According to Statistics Netherlands (CBS), 22.7% of Dutch businesses with 10 or more employees used one or more AI technologies in 2024. That is an average, with big differences between sectors and company sizes. Whether an AI project pays off after six months or ends up in a drawer rarely depends on the model. It depends on which task you choose to tackle with it.
Below are the ten applications with the best prospects for a typical small or mid-sized business (10 to 250 employees, Dutch market). For each one you read what it is, who it suits, what to measure, which tools you need, where it goes wrong and roughly what a first working version costs. Amounts and lead times are indications: what it costs you depends on the integrations and the amount of work.
Want to know what AI is first? Read what is AI or what is an LLM. For the difference between agents and chatbots: AI agent vs chatbot.
The ten applications at a glance
| # | Application | What you measure | First version |
|---|---|---|---|
| 1 | Chatbot on your own knowledge base | share of questions resolved without a person | 2 to 4 weeks |
| 2 | Content production in fixed steps | time per product page or article | 3 to 6 weeks |
| 3 | Sorting mail and drafting replies | time per day in the shared inbox | 2 to 3 weeks |
| 4 | Voice agent for inbound calls | missed calls and appointments that fall through | 3 to 5 weeks |
| 5 | Invoice processing | minutes per invoice and number of corrections | 4 to 8 weeks |
| 6 | Assessing enquiries with a smart intake | time from enquiry to quote | 2 to 4 weeks |
| 7 | Meeting summaries and action items | time for minutes and follow-up | immediate or 2 to 4 weeks |
| 8 | Market research and competitor analysis | hours per research question | immediate or 2 to 4 weeks |
| 9 | Personalised email campaigns | opens, clicks and unsubscribes per segment | 4 to 8 weeks |
| 10 | Searching your own documents | time to an answer with a source | 4 to 10 weeks |
1. Chatbot on your own knowledge base
What it is. A chatbot that answers from your documentation, FAQs, product information and internal procedures. The technique is called Retrieval-Augmented Generation (RAG): the system first searches your sources and only then writes an answer, with a reference to where it came from. That makes it invent less than a general chatbot.
Who it suits. Web shops with many product questions, software companies with an overloaded support desk, service providers who get the same questions every week, and municipalities with a busy front desk. Especially useful when customers expect answers in the evening too and you have no team working then.
What you measure. The share of questions the bot resolves well on its own, how often it hands over to a staff member and how satisfied customers are with the answer. Also measure how many tickets your team gets per week before and after.
Tools. For searching your sources: Pinecone, Weaviate or Postgres with pgvector. As the model: Claude from Anthropic or an OpenAI GPT model. As the interface: a widget on your site or an integration with Intercom, Zendesk or HubSpot.
Where it goes wrong. Poor source documents give poor answers. An outdated FAQ stays outdated, with or without AI on top. Without a handover to a staff member for difficult questions, customers get frustrated. And price agreements are made by a person, never by the bot.
Lead time and budget. A first working version goes live in 2 to 4 weeks, from about €5,000 for a basic setup. More on this application at /ai-klantenservice.
2. Content production in fixed steps for product descriptions and blogs
What it is. An automated chain that turns input (product data, a brief, a keyword) into a product description, blog post or social media post in your brand voice. It is more than one prompt: between the steps there are checks on tone, findability and facts, and an editor at the end.
Who it suits. Web shops with hundreds of products, marketing agencies producing a lot of content, B2B suppliers with a large range, and publishers.
What you measure. How much time a product page or article takes before and after, how much the editor still has to change and how the pages perform in search. McKinsey calculated that about 75% of the value of generative AI sits in four areas, and marketing and sales is one of them.
Tools. Claude or a GPT model for the writing, a layer in Python or TypeScript that runs the steps, an integration with your CMS (Shopify, WooCommerce, Sanity or Contentful) and a brand voice document that goes into every prompt.
Where it goes wrong. AI text without editing reads flat and repeats itself. AI knows nothing about your business, so without a good brief you get generic text. And Google rewards content that helps people, not volume. Quality comes before quantity.
Lead time and budget. 3 to 6 weeks for a working chain, €8,000 to €20,000 depending on complexity and integrations. See /ai-content for the approach.
3. Sorting mail and drafting replies
What it is. Incoming mail is classified automatically (quote request, support, invoice question, job application, spam) and gets a draft reply based on earlier similar mails. A staff member reads it, edits it if needed and sends it.
Who it suits. Any business with a shared inbox that gets a lot of mail every day: customer service, sales, finance, HR and the general info address.
What you measure. How much time your team spends in the inbox per day, how quickly an urgent mail reaches the right person and how often a draft can be sent unchanged. The gain often sits more in the sorting than in the drafts: you see straight away what is urgent.
Tools. The Microsoft Graph or Gmail API for inbox access, Claude or a GPT model for classifying and writing, rules for routing and an overview to spot patterns. For simple cases Power Automate or Zapier will do; for more complex ones, custom code.
Where it goes wrong. Privacy: not everything in a mailbox may be read by an AI model. Do a privacy assessment first, especially for HR or legal inboxes. Second pitfall: staff who press send without reading.
Lead time and budget. 2 to 3 weeks for a first version, €4,000 to €8,000. Works best combined with application 6.
4. Voice agent for inbound calls
What it is. An AI voice that answers the phone, responds to questions, books appointments or puts the caller through to the right person. It feels like a conversation, not a phone menu. Platforms such as ElevenLabs and Vapi make this possible.
Who it suits. Tourism (B&Bs, hotels, attractions), hospitality, hairdressers, dentists, physiotherapists, garages and other service providers that book many appointments. Also for businesses with many foreign customers that cannot staff a reception all day.
What you measure. How many calls you miss now, especially outside office hours, how many appointments fall through and how often the agent completes a call well without a person stepping in. Automatic confirmations and reminders are often where you see a difference first.
Tools. Vapi or Retell to run the conversation, ElevenLabs for the voice, Twilio for telephony and an integration with your calendar (Google Calendar, Outlook or a booking system).
Where it goes wrong. A noticeable pause between question and answer feels unnatural. A poor calendar integration causes double bookings. And not every call belongs with an AI: complaints and negotiations stay human work.
Lead time and budget. 3 to 5 weeks, €6,000 to €15,000 depending on integrations. Worked out for tourism at /ai-toerisme, and the approach at /voice-agents.
5. Invoice processing and bookkeeping
What it is. AI reads incoming invoices (PDF, mail or paper), extracts the data (supplier, amounts, VAT, invoice date, general ledger account), drafts a booking and prepares it in Exact, Twinfield, Yuki or Moneybird. A staff member approves or corrects it, and you use those corrections to improve the system.
Who it suits. Businesses that process hundreds of invoices a month, in construction, retail, hospitality, manufacturing or healthcare, and bookkeeping and accountancy firms themselves.
What you measure. How many minutes an invoice takes now, how often a proposal needs correcting and how that changes over the first months.
Tools. For reading: the vision capabilities of Claude or OpenAI, or specialised tools such as Klippa and Basecone. For the integration: the APIs of Exact, Twinfield or Yuki.
Where it goes wrong. Poor scans of paper invoices. Invoices with an unusual layout, for example from small freelancers, make extraction fail. And responsibility for the books stays with the accountant: AI is a tool.
Lead time and budget. 4 to 8 weeks, €10,000 to €25,000. Specifically for accountancy and bookkeeping firms: /ai-accountants.
6. Assessing enquiries with a smart intake
What it is. A conversational intake that asks follow-up questions based on what the visitor fills in, estimates how promising the enquiry is, puts the data into your CRM and sends urgent enquiries straight on through Slack or Teams. A warm enquiry gets a calendar link, a cold one a series of follow-up mails.
Who it suits. B2B service providers (consultants, agencies, IT companies), software companies and suppliers of complex products, where you want to know what someone is looking for before sales calls. Also for architects, lawyers and accountants.
What you measure. The time between enquiry and quote, the share of enquiries that reach sales and how many of those become customers.
Tools. A custom form in Next.js or Webflow with Claude or a GPT model behind it, integrations with HubSpot, Pipedrive or Salesforce, calendar booking through Cal.com or Calendly, and notifications in Slack or Teams.
Where it goes wrong. Too many questions. An intake must not feel like an interrogation, so stop after five to seven questions. And scoring without explanation: people drop out when they feel they are being filtered out.
Lead time and budget. 2 to 4 weeks, €5,000 to €12,000.
7. Meeting summaries and action items
What it is. A meeting in Teams, Zoom or Meet is recorded, transcribed and summarised in a fixed structure (decisions, action items, open questions, deadlines), and the result lands in the right place: Notion, Slack, Asana or email. Nobody has to take minutes any more.
Who it suits. Anyone in an organisation with many meetings: management teams, sales, project managers, consultants, lawyers and accountants. Especially useful for client calls and project kick-offs.
What you measure. How much time minutes and sending out action items take now, and whether action items get done on time. That last point is often the biggest gain, because they land in someone's task list automatically.
Tools. Ready-made: Granola, Fathom, Otter or tl;dv, or Microsoft Copilot in Teams. Custom: Whisper for transcription, Claude for the summary and n8n or Make to put everything in the right place. More in AI for meeting notes.
Where it goes wrong. Consent: participants must know they are being recorded, and some conversations (legally sensitive, HR) should not be recorded at all. Storage and retention fall under the GDPR. And a summary does not replace being present at an important meeting.
Lead time and budget. Ready-made tools work straight away. Custom work takes 2 to 4 weeks and €4,000 to €10,000.
8. Market research and competitor analysis
What it is. Instead of hours of searching, reading websites and keeping spreadsheets, you let a research tool draw up a report: competitors, their positioning, prices, content, reviews and gaps in the market. The research features in Gemini, ChatGPT and Perplexity look up sources themselves and cite them.
Who it suits. Marketing teams, product managers, founders preparing a launch, advisers on acquisitions and policy makers. Anyone who regularly wants to know what others are doing.
What you measure. How many hours a research question takes now, and how much time you still spend checking and adding to the result. Local knowledge, conversations and judging what really matters stay human work.
Tools. For broad research: Gemini Deep Research, ChatGPT with search and Perplexity. For specialist work: a custom agent with search tools and connections to sources such as Crunchbase or Dealroom.
Where it goes wrong. AI sometimes makes things up, especially numbers and quotes. So always check the sources. What is not online (internal data, gated content), AI will not find. And the judgement stays with someone who knows the context.
Lead time and budget. Ready-made tools work straight away, for around €20 to €40 per user per month. A custom workflow takes 2 to 4 weeks and €4,000 to €10,000. See /ai-strategie for research as part of a strategy.
9. Personalised email campaigns per segment
What it is. Instead of one campaign to your whole list, each segment gets its own version. AI writes the variants based on data per segment (sector, behaviour, purchase history), and your email platform (Klaviyo, Mailchimp, ActiveCampaign or HubSpot) tests which version works best per segment.
Who it suits. Web shops with a large customer base, B2B companies with a big newsletter list, software companies with emails for new and existing users, media and event organisers.
What you measure. Opens, clicks, unsubscribes and revenue per segment, compared with your current campaigns. With a small list the differences are too small to draw conclusions from.
Tools. Klaviyo, Mailchimp, ActiveCampaign or HubSpot for sending, Claude or a GPT model for the variants and a place where your customer data comes together for segmentation.
Where it goes wrong. Variants without a brand voice check produce confusing emails. Without a final edit, an odd sentence will one day land in thousands of inboxes. And the GDPR limits profiling, especially with sensitive characteristics.
Lead time and budget. 4 to 8 weeks, €8,000 to €20,000.
10. Searching your own documents
What it is. You ask a question of your own documents (contracts, policies, project files, legal opinions) and get an answer with a reference to the source. It is the same technique as application 1, but for internal knowledge instead of customer questions.
Who it suits. Law and accountancy firms, engineering firms, consultancies and any business with a lot of knowledge stored in old quotes, reports and files.
What you measure. How long it takes now to find an answer, how often the system cites the right source and how quickly new staff find their way. They can ask the system what they would otherwise ask a colleague.
Tools. A vector database (Pinecone, Weaviate or Postgres with pgvector), embeddings from OpenAI or Voyage AI, a way of splitting documents that suits your document types and Claude or a GPT model for the answer. For larger organisations: Glean or Microsoft Copilot.
Where it goes wrong. Version control: old versions of documents give old answers. Access rights: not everyone may see every document, so those rights must carry over one to one into the search system. And confidential documents do not belong in a cloud model by default. Many organisations with sensitive data require such a solution to run in the EU. The GDPR allows storage outside the EEA with the right safeguards (Dutch Data Protection Authority), so this is their own choice.
Lead time and budget. 4 to 10 weeks, €8,000 to €30,000 depending on scale and security requirements. See /ai-data for broader data applications and /ai-agents if you want to move towards agents.
Start with one application
An AI project in a small business gets stuck more often on doing too much at once than on the technology. So start with one application, do it well, measure what it delivers, and only then take on the next one. Application 1 (chatbot), 3 (sorting mail) and 5 (invoice processing) are a good start for many small businesses: you see results quickly, the rest of the organisation hardly notices and the gain is easy to measure.
What they have in common: they shift work. The repetitive part goes to AI, the part that needs judgement stays with people.
Want to know which application pays off most in your situation? Take the AI Readiness Scan for a first analysis with the most promising applications for your sector.
Curious what AI can do for your business?
Take the free AI Scan and find out in 1 minute.
Frequently asked questions
- What are the best AI use cases for SMEs?
- For many small businesses: a chatbot on your own knowledge base, invoice processing, sorting mail with draft replies, content production in fixed steps and assessing enquiries with a smart intake. What they share: the work is repetitive, takes a lot of time and starts with digital input. Choose based on where most time goes, not on the newest model.
- Which AI use case pays off fastest?
- For most small businesses a chatbot on your own knowledge base, sorting mail or invoice processing: you see results quickly, the rest of the organisation hardly notices and the gain is easy to measure. Measure beforehand how much time the process takes now, so that after a few weeks you know what it delivers.
- How long does a first working AI version take?
- For a well-defined application usually two to four weeks until a first version that a small group of users works with for real. Then you measure what it delivers and expand what works. If it takes much longer, the scope is too wide: one process, not three at once.
- Why do AI projects fail in small businesses?
- Rarely because of the model. More often because of the choice and the execution: too wide a scope, poor source documents, no handover to a person for difficult cases, or a tool chosen before the problem was clear. Start with the process that demonstrably takes too much time and measure from day one.
