How to choose an AI model for the job

There are now dozens of AI models, and new ones arrive every month. Most people do not need to know them all. They need a way to decide: for this task, which kind of model, from whom, processed where, at what cost? Below is a short method to choose an AI model for the job without becoming an expert.

Start from the task, not the model

Models differ, but the task decides what matters. Ask four questions first:

1. How hard is it? A simple rewrite, a lookup and a summary of a short note are easy. A long contract review, a tricky calculation or a multi-step plan is hard.
2. How much text goes in? A paragraph is small. A fifty-page report is large, and some models handle that far better than others.
3. How much does a mistake cost? A slightly awkward email costs nothing. A wrong clause in a contract might cost a lot.
4. How sensitive is the text? Public marketing copy is different from customer records or legal documents.

The answers point you towards a level of model, and a place where it should run.

Three levels are enough for most work

Rather than comparing models one by one, think in three levels. nerix uses the same idea in its Easy picker: Fast, Balanced and Best.

• Fast: quick and cheap. Good for everyday questions, short rewrites, quick summaries, classification and anything you will skim.
• Balanced: a good default for most work: drafting documents, answering from your files, writing emails, analysing a report.
• Best: the most capable and the most expensive. Use it for hard questions, long documents, careful analysis and tasks where a mistake is costly.

Start at Balanced. Move down to Fast when answers are consistently good enough, because you save credits and time. Move up to Best when you see mistakes in reasoning, missed details in long documents or flat, generic writing where you need depth.

Where your text is processed

Choosing a model also means choosing who handles your text, and in which country. This matters for confidentiality, for customer contracts and for rules such as the GDPR.

In nerix, the Easy picker lets you choose where your text is processed: Anywhere, Europe, United States, China or Worldwide, as offered. nerix then picks a model and says which company answers. Some practical points:

• If you work with personal data of people in the EU, prefer Europe, and read your own obligations. Your customers may ask where their data goes.
• Providers differ. Mistral's models are processed in France, in the EU. OpenAI and Anthropic process in the United States. DeepSeek's own API processes data in China, and nerix prints a caution on its picker. Google's processing is worldwide.
• Your workspace can limit providers. Owners and admins choose which companies may answer for their workspace, and can restrict it to ones hosted in your region.

The Security page lists where each provider processes content today. If a provider's location or terms are not acceptable for a type of content, do not allow it for that workspace.

Look at cost honestly

Every answer costs something. In nerix, credits pay for every answer, web search and export, and each answer shows its cost. Larger models cost more per answer, so running everything on the Best level can empty your credits quickly with little benefit.

A sensible habit:

• Use Fast for volume: summarising many short items, tagging, drafting routine replies.
• Use Balanced for your daily work.
• Keep Best for the questions that deserve it, and check whether the answer is actually better.

Look at what a typical week costs and adjust. A small saving per answer becomes large across hundreds.

Speed and length

Larger models usually answer more slowly. For a quick back-and-forth, that delay is annoying. For a one-off report, it does not matter. Match the model to the way you work.

Also check how much text a model can handle at once. If you need to ask questions about a long document, make sure the model can read enough of it, or use a system that finds the relevant passages first, so the model reads only what it needs. In nerix, library answers use the relevant passages from your files and show their sources.

What the model is good at

Models have different strengths. Some are better at writing in a natural voice, some at code, some at step-by-step reasoning, some at working across languages. Maker websites will tell you their model is excellent at everything. Treat that as marketing.

The only evidence that counts is how a model does on your work. Which brings us to the test.

Test with your own examples

Before you settle on a model for a recurring task, run a small test:

1. Collect ten real examples of the task, including two or three difficult ones.
2. Run them on two levels, for example Balanced and Best.
3. Compare the answers side by side. Are they correct? Complete? In the tone you want?
4. Check the details. Names, numbers, dates and sources.
5. Decide. If the cheaper level is good enough on all ten, use it. If it fails on the hard ones, use the better one only for those.

Do not judge by one answer. Models vary from run to run, and one good or bad result tells you little.

Let the system choose, and know when to choose yourself

For most people, the Easy picker is the right place to start: you say how good and where, and nerix chooses. Every answer names the model that wrote it, so you can see what is happening and change it.

Choose an exact model yourself, with the Advanced picker, when you have a reason: a policy that requires a specific provider, a task where you have tested one model and know it works, or a need to compare. Each model shows its level and roughly how many credits an answer costs, so the trade-off is visible.

Questions to ask about any model

If you are evaluating a model or a provider, ask:

• Where is my text processed?
• Is it used to train models, and can I stop that?
• How long is it kept?
• What does it do badly?
• What does it cost per answer at my usage?
• What happens if the provider is unavailable?

In nerix, your documents are never used to train a model, by nerix or by a provider. Check any other tool you use against the same question.

The short version

Match the level to the difficulty: Fast for easy and frequent, Balanced for most work, Best for hard and costly-to-get-wrong. Choose where the text is processed according to how sensitive it is. Watch the cost. Test with your own examples. And prefer a tool that tells you which model answered, so you are never guessing.