact list, and posts the entry to your accounting system via API. Your bookkeeper reviews a list of pre-booked entries instead of typing them one by one. Time saved: three to four hours per week.

Tools: n8n, Make, or Zapier for orchestration, GPT Vision or a specialized OCR tool for extraction, Exact Online or Moneybird API for booking. We build these integrations as part of our business automation service.

Quotes: from CRM data to PDF in one click

Before: a lead asks for a quote. You open your CRM (Pipedrive, Teamleader, HubSpot), copy the contact details and project notes into a Word template, calculate the price based on your rate card, export to PDF, and email it. Fifteen to twenty minutes per quote.

After: you click a button in your CRM. A workflow pulls the contact data, product specs, and pricing rules, sends them to GPT to generate the quote text in your house style, renders it as a PDF, and emails it to the lead. Done in under a minute. Time saved: two to three hours per week for a consultancy or agency that sends ten quotes a week.

Tools: your CRM's API, n8n or Make, GPT for text generation, a PDF renderer, your email platform.

Email triage: who needs to read this, and what's the deadline?

Before: your shared inbox (info@, sales@, support@) gets fifty emails a day. Someone reads each one and forwards it to the right person. That person reads it again to figure out if it's urgent. Thirty minutes of duplicate work every day.

After: an AI agent reads every incoming email, tags it by department (sales, support, finance), extracts any deadline or urgency signal, and routes it to the right Slack channel or person with a one-line summary. Your team sees 'New return request from [name], wants refund by Friday' instead of reading the full thread. Time saved: two hours per week.

Tools: n8n or Make, GPT or Claude for reading and tagging, Slack or Microsoft Teams for routing.

Start with the process that costs you the most hours per week, not the one that sounds coolest. Map it end-to-end first, have someone with integration experience build the first version, and only then trust the workflow with real customer data. Without that experience, you risk an automation that breaks on the first edge case or quietly produces AI errors you only catch weeks later.

Running into this at your own business? We'll spend 30 minutes with you for free, no sales pitch. Book a free intro call

Why most SME AI projects fail (and how to avoid it)

We see three anti-patterns kill most Dutch SME automation projects before they deliver ROI.

Mistake one: buying a generic AI tool without mapping the process first. You sign up for a chatbot platform or an AI assistant, spend two weeks configuring it, then realize it can't handle your approval rules or connect to your ERP. The bottleneck was never the tool, it was the fact that your process (who approves what, when does finance need to see it) was never written down. Fix: spend one hour drawing the workflow on paper before you buy anything. If you can't explain the process to a colleague in five minutes, an AI won't magically figure it out.

Mistake two: assuming the AI will learn your business on its own. GPT knows general knowledge, it doesn't know your product specs, pricing tiers, or internal jargon. If you don't feed it clean documentation, it will guess, and half the guesses will be wrong. Fix: before you automate, collect your knowledge base (product docs, support FAQs, pricing sheets) in one place. Clean it up, remove contradictions, then feed it to the AI. Eight out of ten custom GPT failures we see trace back to messy source documents, not the model.

Mistake three: skipping the integration step. You build an AI tool that lives in a separate tab or app. Your team has to remember to open it, copy data in, copy the result out, paste it into the real system. Within a week, they stop using it because it adds steps instead of removing them. Fix: wire the AI into the tools your team already opens every day. If they live in Exact Online, the AI should write directly to Exact Online. If they live in Slack, the AI should post to Slack. No new logins, no context switching.

This is why we start every project by mapping the current process, cleaning the knowledge base, and designing the integration points before we write a single line of automation code. It's slower up front, but it's the only way to get a tool your team will actually use three months later.

AVG, NIS2, and security: using AI without legal risk

Dutch SMEs need to stay compliant with two big regulations when they use AI: AVG (the Dutch/EU GDPR) and NIS2 (the cybersecurity directive that hit in October 2024).

AVG basics for AI: If your AI processes personal data (customer names, email addresses, purchase history), you need a processing agreement with your AI vendor, you must log what the AI reads, and you need to be able to explain any automated decision to the person affected. Practical steps: use AI vendors that offer EU data residency (OpenAI, Anthropic, and Google all have EU regions), turn off model training on your data (most enterprise plans let you do this), and write a one-page data-processing addendum that lists what data the AI sees and why. The Autoriteit Persoonsgegevens has template language you can adapt.

NIS2 impact: If you're in a critical supply chain (energy, transport, healthcare, food, digital infrastructure), NIS2 requires you to document your access controls, report breaches within 24 hours, and show you've assessed third-party risks. That includes your AI vendors. Practical step: ask your AI vendor where your data is stored, how they handle breaches, and whether they're ISO 27001 certified. If they can't answer those three questions, walk away.

Data minimization: Don't send the AI more data than it needs to do the job. If you're automating invoice booking, send the invoice PDF, not your entire customer database. If you're drafting email replies, send the current thread, not the last five years of correspondence. Less data in the AI means less risk if something goes wrong.

If your AI vendor can't answer 'where does my data live?' in one sentence, that's a red flag. EU-based SMEs should default to vendors that offer EU data residency and let you turn off training.

Custom AI agents vs. standard tools: when to choose what?

Here's the decision framework we use with Dutch SME clients.

Use a standard tool (ChatGPT Plus, Zapier AI, Google Gemini in Workspace) when:

• The task is generic: summarize this document, draft a reply, translate this text.
• You're happy to adapt your process to fit the tool's limits.
• You need an answer today, not in two weeks.
• The data you're working with is public or low-sensitivity.

Standard tools are faster to start and have a lower entry threshold than custom development, but they hit a ceiling fast. If you need the AI to know your pricing rules, compliance checklist or product catalogue, you'll spend more time copying and pasting context than you save. They also still need someone who can write good prompts, validate the output and watch for drift, otherwise the same drag-and-drop tooling produces silent failures.

Build a custom GPT or AI agent when:

• The AI needs to know company-specific knowledge: your product specs, internal jargon, approval workflows, pricing tiers.
• You want it to work across multiple tools without manual copy-paste (read from CRM, write to accounting, post to Slack).
• The task repeats fifty times a week and every minute saved compounds.
• You need audit logs, role-based access, or compliance controls that standard tools don't offer.

Custom agents take more preparation and specialist hours up front, but they fit your process exactly and scale without adding headcount. Building a reliable custom GPT isn't the same as writing a ChatGPT prompt: you have to curate the knowledge base, ground the model in fixed sources to prevent hallucinations and monitor when its output drifts. That's where AI tuning experience earns its build cost back. We deliver these through our custom GPT service: you give us your knowledge base and workflow map, we build the agent, wire it into your tools and hand over the keys.

The break-even point: if a standard tool saves you a few hours a week but you spend almost as much working around its limits, a custom agent that saves more time with no workarounds pays back fast. Just budget for maintenance and monitoring alongside the build, because an agent that stops running halfway through year one has no ROI left.

Conclusion

AI for business works when you start small, pick a process that already costs you hours every week, and wire the AI into tools your team actually uses. Don't buy a tool before you map the workflow. Don't assume the AI will learn your business by magic. And don't skip the compliance step: AVG and NIS2 are real, and the fines are bigger than the cost of doing it right the first time.

If you're not sure where to start, map one workflow end-to-end and ask: which step could a machine do faster? That's your first automation candidate. Do the analysis and prioritisation yourself, then bring in a partner who has built the same stack many times for integration, AI tuning, monitoring and maintenance. The ROI comes from repetition, not complexity, and from automations that keep running long after the launch week.