AI agents for business: what they can really do
AI agents already take work off your hands: customer emails, invoices, stock and questions about your numbers. What really works, why so many projects stall, and how do you use AI so it saves you time and money?
Updated 11 min read
In short
- An AI agent does work, not small talk: it reads an email, looks up the order in your systems and prepares the reply. A chatbot only talks, an automation follows fixed steps.
- The best results come from a combination: fixed steps where possible, AI where understanding is needed. In a test with CRM tasks, agents reached over 83% on tasks with fixed steps, against around 35% on average in longer conversations.
- Your own data makes the difference. An AI that talks to your till, webshop and accounting answers questions about your sales, warns about low stock and books your invoices.
- The pressure is real: 64% of Dutch businesses have staff shortages and 41% of business owners spend at least half the week on admin.
- Start with one task with a clear gain. With a specialised partner, AI projects reach deployment twice as often as when companies build everything themselves.
What is an AI agent?
An AI agent is software that carries out a task from start to finish and decides for itself which steps it needs. It reads a customer email, looks up the order in your webshop, writes a reply with the real status and gets it ready for you. OpenAI describes agents as systems that independently accomplish tasks on your behalf, and adds straight away that a simple chatbot is not an agent1OpenAI, 2025A practical guide to building agentsOpen the source.
Anthropic, the maker of Claude, draws a useful line. In a workflow, the AI follows a route that was mapped out in advance. In an agent, the model decides which steps and which tools to use2Anthropic, 2024Building effective agentsOpen the source. In practice there are three flavours:
| Chatbot | Automation | AI agent | |
|---|---|---|---|
| What it does | Answers questions in a conversation | Runs fixed steps, the same every time | Chooses its own steps to reach a goal |
| Who sets the route | There is no route | You, in advance | The AI model, along the way |
| Good at | Frequently asked questions | Repetitive work with clear rules | Tasks that are slightly different each time |
| At its best | As the first point of contact | With one AI step in it, like reading an invoice | With clear limits and a person checking |
Anthropic advises choosing the simplest solution that works, and only making it more complex when needed2Anthropic, 2024Building effective agentsOpen the source. That’s how we build too: fixed steps where possible, AI where it really adds something. It’s faster, cheaper and more reliable than leaving everything to an agent.
Why businesses are starting now
AI has become normal in a short time. According to Statistics Netherlands (CBS), 17% of businesses with two or more people used AI in 2025, up from 13% in 2024 and 8% in 2023. The bigger the business, the more often: from 14% of businesses with 2 to 10 people to 66% of those with 250 or more3CBS, 2025Bedrijven gebruiken AI vaakst voor marketing of verkoopOpen the source. Smaller businesses are behind, and that’s the opportunity: whoever starts now will soon work faster than the competition.
17%
of Dutch businesses used AI in 2025, twice as many as in 20233CBS, 2025Bedrijven gebruiken AI vaakst voor marketing of verkoopOpen the source
29.8%
of SMEs with 10 to 250 people used AI technology4CBS, 2026Gebruik van AI-technologie door Nederlandse microbedrijvenOpen the source
20%
of EU businesses with 10 or more employees used AI in 20255Eurostat, 202520% of EU enterprises use AI technologiesOpen the source
The pressure to act is high. In spring 2026, 64% of Dutch businesses had staff shortages, and 29.7% of all businesses named more automation as their main measure6CBS (Conjunctuurenquête Nederland, met KVK, EIB, MKB-Nederland en VNO-NCW), 2026Bedrijven zetten meer in op automatisering door personeelstekortOpen the source. Add the paperwork: 41% of Dutch business owners spend at least half their week on admin7Visma & Kantar (via Accountant.nl), 2025Veel ondernemers besteden de helft van hun tijd aan administratieve takenOpen the source. That is exactly the kind of work AI takes over.
Big companies are pulling ahead. In McKinsey’s 2025 survey, 23% of organisations said they were scaling an agent somewhere, and 39% were experimenting with them8McKinsey & Company, 2025The state of AI in 2025: Agents, innovation, and transformationOpen the source. A year later, scaling rose from 27% to 40% at organisations with more than a billion dollars in revenue, while it stayed at 22% at smaller ones9McKinsey & Company, 2026The state of AI in 2026: On the road to ROIOpen the source.
What holds smaller businesses back is mostly a lack of experience (73%), privacy (49%) and uncertainty about the legal consequences (42%)3CBS, 2025Bedrijven gebruiken AI vaakst voor marketing of verkoopOpen the source. In a survey by the Dutch Chamber of Commerce (KVK), half of business owners didn’t know the European AI Act10KVK, 2025KVK onderzoek: ondernemers zetten nauwelijks stappen rond Europese AI-wetgevingOpen the source. Those are exactly the things a good partner sorts out for you.
What AI already does for your business
AI pays off most on tasks with a clear goal, data it can look up and an outcome you can check. For a small business, for example:
- Answering customer emails. The AI reads the question, looks up the order or appointment and writes a reply with the real details. You approve it with one click.
- Processing invoices. Reading supplier invoices, checking them and putting them in Exact or Moneybird. You only look at the exceptions.
- Questions about your own numbers. “What did we sell last week, and why less than last year?” The AI pulls it from your till, webshop and books and explains it.
- Reordering stock on time. The AI forecasts from your sales what runs out soon and prepares the order for your supplier.
- Scheduling and preparing quotes. Emailing back and forth, checking your calendar, and drafting a quote from your price list.
And it’s moving fast. The research organisation METR measures the length of tasks AI models finish half the time. That length has doubled roughly every seven months since 201911METR (arXiv), 2025Measuring AI Ability to Complete Long Software TasksOpen the source. The more AI can take on, the more it matters how it’s set up: with access to your data, clear limits and a person who checks.
Pick a task
- Automation
- AI assistant
- AI agent
AI agent, with your approval
- What it does
- Reads the email, looks up the order in your shop and writes a reply with the real status.
- What you still do
- Approve the draft before it goes out. Once it's reliable, only the tricky ones.
- Freedom for the AI
- Some: it chooses within limits
- Connected to
- EmailWebshop
AI in your own portal
The biggest gain isn’t a standalone chatbot on your website, but AI that works with your data. We build that into a portal: one place where your sales, stock, invoices and customer questions come together, with AI in it that prepares the work.
- Ask your data. “What did we sell today compared to last week?” and get an instant answer from your till and webshop
- Stock alerts before something runs out, with the order already prepared for your supplier
- Drop or photograph invoices, and they’re read into your accounting
- Draft replies to customer emails with the real order status, which you approve
- Everything traceable: what the AI did, and who approved what
This is already running in practice. At a hospitality business, the owner forwards supplier invoices to one address. The AI scans them, checks for duplicates, and after one click they’re in Exact Online. He asks the portal what last week’s revenue was, and the AI only proposes changes to his Google Sheet: they go through once he confirms, and every change is logged. And at a holiday villa, a chat assistant on the website answers guests’ questions with the real availability and prices from the booking system.
Click through the portal demo of a bakery to see how it works.
What works, and what doesn’t
The numbers show clearly where the difference lies. In CRMArena-Pro, a test with realistic sales and customer service tasks, the best agents succeeded on average about 58% of the time on a single question and around 35% in a longer conversation. But on tasks where the steps were laid down, they reached over 83%. They also showed little sense of what was confidential12arXiv, 2025CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and InteractionsOpen the source. The lesson: give AI clear steps and limits, and it becomes much more reliable.
83%
succeeded on tasks where the steps were laid down, in a test with realistic CRM tasks12arXiv, 2025CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and InteractionsOpen the source
58%
succeeded on average on a single question12arXiv, 2025CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and InteractionsOpen the source
35%
succeeded on average in a longer conversation12arXiv, 2025CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and InteractionsOpen the source
Gartner expects over 40% of agent projects to be cancelled by the end of 2027, because of escalating costs, unclear value or poor risk controls. It also warns about agent washing: existing chatbots and automations with a new label. Of the thousands of vendors, Gartner estimates only about 130 are real13Gartner, 2025Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027Open the source. So the difference is in the approach: one clear task, a calculated gain, and someone who actually builds.
A much-discussed MIT NANDA report points the same way. Tools that companies deployed with a specialised partner reached deployment about 67% of the time, against roughly 33% for tools companies built entirely themselves. The report is preliminary and relies on interviews and a survey, but the direction is clear14MIT NANDA, 2025The GenAI Divide: State of AI in Business 2025 (preliminary findings)Open the source.
How to start
- Pick one task that comes back often. Count how many hours it takes per week. Five hours a week is already over 250 hours a year.
- Combine fixed steps with AI. Use AI on the step that needs understanding, like reading an invoice or understanding an email, and let the rest follow fixed rules.
- Give only the access it needs. Let the AI only look at first. Only once it proves reliable may it change or send anything.
- Keep a person checking. Start by approving before anything goes out. Later you only look at the exceptions.
- Measure the result. Time saved, number of errors and what customers think.
- Work with a partner who knows your systems. The value is in the connection to your webshop, till and accounting. That’s custom work.
This is how we approach it: we start with the task that brings in the most, show you how it works with your own data, and expand once it has proven itself. Read more about automations.
The rules: AI Act and privacy
Since February 2025, the European AI Act has asked businesses that use AI to ensure AI literacy among their staff17Official Journal of the European Union, 2024Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act)Open the source. Since July 2026 that has been eased: you must take measures to support that knowledge, without guaranteeing a specific level, and the EU and member states must help SMEs in particular18Official Journal of the European Union, 2026Regulation (EU) 2026/1744 (Digital Omnibus on AI)Open the source.
Since 2 August 2026 there is also a transparency duty: people must know they are interacting with an AI system, at the latest at the first contact, unless it is already obvious17Official Journal of the European Union, 2024Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act)Open the source19European Commission, 2026Commission publishes guidelines on transparency obligations for providers and deployers of certain AI systemsOpen the source. Formally that duty lies with whoever makes the system, but the Dutch Chamber of Commerce advises business owners to simply say so at the first contact when customers talk to an AI assistant20KVK, 2026Moet je laten weten dat je AI gebruikt in je bedrijf?Open the source. We build that in by default.
- Tell customers they are talking to an AI assistant
- Make sure your team knows what the AI can and can’t do
- Choose a provider that processes your data in the EU, and sign a data processing agreement
- Give the AI access only to the data it needs
- Keep a person reachable for customers who want one
What does it cost, and what does it bring?
The cost is rarely in the AI itself. You pay for a language model per piece of text it processes, and with a few hundred emails or invoices a month that is usually a small item. The real work is in the connections to your systems, the rules for what the AI may and may not do, and the upkeep when your software changes.
Escalating costs are the first reason Gartner gives for agents being cancelled13Gartner, 2025Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027Open the source. So work out in advance what the task costs you in hours now, and what it brings in when those hours free up. At Forwarddd you get a proposal with the price up front, and a web app or customer portal starts at €1,450 one-off.
Frequently asked questions
A chatbot answers questions in a conversation, but does nothing else. An AI agent carries out a task: it looks up data in your systems, chooses its own steps and prepares or does something, like a reply with the real status of an order.
Answer customer emails with the right order details, read and book invoices, answer questions about your sales figures, reorder stock on time, schedule appointments and draft quotes. It pays off most on tasks that come back often.
It depends on how you build them. In a test with realistic CRM tasks, agents reached over 83% on tasks with fixed steps, against around 35% on average in longer conversations. With clear steps and a person checking, it works reliably.
Yes. Since 2 August 2026, the European AI Act requires that people know they are interacting with an AI system, unless it is already obvious. The Dutch Chamber of Commerce advises saying so at the first contact.
In a portal, AI works with your own data: you ask questions about your sales, get a heads-up before stock runs out, drop in invoices that land straight in your accounting, and approve draft replies to customer emails. You stay in control of what the AI may do.
Using the language model is usually a small item at low volumes. The real costs are in building the connections and maintaining them. At Forwarddd you get a proposal with the price up front, and a customer portal starts at €1,450 one-off.
Sources
- 1OpenAI, 2025A practical guide to building agents
- 2Anthropic, 2024Building effective agents
- 3CBS, 2025Bedrijven gebruiken AI vaakst voor marketing of verkoop
- 4CBS, 2026Gebruik van AI-technologie door Nederlandse microbedrijven
- 5Eurostat, 202520% of EU enterprises use AI technologies
- 6CBS (Conjunctuurenquête Nederland, met KVK, EIB, MKB-Nederland en VNO-NCW), 2026Bedrijven zetten meer in op automatisering door personeelstekort
- 7Visma & Kantar (via Accountant.nl), 2025Veel ondernemers besteden de helft van hun tijd aan administratieve taken
- 8McKinsey & Company, 2025The state of AI in 2025: Agents, innovation, and transformation
- 9McKinsey & Company, 2026The state of AI in 2026: On the road to ROI
- 10KVK, 2025KVK onderzoek: ondernemers zetten nauwelijks stappen rond Europese AI-wetgeving
- 11METR (arXiv), 2025Measuring AI Ability to Complete Long Software Tasks
- 12arXiv, 2025CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions
- 13Gartner, 2025Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
- 14MIT NANDA, 2025The GenAI Divide: State of AI in Business 2025 (preliminary findings)
- 15Klarna, 2024Klarna AI assistant handles two-thirds of customer service chats in its first month
- 16Bloomberg (via The Star), 2025Klarna rethinks AI cost-cutting plan with call for real people
- 17Official Journal of the European Union, 2024Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act)
- 18Official Journal of the European Union, 2026Regulation (EU) 2026/1744 (Digital Omnibus on AI)
- 19European Commission, 2026Commission publishes guidelines on transparency obligations for providers and deployers of certain AI systems
- 20KVK, 2026Moet je laten weten dat je AI gebruikt in je bedrijf?