AI in practice: 5 applications that already save time today
Beyond the hype: these are five AI applications that already, today, take work off your hands.
Five applications already take work off your hands: an assistant for customer questions, first drafts of texts, data from emails and PDFs put in order, a knowledge base on your own documents, and routine steps such as summarising a call. They only work with good data of your own and a person who checks. Say so when someone is talking to AI, and make agreements before personal data goes in.
Almost every business has work that comes back every week. Quotes that look alike. The same questions from customers. Data from an email that someone retypes into another system. That is where AI pays off, not in a chat window on your homepage.
In this article we go through five applications that work today. For each one you will read what it does, what you need for it, where it goes wrong and who stays in control. After that, what the EU AI Act and the GDPR ask of you. This is general information, not legal advice.
Which five applications already save time?
The five are: an assistant for customer questions, a first draft of a text, scattered data put in order, a knowledge base on your own documents, and routine steps taken off your hands. They have three things in common: the work comes back often, it has a fixed form, and you can check the result. The table puts them side by side; after that we work through them one by one.
| Application | What goes in | What comes out | Biggest risk | Where a person checks |
|---|---|---|---|---|
| Assistant for customer questions | Your own texts, prices, terms and standard answers | An answer to a question, or a handover to a member of staff | An answer that sounds convincing and is wrong | With every question it is not sure about, and always when the customer wants a person |
| First draft of a text | A briefing, keywords, your tone and good examples | A draft to work on | Errors in facts and prices, or text that sounds like nobody | Before the text goes out |
| Data put in order | Emails, PDFs, forms and scans | Fields in a fixed format, such as name, date and amount | A misread figure that nobody notices any more | With every field flagged as uncertain, and with spot checks |
| Knowledge base on your own documents | Manuals, procedures, files and frequently asked questions | An answer with the place where it is written | An answer from an outdated or wrong document | Whoever owns a document keeps it up to date; the reader checks the source |
| Routine steps | A one-line instruction and access to your systems | A summary, a sorted request or a prepared draft | Something goes to the wrong customer or is sent without a check | Anything that goes to a customer or costs money only leaves after a person has seen it |
1. An assistant for customer questions
What it does. An assistant answers the questions that keep coming back to you: opening hours, delivery times, how a return works, what a service includes. If it does not know something, the question goes on to a person.
What you need. Approved texts to draw from: your prices, your terms and the answers your team already gives. If those do not exist yet, write them out first; that helps your website as well.
Where it goes wrong. A language model can give an answer that sounds certain and is wrong. An assistant can keep a customer going round in circles with no way out to a person. And an assistant on your homepage often gets in the way: someone looking for a price or an opening time wants to read it on the page, not start a conversation first. That is why we would rather put an assistant in a customer portal or in support with many repeated questions. More on that under AI integrations.
Who stays in control. A person. The handover to a member of staff is part of the design, and you read back conversations regularly. On 2 October 2025 the Dutch Authority for Consumers and Markets (ACM) and the Dutch Data Protection Authority (AP) jointly called on businesses to always give customers the option of speaking to a member of staff, to make clear when someone is talking to a chatbot, and to make sure the chatbot does not give incorrect, evasive or misleading information.
2. A first draft of a text
What it does. You give a briefing or a few keywords, and out comes a draft: the text of a quote, a product description, a message you send every week.
What you need. A few good examples of how you write, the facts that must be in it (prices, sizes, terms) and a fixed structure.
Where it goes wrong. A draft often sounds good and says little. It can mention a price or a feature that is wrong, because the model fills a gap with something that sounds plausible. And what makes your business different, you have to put in yourself.
Who stays in control. Whoever sends or publishes the text reads it first and is responsible for it. That is how Signal is set up, the news part of our own product Basestep: a language model writes the first draft, a person edits, decides and publishes. How to write texts that visitors actually read is covered in our article on web copy.
3. Scattered data put in order
What it does. Information that is now spread across emails, PDFs and loose files is read and written away in a fixed way. Think of a request by email whose name, address, date and wishes end up in the right fields of your planning.
What you need. A fixed list of fields: what do you want from each document, and in what form. Real examples to test on, including difficult ones. Poor scans need cleaning up first.
Where it goes wrong. A misread figure does not stand out. A date in the wrong order or an amount without a decimal point looks just as neat in a table as a correct one.
Who stays in control. Whatever the AI is not sure about stays flagged, and a person checks those fields. On top of that you regularly take a sample of fields that were not flagged.
4. A knowledge base on your own documents
What it does. Knowledge that is written down somewhere and that nobody can find becomes findable. A knowledge base with AI searches those sources and answers an ordinary question, with the place where the answer is written.
What you need. One area of knowledge to start with, not everything at once. The documents that belong to it, and for each document someone who keeps it up to date. Perfect documentation is not required; tidying up is part of the build.
Where it goes wrong. A knowledge base is only as good as the documents in it. If there are two versions of a procedure, the answer can come from the old one. And if everyone can ask anything, someone can get an answer from a document they are not allowed to read.
Who stays in control. The reader, because every answer shows its source and they can open it. The owner of each document, who removes old versions. And your organisation's access rights still apply: whoever may not read something does not get an answer from it either.
5. Routine steps taken off your hands
What it does. A one-line instruction, and a series of steps is carried out: summarising a customer call and filing it, sorting and completing an incoming request, preparing a draft quote from your own price structure, scheduling messages. You can read more about a set-up with several steps in AI agents at work.
What you need. Access to the systems where the work happens: calendar, email, customer records, documents. The rule is: as much access as the task needs, and no more.
Where it goes wrong. The more steps without a pause, the greater the damage from a single mistake. A summary that ends up with the wrong customer. A message that goes out before anyone has seen it.
Who stays in control. A person, at a fixed point in the process. Anything that goes to a customer or costs money only leaves once a person has seen it. Where that line sits, you decide in advance. More examples are under AI automation for your business.
Why is personalisation not on the list?
Content that changes per visitor is deliberately left out here. It needs data about your visitors, and the GDPR rules further down apply to that: a legal basis, a purpose set in advance and as little data as possible. A clear page per audience is usually simpler. Our own product Basestep has no trackers and therefore no cookie banner. snelrepareren.nl also has no cookie banner and no tracking cookies.
How much time does it save?
That depends on how often the work comes back, how long it takes now and how much still has to be checked after the AI. We do not give a fixed number: we have not measured it at a client. On our service page, too, we do not promise saved hours.
What you can do: measure it yourself. Write down how often a task occurs each week and how long it takes. Then run a trial on your own data for a few weeks, and keep track of how long the checking takes. Then you know what it delivers.
How do you choose where to start?
Start with one process, not five. Put every candidate through these questions:
- Does the work come back every week? Something that happens twice a year is not worth setting up.
- Does it have a fixed form? The more judgement and negotiation, the less it fits.
- Can you check the result? If not, do not hand it to an AI.
- What does a mistake cost? If a mistake costs more than the time you gain, leave it with a person.
- Which data goes in? If it involves personal data, or data about health, start with agreements, not with building.
The dullest process is often the best first choice.
What does the EU AI Act say?
The European AI Act (Regulation (EU) 2024/1689) dates from 13 June 2024 and has largely applied since 2 August 2026. For the five applications in this article, the main point is Article 50: saying that AI is involved. An amendment of July 2026 (Regulation (EU) 2026/1744) left Article 50 largely in place.
- An assistant that talks to people. People must know they are talking to an AI system, unless that is already obvious to an attentive user from the situation. The obligation lies with the provider of the system; make sure together with your developer that the notice is there.
- When to tell them. At the latest at the first interaction, clearly, and in a way that also works for people with a disability.
- Generated images, audio and video. With a deepfake you disclose that the content has been artificially generated or manipulated. The providers of those AI systems must mark generated content in a machine-readable way. For systems already on the market before 2 August 2026, there is a deadline of 2 December 2026 for that.
- Generated text. Text that informs the public on matters of public interest, and that was generated or manipulated with AI, must be disclosed as such. That obligation does not apply if a person has reviewed the text and someone holds editorial responsibility. Whether your text falls under it depends on the subject; you review it in any case.
- Knowledge in your team. Since the amendment of July 2026, Article 4 asks you to take measures to support the AI literacy of those who work with AI systems on your behalf.
The European Commission has a voluntary code of practice and a set of icons for labelling generated content. The obligations of Article 50 itself are not voluntary.
When does it get heavier?
With decisions about people. The Act lists, among others, AI for recruiting and selecting staff and for assessing the creditworthiness of individuals as high-risk systems. Much stricter requirements apply to those, which under the July 2026 amendment apply from 2 December 2027. The five applications above do not fall under this, as long as you use them for preparatory work and not to decide about people.
What does the GDPR say?
As soon as personal data goes into your AI application, the GDPR applies, also during a trial. The Dutch Data Protection Authority (AP) names the rules that weigh most with AI:
- A legal basis for using the data.
- A purpose set in advance, and the data used only for that.
- As little data as possible, with a retention period set in advance.
- Transparency towards customers and staff about what you do with their data.
- Security through technical and organisational measures.
- A person for important decisions. According to the AP, you may not take a decision that matters to someone in an automated way.
With a high privacy risk, a DPIA is mandatory: a data protection impact assessment carried out in advance. The Dutch AP works with a list of nine criteria: if two or more apply, you must do one.
What if an employee uses a chatbot on their own?
On 6 August 2024 the AP reported several data breaches caused by employees entering personal data of patients or customers into an AI chatbot, against the agreements. If an employee does that on their own initiative, it is a data breach. If it is part of your policy, it is not a data breach, but often not legally allowed. The AP advises making clear agreements: may a chatbot be used, and if so, which data may and may not go in. You can also arrange with the provider that it does not store the data entered.
For an application we build, we put in writing beforehand which data the AI may see, where it goes and what stays out.
What is involved if you have it built?
We start an AI integration with one conversation about where the same work keeps coming back in your week. Then we write out what goes in, what comes out and where a person checks, and you go through that with your team. We build small, on one process, on top of your existing systems, and you first use it for a while on your own data before it goes into wider use.
Expect two kinds of cost: a one-off build price and the use of the AI, which costs money every month and depends on how much goes through it. What the build roughly costs is on our pricing page. Broader choices, such as which process gets its turn first this year, are in AI at work in 2026.
When this does not apply
If you have little recurring work, or work that is different each time, setting it up costs more than it delivers. If an ordinary form already works well, keep it. If you work with health data, with decisions about staff or with credit, stricter rules apply than described here, and this article is not enough.
What we do not know: we have not measured at any client how much time one of these applications saves, so we give no number. The rules changed again in 2026 and may change after 3 November 2026; the European Commission is updating its guidance. This article is general information, not legal advice. When in doubt, ask a lawyer.
Frequently asked questions
Which AI application pays off fastest for a small business?+
Do I have to say that a chatbot uses AI?+
May I put customer data into ChatGPT or another chatbot?+
Do I have to label a text written with AI as AI text?+
How much time does AI save?+
Wondering which work of yours qualifies first? Tell us which task comes back every week, and we will tell you what AI can do there, what it costs and what it does not solve. Even if the answer is that it does not pay off. You work directly with Nawid; for each project he puts together a fixed team of specialists.
Sources
Consulted on 3 November 2026.
- EUR-Lex, Regulation (EU) 2024/1689 (AI Act) of 13 June 2024, Official Journal 12 July 2024: Articles 3, 4, 50, 113 and Annex III
- EUR-Lex, Regulation (EU) 2026/1744 (digital omnibus on AI) of 8 July 2026, Official Journal 24 July 2026
- European Commission, Code of Practice on Transparency of AI-generated Content (updated 19 October 2026)
- Ondernemersplein (Netherlands Enterprise Agency), Transparantieverplichting bij het werken met AI (in Dutch)
- Dutch Data Protection Authority, Let op: gebruik AI-chatbot kan leiden tot datalekken (6 August 2024, in Dutch)
- Dutch Data Protection Authority, Oproep ACM en AP over AI-chatbots (2 October 2025, in Dutch)
- Dutch Data Protection Authority, Regels bij gebruik van AI en algoritmes (updated 11 July 2025, in Dutch)
- Empireyo, AI integrations
- Empireyo, Basestep case
- Empireyo, ArtsenPool case
Ready to work smarter?
We scan where AI saves the most time and build it safely into your processes.
Your request has reached us. You will usually hear back within two working days. Prefer direct contact? Email hi@empireyo.com or message us on WhatsApp with the chat button at the bottom right.