AI Model

The AI Model tab is where an admin sets the starting point for Centrus across the organisation. It holds six shared settings: Organisation Context, Organisation AI Model, Organisation Temperature, Organisation Live voice, Organisation image model and Organisation image quality.

Those six are the defaults. Anyone can set their own model, voice, pictures and a personal context, and a personal choice is used instead of the organisation's for that person. The same form also shows the personal fields for the admin who is signed in, under Personal Context and, once you open Show advanced, under Personal AI Model. Changing those changes that admin's own account. It does not change anyone else's personal settings.

Comparing models, and what each answer costs, is covered in Choose the AI model. How to write a personal context, and how a personal choice wins, is in Personal AI settings. This guide is the organisation page: what to put in the shared context, how that context is combined with a tag and a person, and what each control on the tab does.

Open Settings in the menu on the left, then AI Model. The page heading is Settings, and the line under it is Update your account details. The heading on the tab is Model settings.

1. 1Click Settings in the menu on the left.
2. 2Click AI Model. It sits after Account, Billing and Branding.
3. 3Read Organisation Context at the top. Personal Context is directly under it.
4. 4Click Show advanced to open the model, temperature, voice and picture settings. The button then reads Hide advanced.

Only an admin sees the tab. Anyone else who opens the address is sent back to their usual page, and the server only accepts a save of the organisation settings from an admin. Train access does not open this tab. To give someone admin access, see Edit users and roles.

On a phone the same tab is there for an admin. The row of tabs scrolls sideways, and the form sits in one column. The labels and the buttons are the same as on a computer. Most admins set this up on a computer, which is the layout drawn here.

Organisation Context is text Centrus is given before it answers. The line under the label reads: This context applies to all users in your organisation. Use it for things that are true of the business on every chat: what you do, how answers should be written, and a standing limit you want kept. It does not replace your documents. Centrus still searches the files on the chat's tags. The context tells it how to use what it finds.

The information icon beside the label opens What does Organisation Context do? The window calls it the default background and instruction set for everyone in the organisation: the business, its goals, and the way you want answers written. There is no button except the close icon, and closing it does not save anything.

The box is a fixed height and does not grow. Longer text scrolls inside it. There is no character counter and the page does not show a limit, unlike tag context, which stops at 1,000 characters. Centrus stores whatever you save, and it sends the whole text with answers, so keep it to a short paragraph. A page of instructions crowds the documents and the other layers.

A useful context names the business, says what it maintains, and gives a few rules about the shape of the answer. Meridian FM uses this:

Write it as instructions to Centrus, in plain sentences. These belong here:

• What the organisation does, and the trades or contract types it covers, so answers stay inside that work.
• How every answer should be written: British English, pounds, the site named first, or a list rather than a long paragraph.
• A standing limit for the whole business, such as not changing a contracted response time unless the person asks.

These belong somewhere else:

• A price, a part number or a clause. That belongs in a file on Train, so Centrus can quote it and you can replace the file. See Upload files.
• A correction to one wrong answer. That is a lesson, and it can apply to the organisation, to a tag, or to one person. See Teach Centrus from a chat.
• A rule for one group of documents, such as how engineers should read call-out rates. That is tag context, up to 1,000 characters, on the tag. See Tag context.
• How one person wants their own answers written. That is personal context, and it is added on top of this one. See Personal AI settings.

A new organisation starts with a standard paragraph already in the box, and with the model set to GPT-4.1 rather than the recommended GPT-5.6 Terra. Replace the paragraph with your own when you first set Centrus up, and check the model at the same time. Reset to default puts that standard paragraph back into the box. Nothing is saved until you click Save changes.

Centrus does not pick one context and ignore the others. On each answer in a chat it is given the layers in a fixed order, and it is told to use them together with what it found in the documents. Your organisation context does not switch a tag off, and a personal context does not switch the organisation off.

For a chat, the order is: the organisation name, then Organisation Context, then organisation lessons, then the context for the tags on that chat, then the lessons for those tags, then the person's first name, last name and position from the Account tab, then their personal context, then their personal lessons. Lessons are there to build on what the documents say. Centrus reads up to 50 accepted lessons of each kind, the most recently accepted first. Lessons themselves are covered in Teach Centrus from a chat.

Which tags are included follows the chat, not the person's whole set of tags. A chat tagged Engineers includes the Engineers context. It does not include Office, even if Sarah also holds Office, unless Office is on that chat too. When more than one tag on the chat has context, Centrus is given each of them, numbered. A tag with a blank context is left out. Tag context is edited on the tag, not on this page. See Tag context.

An empty organisation box still leaves the organisation heading on the answer, with nothing under it to follow. An empty personal box is left out entirely. Clearing Personal Context and saving stops sending that layer for the person who cleared it. Colleagues are unchanged.

The same layers are used on the web, on WhatsApp and in Microsoft Teams, because those conversations are chats. The personal layer is the person's own, so a WhatsApp conversation uses the organisation context, the tags on that conversation, and that person's personal context.

An agent is given the organisation context when the box has text, together with the agent's own instructions and the tags the agent is set to search. It is not given someone's personal context. The model on an agent can also be set in the Agent Builder, and that model is used for the agent instead of the organisation model. See Open Agent Builder.

Put a rule in the narrowest layer that should have it. A rule for every chat goes in Organisation Context. A rule for one set of documents goes on that tag. A rule for one person's answers goes in their personal context. A fact that should be quoted from a file goes in the file.

Click Show advanced. The first control in the panel is Organisation AI Model. The information icon beside it explains that this sets the default model for everyone, unless a person selects a personal model, in which case their choice overrides it for their account only.

The list is the same list people see when they change model from a chat, in the same order. GPT-5.6 Terra is marked RECOMMENDED. Each row has four marks, for web search, file search, reasoning and image generation, with a tick or a cross, and an information icon. The icon opens a window with the description, a capabilities list, the release date, and Message credits per message, plus 2 when a document is built in the sandbox.

Which model to pick, what each one can do, and how reasoning effort works, is set out in Choose the AI model. The credit for each model, with Terra at 2 for one main answer, is in What each answer costs.

A new organisation starts on GPT-4.1, not on Terra. GPT-4.1 does not reason. If you have just signed up, open this list and move to Terra unless you have a reason to stay. People who already have an older model keep it until an admin changes this setting, or until they pick their own.

The choice here does not change someone who has already picked a Personal AI Model. It changes everyone who has left theirs on Use organisation AI model. Who that includes, including WhatsApp, Microsoft Teams and shared chats, is in Who a change affects.

Under the model, Organisation Temperature shows a number to two decimal places, such as Organisation Temperature: 0.50. Temperature is how much the wording of answers varies. Lower stays closer to the same answer each time. Higher varies more.

Most models do not accept it. For those, the slider is hidden and the page reads Temperature is not supported for the selected reasoning model. GPT-5.6 Terra, Sol, Luna and GPT-6 Astra are in that group. The number is still shown, and it is still stored, but it is not sent with the answer.

The slider appears for GPT-4.1, and also for GPT-5.2 and GPT-5.1. GPT-5.2 and GPT-5.1 only accept the number when reasoning is set to None. At any other reasoning level the stored number is ignored. Reasoning is chosen on the message bar for that browser tab, and it is not saved on this page. See Reasoning effort and temperature.

The information window, What does Organisation Temperature do?, describes the same idea: lower values, around 0.1 to 0.3, for more consistent answers, higher values, around 0.7 to 1.0, for more varied ones, and a moderate setting around 0.5 for most work. A new organisation starts at 0.50. The window also says a person's own temperature replaces the organisation number once they have chosen a personal model. What the server actually does is slightly wider than that, and it is set out under Personal overrides below.

If neither the person nor the organisation has a number stored, Centrus uses 0.50, and then only if the model that answers accepts temperature. For everyone on Terra, which is the usual case, you can leave this where it is.

Organisation Live voice is the voice Centrus uses in Voice mode when the person has not chosen their own. The line under the list reads: Used when someone starts Live and has not set a personal voice. Settings and some messages call Voice mode Live. It is the same feature. How to start it, and what it costs, is in Talk to Centrus.

There are 13 voices. Vesper, a British voice, is the default and is marked RECOMMENDED. The open list shows an accent and a style beside each name. The full set, and the fact that the voice changes the sound and not the language, is listed in Talk to Centrus. A sample that fails to play shows Could not play this voice preview. You can still choose the voice and save it without hearing it.

Voice mode is only in web chat. It is not available on a WhatsApp or Microsoft Teams conversation. It costs 2 message credits for each minute or part of a minute, on top of the usual cost of each answer it looks up. The play button that reads an answer aloud is a different control, and it does not use this voice.

Someone who has set Personal Live voice keeps that voice. Use organisation Live voice, on their Account tab and in the personal half of this panel, follows whatever you save here. A change applies the next time they start Voice mode. It does not change a conversation that is already open.

Two lists set the pictures Centrus generates in web chat, when the picture icon on the message bar is switched on. They do not apply on WhatsApp or in Microsoft Teams. How to ask for a picture, and where the result appears, is in Search the web and generate images.

Organisation image model has two choices. Flare is marked RECOMMENDED in the open list. The line under the list reads: Flare is fast everyday generation. Sunburst is slower, with the highest quality and most precise edits. The closed list shows the name only, so Flare does not repeat the badge once it is selected.

Organisation image quality sets how much extra is spent when a picture is produced. The closed value shows the level and the extra credits, for example Auto (+2 credits). The line under the list reads: Image generation uses extra credits when an image is produced: 1 (low) to 10 (max). Auto and medium are 2. The chat model's usual credits still apply.

Words in the message can replace the saved quality for that picture only. A request that says high quality, or draft, uses that level instead of the one saved here. The line that says so sits under Personal image quality, not under the organisation list: A prompt that clearly asks for high or low quality still overrides this for that message. The same rule applies when the person is following the organisation quality.

A personal image model or quality, if someone has set one, is used instead. On their list, Use organisation image model and Use organisation image quality show your current choice in brackets, for example Use organisation image model (Flare). Leaving those selected means a change you save here reaches them on the next picture.

Nothing on the tab is stored while you are still editing. Discard changes puts every field back to the last saved values, including your own personal fields on this form. Save changes writes the organisation settings, and it also writes your personal fields when they differ from what is already stored.

1. 1Edit Organisation Context, or open Show advanced and change the model, temperature, voice or picture settings.
2. 2Leave Personal AI Model on Use organisation AI model if you want your own chats to follow the model you just set. Pick a model there only if your own chats should use something else.
3. 3Click Save changes at the foot of the page.
4. 4Wait for Updated. If you also changed Personal Context, you will see Your AI context has been updated successfully. as well.

The organisation half reaches everyone who is still following the organisation, from their next message. Earlier answers stay as they were. People who have chosen their own model, voice or pictures keep those, and shared chats keep the owner's model. Who a model change reaches is in Who a change affects. The personal half reaches only the admin who clicked save. Colleagues set their own on Account. See Personal AI settings.

Save changes is unavailable while the organisation save, or the personal context save, is still running. If the organisation save fails, a notification shows Something went wrong, or the message the server sent. If the personal context save fails, the line is Failed to update your AI context. Please try again. The organisation half may already have saved. If the personal model, voice, pictures or temperature fail, the line is Failed to update user, unless a more specific message was sent. Click Save changes again.

Reset to default fills the form with Centrus's starting values, and it does not save by itself. Organisation Context is replaced with Centrus's standard wording, temperature goes to 0.50, voice to Vesper, the image model to Flare and the quality to Auto. It also clears your own Personal Context, personal model, personal voice and personal picture settings. Your personal temperature stays as it was. Pick the organisation model again from Organisation AI Model before you save. Click Discard changes if you pressed it by mistake.

If the voice or picture fields were empty, opening the page already shows Vesper, Flare and Auto. The next Save changes stores those resolved values along with anything else you edited. You do not have to open Show advanced for that to happen.

The lower half of Show advanced is the signed-in admin's own settings, the same ones everyone else has on the Account tab. Personal Context at the top of the page is the same text as Personal AI Context on Account. The placeholder on both boxes is Add personal context to customise AI responses for you... Saving from either place writes the same field.

The information window here is titled What does Personal Context do? It says Personal Context is combined with the organisation context, but only applies to your own conversations. Use it to tailor the AI's responses to your specific needs, preferences, or role, without affecting other users in your organisation. How to write one is in Personal AI settings.

Each personal list that offers Use organisation follows the organisation when that option is selected, and overrides it when a specific choice is selected. The line under the model, the voice, the image model and the image quality says so. Use organisation AI model does not name the organisation model. The image lists do: they show the current choice in brackets, such as Use organisation image model (Flare). Personal Temperature has no Use organisation choice. Its slider stays hidden, and the page says temperature is not supported, while Personal AI Model is Use organisation AI model, even if the organisation model is one that accepts temperature.

Temperature works a little differently. The window What does Personal Temperature do? says the personal number applies only when a personal model is set. In practice, if a number is stored on the account, that number is used whenever the model that answers accepts temperature, with or without a personal model. Otherwise Centrus uses the organisation number, or 0.50 if that is empty too. The form shows 0.50 when nothing is stored yet, and saving while it shows 0.50 stores 0.50. There is no control that clears a stored number back to "follow the organisation". Once it is stored, Organisation Temperature no longer reaches that person on the models that accept it. GPT-5.6 Terra does not accept it, so for most people the number has no effect either way.

The order for model, voice and pictures, and the fact that words in a message can still override picture quality, is in Which setting wins.

These are the messages on the AI Model tab. Messages from the model window inside a chat are in Choose the AI model. Messages from the Account tab are in Personal AI settings.

Anyone with an account follows these settings in chat. Only an admin can open the tab and change the shared half.

People without Train access still follow the organisation context, the model and the voice. Train access is about files, tags and skills. To change who is an admin, see Edit users and roles.