The AI transformation: how companies and public authorities are becoming more digital

The first step into the AI age is the start. It starts with the question of AI strategy. There are two options, which can also be mixed: Cloud (ChatGPT) or in-house AI. Both have their own advantages. However, the cloud does not deliver the better results. Some also ask about data security and costs.

Introduction

AI offers incredible opportunities. This is why many companies and authorities are thinking about using AI to make business processes more efficient or to enable them in the first place.

In recent weeks, the author of this article has had the opportunity to give numerous presentations, workshops and training courses for companies, public authorities/administrations and at conferences. This has led to the conclusion that an AI strategy is of interest to many. Incidentally, the photo also comes from one of these lectures. It was taken in Düsseldorf/Germany at a meeting of the construction industry, which can also benefit from AI and was therefore keen to find out more.

Part of a AI-strategy is asking whether one is willing to send their treasure trove of data to Microsoft or OpenAI, or if one would rather like to cultivate this treasure themselves. The question can be answered more easily when one knows that an optimized AI not only delivers better results than ChatGPT but also enables full data security.

Which AI strategy suits your organization best? That depends. There are different AI user groups. Depending on the group, a different AI strategy makes sense. The groups are:

• Private user: You don't need a AI strategy and shouldn't read this post. It would be more recommendable to take a look at other blog articles on Dr. GDPR.
• Individual users / individual entrepreneurs: Their economic willingness depends on the benefit. It is assumed that they do not need a AI strategy or cannot implement one. They will land on ChatGPT, unless you trust OpenAI or Microsoft not.
• Small companies: If AI is of great benefit to you, an AI strategy would be advisable. Otherwise not.
• Larger companies or public authorities: An AI strategy is not only sensible, it is mandatory. They are the first target group of this article.

What is an AI strategy?

A strategy is a long-term plan.

They want to achieve business goals. It shouldn't matter whether with or without AI. Only: With AI, many things are possible that were previously either impossible or too expensive or time-consuming.

An AI strategy helps you to find an AI system for a problem that you want to solve. The AI strategy also helps you to find the problems you want to solve.

The first step is therefore: Find the problems you wanted to solve. Instead of problem one could also speak of task, business process, project or application case.

The AI strategy of your organization

We go through the strategy step by step and stick to the important points.

The first step is to identify use cases, i.e. the problems that your company or authority wants to solve.

Recognize use cases

What processes are in place in your company? Since AI works with data, the processes that are to be handled with AI must have a database.

Two cases must be distinguished here:

1. You want to process paper with AI. To do this, this paper must be scanned. It can then be processed with AI. Even OCR works better with AI than before. After the paper has been scanned, we can go to case 2 (which can also include OCR with AI).
2. You want to process existing data, files (especially documents or images) or user input. Let's get started in the AI age.

Digitized data or information is the basic prerequisite for AI applications. This has now been fulfilled. Which use cases are possible?

AI can process all types of data. These types of data are called modalities. Examples of modalities are:

• Text
• Image
• Video
• Audio
• Sensor values

An AI system is based on a neural network, just like your brain. Such a neural network can do quite astonishing things, just like you can. Qualitatively there is no difference between your brain and that in the computer (some see it differently, but do not mention any arguments).

What are good use cases?

Good use cases are those that can be handled easily or significantly better or cheaper with AI than with previous methods. Examples of such use cases are:

• Knowledge assistant for your employees or citizen inquiries: Most people struggle to understand a legislative text or legal regulation. AI reduces this barrier. Another example: Citizens want to know how high they can build in area X: The AI gives the answer based on your documents. Hallucinations are what they should expect from ChatGPT, but not from an optimized AI (Offline-AI).
• Extracting knowledge from documents: That works very well. What do you then with this knowledge? If you don't know that, skip this point. Further suggestions follow.
• Enhance your offers: Your offer lists multiple pages of part numbers with brief descriptions in a Word file. Wouldn't it be better to make everything more tangible and easier to understand? The AI finds images from the manufacturer of the products included in your offer. Long descriptions, product data sheets, and technical specifications can also be automatically inserted into your offer.
• Compress knowledge: You don't want to watch this training video; you don't want to read these 50 A4 pages of text. The AI translates language into text and long texts into a summary. Creating podcasts from texts is also possible.
• Edit images: Your city wants to feature its landmark with a flood situation to sensitize citizens or illustrate the current flood. To do this, you take a photo of your landmark, click on the foreground a few times to mark it, and tell the AI: "Add flooding". This approach is actually better than having ChatGPT generate an image of your landmark with flooding. Because a landmark is what it is. It can only get worse through artificial generation.
• Protocoling and Live-Translation: Long meetings or sessions can be recorded live via video, translated into text or even other languages in real-time. The target language can then be read out as needed. There is already a very good solution from Germany that is used at larger congresses.

You can best solve these applications with an optimized AI. For general tasks and knowledge questions, please use a chatbot of your choice. For serious and concrete business applications that also need to take your company knowledge into account and are secure, an optimized AI is the best option.

An optimized AI uses open source AI models and is optimized for your use case. The result is high quality and reliability in AI output.

How do you optimize an AI?

There are several options for this:

1. Prompt tuning: In contrast to ChatGPT users, this takes place in the background. The user does not have to deal with it.
2. Automation: In contrast to ChatGPT, your optimized AI runs largely automatically. This makes it possible to optimize sub-processes that need to be run through for the final result.
3. Training (more precisely: fine-tuning): We teach your AI what is good and what is bad. To do this, we need examples that consist of input (actual) and output (target). For a translation app, for example, an example would be a pair of a) German text and b) English equivalent.

The possibilities for achieving better results are thus diverse and powerful.

Check requirements

Once it has been recognized that use cases exist, the first goal should be a prototype. Ready-made market solutions are an exception. With AI products, the question is who offers them. A server located in Germany is no guarantee of data security. A Taiwanese server operator can access this server at any time. The server location is technically irrelevant, even if the location in Germany offers a little more than other locations. At least you can visit the server easily.

For a prototype, but also for the finished AI solution, you need hardware. The best use cases for entry do not require any special hardware, however.

For fully-fledged AI applications, you need AI servers (one or more, depending on the number of users). AI servers are characterized by their AI-capable graphics card. This usually comes from NVidia. Seek advice on this to find the graphics card that offers the best price/performance ratio.

AI applications are calculated on special graphics cards: In contrast to processors (CPUs), these can perform many thousands of calculations in parallel.

The graphics cards in AI systems are therefore only used for calculation and not for outputting a video signal.

There are three good options for servers:

1. They operate their own data center. Many companies and administrations have one of those. There, your AI servers are located. That's the On-Prem solution.
2. They buy their AI servers themselves (get advice). The servers are then located in a data center in Germany, which you rent into. This is called Colocation.
3. They rent a AI-server from a purely German provider with a German contract for data processing and server location in Germany. Let them assure you that your data will not be sent outwards by the AI on the server. Any serious AI-provider will do that. The Rental server is such a safe choice.

The quickest way to get started with AI is a pure knowledge search that finds results semantically. Searches are therefore not performed by comparing letters, but by comparing meanings. If the results are displayed in the form of a hit list, you do not need any special hardware.

From a hardware point of view, the quickest way to get started is to rent a server. The costs are low, so that a later switch to your own server makes economic sense. From a technological point of view, this change can also take place with manageable effort.

You also have an advantage with a quick start with rental hardware. A few months later, better hardware may be available at the same price or for less money. You save money by purchasing hardware at a later date. At the same time, the rental server gives you fast results.

From prototype to live system

The prototype is used to demonstrate the performance of the AI system for your use case. Optimizations will also take place here. You should know one thing about this:

A prototype can therefore be realized with a manageable budget. The finished version can be easily created from this. The work essentially consists of providing a user-friendly user interface and the like.

Upgrading your intelligence

You can't really upgrade your intelligence, which is hopefully located between your ears, just like that. But you can improve your intelligence, which is located in your AI system, on a regular basis.

It is very easy to replace the electronic brain:

1. Download new electronic brain (AI model). Excellent AI models are available for this, which are already significantly better than ChatGPT in some areas.
2. Replace the old AI model with the new one from step 1: In the best case, this operation consists of copying and overwriting files.

You can do more than OpenAI and Microsoft can do for you. For the enormous universal intelligences are not only mediocre or poor experts in your specific use cases; they are also very cumbersome. You wouldn't even change a Albert Einstein that easily. It would be good if that were possible. Because if Mr. Einstein is supposed to change the tires on your truck, then he starts crying.

You can therefore easily replace the AI brains of your optimized AI systems on an ongoing basis. ChatGPT, on the other hand, is rarely updated. Above all, you prefer to decide for yourself when an update takes place. Otherwise, you may be a little surprised at how the AI system suddenly behaves just because OpenAI suddenly thought it needed an update.

Further advantages of optimized AI systems

Cloud Services like ChatGPT or Mistral have the following cost structure and resulting consequences:

1. Free account: Free and mediocre. But very good for the question of what the second highest mountain in the world is. But you can also find out differently.
2. Flat rate for individual users: For around 25$, each user can type something into a box on their own responsibility and receive an answer. The user knows nothing about his colleagues. Colleagues may exchange information in the corridor. Does that really make sense?
3. Usage-based fees: A programming interface allows you to query a cloud AI. You pay money per call. The more data you enter, the more expensive it becomes. If you want to enter your company documents, it becomes very expensive. How often do you have to call up a universal AI to get a result? Counter question: How long does it take your colleague to complete a certain task that they have never done before? Answer: You don't know. Not even your colleague knows.

Optimized AI has a different cost structure:

1. Rental model: You rent an AI service (Software as a Service, SaaS): There are several pricing models here. One possible model is a low flat rate. It doesn't matter how often your AI is called up and how many company documents are processed: The costs are always the same.
2. Purchase model: You have an AI service developed for you. This costs money once and a little for maintenance (if desired). It doesn't matter how intensively you use the AI: The costs are always the same.

Fixed costs have several advantages over usage-based charges:

1. The costs can be planned because they are known.
2. The costs are manageable.
3. Your system can work 24/7. Recommendation: Torture your system, because it can think for itself and improve itself.

Above all, the 24/7 operation is a strategic advantage!

Very important: The data is with you. An optimized AI does not need to retrieve data from anywhere else or send it anywhere. Unless you want it to or the use case requires it.

For example, it is necessary for the AI to tap into the Internet if you want search results to be included in AI responses. Or you want to send the AI results to your database server or your DMS. Then, of course, the optimized AI must send data to these systems – because you want it to!

Conclusion

Several aspects are important for an AI strategy.

But the most important thing is: get started! Achieve results quickly.

The first question is: What use cases are there in your company or authority? This is where a workshop comes in handy.

This is followed by the question: Do you want to send your data to cloud services or not?

Everything else follows from this. If you have any questions, please contact the author!