You use ChatGPT, but
nothing has changed in the company?

In addition to the risk of fines, you run the danger of deploying systems that will later be prohibited or will require costly modifications. Setting the rules for AI correctly right at the start will save you costs in the future.

From the point of view of security and the AI Act this is highly risky. Free versions (e.g. standard ChatGPT without an Enterprise licence) may use the data entered to train the model further. If your employee enters sensitive data, trade secrets or clients’ personal data there, it becomes part of the public cloud. We will help you set up secure systems and procedures that lock your know-how inside the company.

We start with a short audit of your processes. We look for repetitive tasks, work with large volumes of text or data and areas with a high error rate. These are the places where AI delivers results within a matter of weeks. Our 10 years of experience in digitalisation allow us to distinguish real benefit from technical enthusiasm quickly.

The AI Act is an EU regulation that regulates the development and use of AI. Although some obligations are being phased in gradually, it is essential to have the rules in place now (AI governance). If you are now deploying new systems without regard to the AI Act, you risk having to redesign them at great cost or switch them off in a year’s time.

This is one of the most common concerns. We deal with it by setting up processes, using corporate licences or using API interfaces with local models, where your data will not be used for further training. Our services also include setting up an internal policy for the safe use of AI.

Training is not just about theory. We focus on practical workshops. Employees try out on their own tasks how AI can help them with writing e-mails, analysing spreadsheets or preparing reports. The aim is to remove the fear of the technology and replace it with a practical skill.

The costs depend on the complexity. Many effective tools work on a monthly subscription basis and deploying them is a matter of days. For complex bespoke systems we prepare a return-on-investment (ROI) analysis so that you know exactly how long it will take for the time saved or the increased efficiency to repay the investment.

Using ChatGPT in a browser is an individual tool for increasing personal productivity. Genuine integration means connecting AI models with your internal data (CRM, ERP, databases) via an API. Thanks to our 10 years of experience in digitalisation, we will not just help you with “writing prompts” – we will design procedures and systems that automatically sort e-mails, analyse orders or generate reports without your data being exposed outside your company

We build on the methodology of our courses at Mediatoring – we always adapt the training to the specific audience. We have modules for marketing (GEO and content creation), for sales staff (lead analysis and personalised outreach) and for management (strategy and the AI Act). The aim is not to teach people to operate one tool but to change their working workflow so that AI saves them hours every week.

The AI Act introduces an obligation of “AI literacy” for all organisations that use AI. If you now introduce processes without regard to this legislation, you run the risk that your solution will retrospectively be declared non-compliant. We will help you set up an ethical and legal framework right at the start, which is far cheaper than correcting mistakes later and facing the threat of heavy penalties.

Yes, that is a key part of our technology-independent consulting. We will assess whether it is more advantageous for you to use an existing platform (such as Microsoft 365 Copilot) or to build your own solution based on open-source models (e.g. Llama, GPT OSS and so on) running on your secure server. We decide on the basis of a cost analysis, data security and your specific needs. We are not dependent on any of the solution providers; we are on the side of the customer and their needs.

We do not underestimate this psychological aspect. Our training combines technical instruction with “change management”. We present AI as an assistant that removes routine, boring work, not as a substitute for human creativity and judgement. Thanks to practical demonstrations taken from their own working day, employees quickly understand that AI is a tool that increases their own value on the labour market.

At least partly, yes. Language models only work where they have access to data in a usable form – not in ten versions of the same Excel file. In practice the two go hand in hand: when a process is digitalised, the data is unified and completed at the same time, so AI can be connected to it straight afterwards. That is why we do not start by buying an AI tool but by putting processes and data in order.

Yes, and it is usually the fastest first step. Receiving and matching documents, approvals, monitoring deadlines, transferring data between systems and regular reports can all be automated. With documents, AI is useful today for extracting data and passing it on. At the same time we leave checking to a human wherever an error would be expensive.

Our experience is different: routine disappears, not people. AI and automation take over retyping data, searching documents and putting reports together – in other words work that nobody enjoys. People move towards customers, checking and decision-making. For this to work, every deployment must also include explanation and training, otherwise the team will go back to Excel.

Most of all where work involves text and documents: searching company guidelines and contracts, preparing answers to repeated queries, extracting data from invoices and orders, checking inputs and summarising long materials for management. The prerequisite is access to up-to-date company data and clearly defined boundaries for what AI may and may not do.

First we try the simpler route: a well-constructed prompt and connecting the model to your documents and data. Fine-tuning makes sense when you need stable output in a precisely defined structure or style, when the model has to understand your terminology, or when, with a large volume of queries, a smaller fine-tuned model is more economical to run than calling a large cloud model. We decide on the basis of measurement, not on a hunch.

Yes. We deploy open-source models on your server or in a private cloud in the EU, so neither queries nor documents go out to public services. Deployment includes setting up access, logging and backups. We choose the solution according to the hardware you have and how fast you need answers – some tasks can be handled by a single server with a more powerful graphics card.

The most valuable material is real examples: pairs of query and correct answer, completed documents, approved texts or historical decisions. In practice this means hundreds to thousands of good-quality examples, and quality matters more than quantity. We will help you sort the data, remove personal information and prepare it into a training set. If you have too little, we will start by connecting the model to your documents and postpone fine-tuning.

It depends on the size of the model and the number of users. Smaller open-source models run on a single server with a more powerful graphics card; larger ones need dedicated hardware or a rented private cloud. Before buying anything we calculate how many queries a day you will have and compare the cost of running it yourself with the cloud option.

Your data remains yours and we do not use it to train anything else. The trained model and the training set belong to you, including documentation of how they were created. With open-source models we check the licence so that commercial use is in order, and with cloud services we configure things so that your inputs are not used for further training.

Not from the text any more. Generative AI writes without errors, can imitate your jargon and a colleague’s signature and can even manage a voice recording. So rely on process, not on impressions: confirming changes to bank details through a second channel, two-factor login and the rule that nobody sends money or passwords on the basis of an e-mail or a phone call. That is why we add examples of AI-generated attacks to our training.

A list of approved tools, a clear definition of the data that must not be entered into them (personal data, contracts, source code, know-how), a rule on human checking of outputs and a responsible person people can turn to. Alongside that, instruction of employees and a record of who completed the training and when. We write the guideline on one or two pages so that people actually use it.

We recommend it even if the legislation did not require it. Without a list you do not know where company data is going and what an external service is storing about your company. The record is usually a simple table: tool, purpose, who uses it, what data goes into it and who is responsible for it. It helps with security questionnaires from customers and with meeting AI Act obligations.

Yes, if the data must not leave the company. We can deploy an open-source model on your server or in a private cloud and fine-tune it on your documents. The data then stays with you and what you mainly have to deal with is access rights and backups. For most companies, though, it is quicker to start with a cloud service with the rules set up, and to move to local operation only when there is a measurable reason.

It is easy to check: ask the same question your customer would ask. AI tools answer from sources that can be traced, so they are influenced in a similar way to search. We add missing facts to your website and to the sources the models use, correct outdated information and keep checking what the answers look like. The change usually shows within a matter of weeks.

Unfortunately, yes. In a few minutes generative tools can produce dozens of serious-looking reviews, articles or a voice recording that sounds like the managing director. That is precisely why reputation is monitored continuously – the sooner a fake wave is caught, the easier it is to deal with at source and the smaller the chance that search engines and AI answers will take it up as fact.

You can manage the basics: alerts for the company name and the names of directors, monitoring reviews and regularly asking AI tools what they say about you. An agency makes sense at the point where content has to be removed at source, where you need to deal with platforms and registrars or systematically build content that pushes negative results down. We are happy to hand over the monitoring set-up for you to manage yourself.

General and informational queries are answered by the overview directly in the results, and some traffic disappears as a result. But for queries where someone is choosing a supplier or a product, people still click through. That is why we build content around queries with purchase intent and add facts that can be verified – those are exactly the pages that models cite and send users to.

As a helper yes, as the author no. Models do not know the facts about your company and will fill them in their own way, which shows both in the results and in customer trust. What works is a combination: the material and figures from you, AI for the outline and variants, a human for fact-checking and the final text. Content like that also holds up in the answers of AI search engines.

Yes, and it is one of the best uses. The model works with your catalogue and a template, turns parameters into readable text and can handle a volume there would otherwise be no time for. The prerequisites are good input data and checking a sample – for parameters where mistakes are not acceptable, such as dimensions or composition, we check everything.

Indirectly, yes. Models draw on sources that can be traced on the web and are rated as trustworthy. When you are written about on sites that carry weight, the chance grows that AI will cite you as an example of a supplier or a source of information. A mention without a link also helps, because the model associates the company name with the topic.

It speeds up the preparation of variants of subject lines and texts for individual segments, helps to estimate who will be interested in an offer and helps to choose a suitable sending time. It does not replace the decision about what to send and to whom, nor a good database. Personalisation applied to poor data merely sends out the wrong message faster.

We use it for routine work: preparing materials, variants of texts, refining keywords, summarising data and reporting. It saves hours, but a human always checks the output. The second thing is visibility: people ask ChatGPT or Perplexity and the answer is created from publicly traceable information. That is why we also make sure AI tools have the right data about you.

No. AI speeds up production and analysis, but it will not decide for you where to put your budget and it will not spot that the data has been measured incorrectly. Without a brief and checking, a great deal of content is created with no effect. It is useful where volume and repetition are involved. Decision-making, strategy and responsibility for the result remain with people.

Usually adjustments are enough, not a new website. What matters is a comprehensible structure, consistent information about the company, prices and terms in one place, and content that answers customers’ real questions. AI tools and search engines can draw on a website like that. We only recommend a complete redesign when the website is holding back ordinary visitors as well.

Yes, it is one of the most frequent topics today. We discuss which tasks it makes sense to hand over to AI, which tools to choose, what you may enter into public models and where a solution running on your own premises is better. Related to this are rules for the team and a record of who uses AI for what.

It depends on two things. First, whether the website denies them access in robots.txt, and second, whether the content is available without running JavaScript. When text is only loaded by a script, some tools will not read it. We check both when reviewing indexing and you receive the result in writing.

It depends on what the website does. If you want enquiries and recommendations, it usually pays to let them reach ordinary information about services, prices and contacts. With paid content or original research it makes sense to restrict part of it. The decision is commercial, not technical, and we recommend making it consciously.

It helps with sorting large lists of addresses, finding patterns in server logs and quickly rewriting weak texts. But the decision about what belongs in the index and what does not has to be made by a person who understands the business. Blindly generating content usually makes the situation worse, because it adds further duplicates.

Yes, and it is one of the main reasons why companies ask for PR today. AI assistants put their answers together from public sources. If nobody writes about you, the model has nothing to draw on. Consistency of information and factual texts help, not a quantity of superlatives.

You can. A common option is for you or AI to prepare the first version and for us to adapt it to the requirements of the media outlet. Purely generated text, however, is often rejected by editors or is published with no effect, because it contains nothing specific. What works best is your own figures, case studies and practical experience.

A chatbot answers a query; an agent completes a task. It is given a goal, plans the steps itself, reaches into your systems via API, verifies intermediate results and writes the output where it belongs. Typically it processes an order, prepares material or checks a contract against rules. We always build an agent with a log of steps and with human checking wherever an error would be expensive.

Yes. First we describe the process and decide which steps the agent can handle alone and where a human must decide. Then we build the application on top of your ERP, CRM, e-shop or document store via API. This includes permission management, a log of all steps and a test run on a sample of data before switching to live operation.

We do both. For demand forecasting, classification, customer scoring, anomaly detection or predictive maintenance, classic machine learning models are more accurate, cheaper and easier to explain than language models. We build them directly on your data sets and evaluate them against your existing way of making decisions so that the benefit is measurable.

Yes. From camera images we handle defect detection, completeness checks, reading labels, codes and serial numbers, or counting items. The procedure is always the same: collecting and labelling images, training the model, verifying it on the real production line and only then deploying it. The model can run directly at the machine, so it does not need a permanent connection to the cloud.

We package the model into a service with an API, deploy it on your server, in a private cloud or on an edge device at the machine, and connect it to the systems that are to work with it. We version both models and training data, so a previous version can be restored at any time. We monitor availability, response time, costs and, above all, the accuracy of the outputs.

Data changes over time and with it the accuracy of the model declines. That is why from the outset we measure the quality of outputs against a control sample and watch for deviations in input data. When accuracy falls below the agreed threshold, you receive an alert and we retrain the model on current data. For some tasks retraining is scheduled regularly from the start.

Both, depending on the task. With language models, fine-tuning an existing open-source model on your terminology and data is almost always sufficient, because it is an order of magnitude cheaper. Training from the ground up makes sense for specific machine learning and computer vision tasks using your own data. We always first check whether the simpler route is enough.

We need images from the environment where the model will run, ideally in the same lighting and at the same angle as in operation. For simpler detection, hundreds of labelled images per category are enough; for fine defects more are usually needed. More important than quantity is that the data also contains rare cases and real defects, not just perfect items.

Yes. Smaller models, particularly in computer vision and anomaly detection, run on an edge device right at the production line. The advantages are low latency, operation independent of the network and the fact that data does not leave the plant. Only results and metrics are then sent to head office, not raw images.

The agent is given only the permissions it needs for the task and accesses data through an interface with a limited scope. Every step is logged, so it can be traced afterwards what the agent read and what it wrote. For sensitive steps, approval remains with a human. We set this up together with your IT team and reflect it in your internal guideline.