Local, cloud, or hybrid AI: how to choose

Summarizing documents, preparing meeting notes, finding information, or drafting a first version: AI can already help with many everyday tasks.

One question quickly follows. Should you use an online service, install AI on a computer, or combine both?

There is no single answer for everyone. The right choice mainly depends on four things: the documents you give the AI, the simplicity you want, the internet connection available, and the true cost.

The three options in one minute

Five questions that make the choice easier

1. Which documents will you give the AI?

Start here. A public text does not carry the same concerns as a customer file, contract, HR record, or medical document.

For ordinary content, a professional cloud service may be the most practical choice. For personal, confidential, or strategic information, a local solution can reduce exchanges with an outside provider.

The French data protection authority says the deployment method should follow the use case and the data involved. It also recommends sending only the information that is genuinely needed.

2. Is simplicity the priority?

Cloud services require little preparation: create an account, gain access, and the service is ready. Updates and computing power are managed by the provider.

Local AI needs more attention. You need a suitable machine, a model to install, updates to apply, and occasional problems to solve. It is a better fit for a person or team prepared to maintain it.

3. Do you need to work without internet access?

Fully local AI can keep working on a train, at a remote site, or during an internet outage. That is a concrete advantage for mobile work and places with an unreliable connection.

A cloud service depends on the internet and the provider’s availability. In return, it can be reached from several devices without relying on one particular computer.

4. What level of quality do you actually need?

Cloud services generally provide quick access to the newest and most capable models. They are useful for long, varied, or difficult requests.

Local AI may be entirely sufficient for summarizing documents, classifying requests, extracting information, or writing in a known format. The best test is to try a few representative tasks and compare the results.

The right model is not necessarily the most powerful one. It is the one that completes the work correctly with the fewest revisions.

5. What is the full cost?

Cloud services are generally priced through subscriptions or usage. You can start without purchasing special hardware.

Local requests may not create a separate bill, but the computer, electricity, installation, and maintenance still cost money. A local solution that is rarely used can therefore cost more than an online service.

Three everyday situations

A small team wants to save time

It wants to prepare meeting notes, summarize documents, and draft first versions without installing hardware.

Simplest choice: cloud AI. A professional service provides a quick start. The team should still check its privacy rules and avoid sharing information the task does not need.

A firm works with confidential case files

Its documents contain personal, contractual, or strategic information. The team wants to limit how far those documents travel outside its environment.

Option worth studying: local AI. It may fit that constraint better, provided the machine, access, and backups are protected. “Local” does not automatically mean “secure.”

A company has very different needs

Sensitive documents need tighter control, but teams also want capable AI for general research, ideas, or non-confidential content.

Most flexible choice: hybrid AI. Sensitive tasks stay local. The others can use the cloud. A person remains in control before a message, publication, or hard-to-reverse action.

Why hybrid is often a practical compromise

A hybrid setup avoids forcing the same choice everywhere. The rule can stay simple:

• sensitive case files are handled locally;
• ordinary, non-confidential requests use the cloud;
• when there is doubt, nothing is sent without approval.

This approach only works when the rule is understandable. Everyone should know which service receives the documents and when cloud AI is used.

What to check before adopting a service

Whether the choice is local or cloud, five checks prevent most surprises:

1. The data: what information is sent, and what can be removed?
2. Retention: how long are requests, documents, and answers kept?
3. Use: is the data used to improve or train the service?
4. Access: who can read documents or approve an action?
5. Continuity: what happens if the computer, internet connection, or service is unavailable?

A professional offer and a consumer account may follow very different rules. For example, OpenAI says it does not train on Business and API data by default. Always check the exact product in use, because terms can vary from one service to another.

The French cybersecurity agency’s generative AI guidance also explains that protection depends on the whole system. Access, devices, backups, and everyday practices matter as much as where the AI runs.

How this works with Stellary

Stellary helps organize missions and choose which tools an agent may access. With Stellary Desktop, a mission can also use AI or a tool available on a connected computer.

• LM Studio can provide models installed on a machine;
• Codex and Claude Code can work in a project stored on the computer;
• depending on the chosen configuration, the model may be local or use a cloud service;
• permissions and approvals defined in Stellary continue to apply.

The local AI apps guide explains how to connect a machine and make a model available in a workspace.

Official sources

• CNIL — Questions and answers on using generative AI systems
• ANSSI — Security recommendations for a generative AI system
• OpenAI — Business data privacy, security, and compliance
• Mistral — Using offline models