Which AI models are there, and when do you use which one?

One moment you hear about ChatGPT, the next about Claude or Gemini. Meanwhile Llama and Mistral come up as well. Anyone who wants to use AI in their own software soon wonders which model is best. That question is understandable, but for organisations it is usually too narrow.

That is why this blog does not give a ranking. We briefly list the best-known models and then look at the question that gets you further: when does which kind of model fit into an application you already have or want to have built?

The best-known AI models at a glance

Broadly speaking, you can divide the best-known models into two groups.

The first group are generic models:

• ChatGPT from OpenAI;
• Claude from Anthropic;
• Gemini from Google DeepMind, integrated into the Google ecosystem.

ChatGPT, for example, runs in the cloud, which makes it less suitable for organisations that want to stay on-premise. With this group comes the question of which data may go to an external AI service.

The second group are open source models. You can run Llama from Meta locally or in your own cloud environment. That gives you more control over your data, but it also takes more technical knowledge and maintenance. Mistral is a European provider of compact models, which are also suitable for applications that have to run locally.

Use this overview as a starting point, not as a ranking.

Why the model is not the first question

Which model is best? That question actually comes too early. First it has to be clear which work can be done better, and under which conditions. Three questions help:

• Which task currently takes a lot of time or is prone to errors?
• Which data may the AI model use, and which not?
• Who checks the output before it affects customers, employees or decisions?

Only once those answers are there can you decide whether you are better off with a generic model, an open source model or a combination within your own application.

Choosing a model only becomes valuable once it is clear which process you want to improve, which data is available and how the outcome is checked.

When does which kind of model fit?

Those three questions help you narrow down the choice. For business software, these are often the most important decisions.

Where may your data go? If data may go to an external AI service, a generic model is an option. Is it privacy-sensitive information that has to stay within your own environment? Then a model you can run yourself is the more obvious choice.

What does AI have to work with? Does AI have to work with existing systems, such as a CRM, ERP, document storage or customer portal? Then that integration is part of the choice. If your organisation already works with Google, that can be a factor.

May AI only make suggestions, or also act? You can give AI a supporting role, with suggestions that a user can always review. You can also let AI carry out actions itself. The more AI does on its own, the more weight the question of who checks the outcome carries.

Who maintains it after go-live? Who monitors quality, errors, costs and further development once the AI feature is live? If you choose a model you run yourself, maintaining that model comes on top of that.

A combination of models within your own application can also be the outcome, each for the task it suits.

An example: reading documents

Take the automatic reading of documents, such as invoices, forms or contracts that are still processed by hand.

The considerations above then determine what is needed. Do the documents contain personal data? Then the question of where data may go weighs heavily. Does the result have to end up directly in an existing system? Then the integration with that system is part of the choice.

The documents themselves make a difference too. If documents follow fixed formats, a model can be trained to recognise specific fields and process them directly. If the input is less predictable, AI can still assess whether something is correct, legible or deviates. Instead of checking everything by hand, people then only look at the exceptions.

In this example, the model is one part. The real gain lies in integration: AI that is part of the process itself, not a separate step alongside it. How to take that first step is explained in Integrating AI into existing software: where do you start?

From choosing a model to software that keeps working

Choosing an AI model is one step. The real value only appears when AI becomes a reliable part of your application, workflow or decision-making. That application has to be built, and then keep working.

That is where our role lies. We build and maintain web applications and mobile apps, and we develop traditionally, with the help of AI.

The considerations in this blog apply to new software, but also to an application you already have. For existing applications we offer maintenance and support, also when another party built the application. Why maintenance matters so much once software becomes part of daily processes is explained in AI makes building easier. But who maintains it afterwards?

The best model does not exist, the best application does

There is no single "best AI model". For an organisation, the more useful question is: which work gets better from this, and what does that take?

Start with the task that takes time or is prone to errors. Decide which data may be used and who checks the outcome. Only then look at which kind of model fits.

Do you want to have new software built and take these considerations into account from the start? See how we develop web applications.