A spare parts distributor showed us a spreadsheet with 4,200 products and no descriptions. Each description took roughly ten minutes of manual work, which added up to nearly 700 hours. The project had been postponed for two years because nobody had 700 spare hours.
We did something simple. We took the technical specifications from the supplier, fed them to a language model with a clear template, and had drafts within four days. A person reviewed and corrected every description, which took about two minutes per product instead of ten. The project finished in six weeks.
That is the difference between a useful and a useless application of artificial intelligence. Nobody published the text directly. AI did the heavy lifting while a human stayed accountable for the result. Every time that second half gets skipped, problems follow.
What AI can and cannot do
Modern language models are remarkably good at transforming text: summarising, rewriting, classifying, translating, extracting data from unstructured text. They are also good at drafting to a given template.
They are weak on facts they were not given. A model asked about a specific price, deadline or legal provision will produce a confident sounding answer that may be wrong. Every business application therefore needs to supply the model with the information it needs rather than relying on its memory.
The rule that prevents most trouble
Use artificial intelligence where errors are cheap and easy to spot, or where a person checks the output before it is used. The reverse, tasks with a high cost of error and no review, should not be automated no matter how convincing the demo looks.
Where AI delivers measurable value
Customer service
This is not about a chatbot answering on your behalf. It is about a prepared reply that a staff member reviews and sends. Response time drops considerably while the tone stays consistent.
The second form is classification. Incoming enquiries get routed automatically by topic and urgency, which removes the manual sweep through a shared mailbox.
Website content
Product descriptions, article drafts, meta descriptions, image alt text. For an online store with thousands of items, this is the difference between done and permanently postponed.
An important caveat: generated and unreviewed content published at scale damages search rankings. Search engines judge usefulness, not origin. Text that adds nothing new performs poorly regardless of who wrote it.
Document processing
Extracting data from invoices, contracts and delivery notes. Instead of manual entry, the system reads the document, pulls the required fields and passes them into accounting software. A person confirms anything uncertain.
Data analysis
Summarising customer reviews, spotting recurring complaints, grouping enquiries by topic. Tasks that require reading a lot of text and that nobody does because they are tedious.
Which tasks to hand over and which to keep
| Task | Suitable for AI | Why |
|---|---|---|
| Product description draft | Yes, with review | Cheap errors, high volume |
| Enquiry classification | Yes | Easy to verify, repetitive |
| Content translation | Yes, with editing | Good quality, needs context |
| Quotes with pricing | No | Expensive errors, needs judgement |
| Legal or accounting advice | No | Liability and accuracy |
| Handling a complaint | Draft only | Tone and context are decisive |
| Invoice data extraction | Yes, with confirmation | Structured task |
The logic shows in the right hand column. AI is good at volume, people are good at judgement. Projects that fail usually confuse who is responsible for what.
Introducing AI without wasting money
Start with a task you hate
The best first project is something that consumes hours weekly, repeats, and requires no judgement. Usually that means data entry, writing near identical replies, or preparing reports.
Measure before and after
Record how long the task takes today. Without that number you cannot say whether AI helped. Many companies adopt tools and end up with a feeling of progress that nothing confirms.
Do not buy a platform on step one
A large share of useful applications are built with a subscription costing a few euro a month plus a little automation connecting your existing systems. Specialised platforms make sense once the volume is proven.
Write internal rules
Which data may be sent to an external tool and which may not. Customer personal data, contracts, financial information and health data all require care and sometimes an explicit choice of a provider that does not train on your data.
Underestimated risks
Data leakage. An employee pasting a contract into a free tool to summarise it may push information outside the company. Rules and training cost less than the incident.
False confidence. Models sound equally certain whether they are right or wrong. Absence of hedging in the text is not a sign of accuracy.
Vendor lock in. A workflow built entirely on one tool is exposed when prices or terms change. Keep your data in a format you own.
GDPR compliance. Processing personal data through an external AI tool is processing like any other. It needs a legal basis, transparency and an agreement with the provider.
Uniform content. If everyone in a sector generates text the same way, the result is websites that sound identical. Differentiation comes from your experience and specific examples, not from the tool.
AI and search visibility
The change that affects every website is not in generating text but in being found. More people now get an answer directly in the search result or inside a chatbot without opening a site.
That shifts the requirements for content. The pages that win give a clear, specific and verifiable answer, contain data found nowhere else, and are technically readable by a machine. That direction is covered in detail in our piece on SEO and GEO.
The practical consequence: content written simply to exist loses value quickly. Content carrying your specific experience, numbers and examples becomes more valuable.
Frequently asked questions
Will AI replace my staff
More often it changes the composition of the work rather than removing it. The repetitive parts disappear while judgement, client relationships and accountability remain. In practice, teams using AI absorb more volume with the same headcount rather than shrinking.
What does introducing AI cost a small company
The first working applications usually run on tool subscriptions in the range of tens of euro per month plus one off work to connect them to existing systems. Large costs appear with custom models and specialised platforms, which is rarely necessary at the start.
Can AI write my website content
It can draft and speed the process up considerably. Publishing without editing is a risk though: the text sounds generic, repeats what already exists elsewhere, and carries none of your experience. The working model is AI for the skeleton, a person for the specifics and the checking.
Is it safe to feed company data into AI tools
It depends on the tool and the settings. Business plans from serious providers generally do not train on submitted data and offer a processing agreement. Free versions often do the opposite. Check the terms before setting a rule for the whole team.
Where do I start without a technical person
With one task that eats time and requires no judgement. Write it down step by step, test it manually with a tool for a few weeks, and only then think about automating. Most failed projects begin by automating a process nobody has described.
Next step
Artificial intelligence does not solve problems you do not have. It solves the volume problem: many products without descriptions, many enquiries without a fast reply, many documents awaiting manual entry. If any of that sounds familiar, it is worth a look.
If you want help judging which of your processes are worth it and which should stay manual, get in touch. We review the specific tasks, propose a first step with a measurable result, and send an individual quote within 24 hours.
