What Is an AI Model?
A plain-language guide to the "models" behind your agents, and why Outermind uses more than one of them.
The Short Version
- An AI model is the part of an agent that reads and writes. When an agent reads an email, drafts a reply or decides what to do next, it is asking a model. You will also see models called large language models, or LLMs. They are the same thing.
- Models are made by AI companies, which Outermind calls providers: Anthropic makes Claude, OpenAI makes GPT, Google makes Gemini and xAI makes Grok. Outermind does not make models. It connects your agents to them.
- You do not need to understand models to use Outermind. Pick a setup on the AI Models page and Outermind chooses a suitable model for every kind of work and keeps those choices up to date.
What a Model Does
A model takes text in and gives text back. You hand it a question or an instruction along with the relevant material, and it writes a response. Some models also take in or produce pictures.
That is all a model does by itself. It has no access to your mailbox, your files or your systems. Outermind is the part that:
- gives the model the information it needs for the task in front of it,
- tells it what kind of agent it is working for and which rules apply,
- carries out any action through the tools and approvals you have set up.
A useful way to picture it: the model is the reasoning, and the agent is the employee who uses that reasoning to do a job.
Why There Is More Than One Model
Providers offer several models, and they trade off against each other. A bigger model is more capable, but it is slower and costs more to run. A smaller model is quick and inexpensive, but it is not the one you want planning a complicated piece of work.
Using the most powerful model for everything would be like asking your most senior person to sort the post. So Outermind groups models into four levels and gives each kind of work the level that fits it.
| Level | In plain words |
|---|---|
| Frontier | The strongest everyday model right now. |
| Balanced | Nearly as capable, for less. |
| Fast | Quick and cheap for simple jobs. |
| Image | Makes pictures. |
The Nine Jobs
Outermind splits an agent's work into nine jobs. Each job has its own model, so a hard job can use a strong model while a simple one uses a fast model.
| Job | What it covers | Usual level |
|---|---|---|
| Thinks and plans | Working out what a request needs and how to get it done | Frontier |
| Writes | Drafting emails, documents and reports | Frontier |
| Chats | Conversation with you and your team | Balanced |
| Creates images | Making illustrations, diagrams and other pictures | Image |
| Sorts email | Deciding what an incoming message is and where it goes | Fast |
| Screens for risk | Checking content against your safety rules | Fast |
| Summarizes | Condensing long material | Fast |
| Reflects overnight | Reviewing the day's work to improve what agents remember | Balanced |
| Handles everything else | Small background tasks | Fast |
The exact level for a job can differ between setups. The AI Models page always shows the level and the model each job is using.
Where the Models Run
There are three ways to give your agents their models. They differ in where your agents' requests are sent and who bills you.
| How you connect | Where the models run | Who bills you |
|---|---|---|
| A provider's API key (Anthropic, OpenAI, xAI and others) | On that provider's own servers | The provider |
| Microsoft Foundry | In your own Azure subscription, inside the Microsoft cloud | Microsoft, on your Azure invoice |
| Outermind Credit | Outermind runs them for you | Outermind, on your Outermind invoice |
If keeping your agents' AI traffic inside the Microsoft cloud matters to your organization, see Where Your Agents' AI Traffic Goes for a diagram, and AI Data Privacy and Residency for the details and their sources.
Setups: Choosing Without Choosing Nine Times
A setup fills all nine jobs in one step. The AI Models page starts by asking what matters most to you, and offers three routes.
- Keep it inside Microsoft. Pick one of the two Microsoft Foundry setups. The models run in your own Azure subscription.
- Select a Frontier Lineup. Pick one provider's setup. Outermind gives each job that provider's newest model at the right level and keeps it current.
- Mix and match. Connect more than one provider, then choose any model for any job. Each job has a picker that compares your available models by quality and price.
You can start with one provider and add others later. Nothing is locked in: you can change a single job, or the whole setup, at any time.
Keeping Current
Providers release new models often. When a newer model appears at a level, Outermind tests it on the kind of work your agents do before any job moves to it. You choose how updates reach you:
- Always current. Jobs move to the newer model once it has passed those tests.
- Asks before changes. Outermind suggests the change and waits for you to approve it.
- Versions managed by you. Nothing changes until you change it.
Words You Will See
| Word | Meaning |
|---|---|
| Model or LLM | The AI that reads and writes for an agent. |
| Provider | The company that makes a model, or the service that runs it for you. |
| API key | A password-like code from a provider that lets Outermind use your account with that provider. |
| Level | A group of models of similar strength: Frontier, Balanced, Fast or Image. |
| Job | One of the nine kinds of work an agent does, each with its own model. |
| Setup | A ready-made choice of models for all nine jobs. |
| Effort | How long a model is allowed to think before it answers. More effort is slower and costs more. |
Common Questions
Which model is the best? There is no single best model. It depends on the job. The Rankings page (Workforce > Intelligence > Rankings) shows how models score on the kinds of work your agents do.
Can I change my mind later? Yes. You can change one job, switch to a different setup, or add another provider whenever you like.
What happens to the data my agents send to a model? That depends on how you connect. See AI Data Privacy and Residency.