AI Assistant Configuration
Create and manage your organization's AI assistants — their instructions, knowledge, tools, model, and who can use them. As an admin, you decide what each assistant knows, what it can do, and which people can reach it.
:::info Who can use this By default only Admins can create, edit, or approve assistants and manage access. Your organization can also give some or all of these actions to other people through a custom role — see Roles & Permissions. Team members see the assistants they're allowed to use, but can't change them. :::
Open the AI Assistants page
In the sidebar, open WorkSpace → AI Assistants. The page header reads AI Assistants with the subtitle "Configure assistant permissions and knowledge base access"
With full access you'll see three tabs. If your role covers only part of assistant management, you see only the tabs and buttons it allows.
| Tab | What it's for |
|---|---|
| Assistants | The assistant list — search, view, edit, delete, and turn each one on or off |
| Pending Approvals | Review assistants that were delivered to your organization and set their access before they go live |
| Access Requests | Approve or reject requests from users who want access to a restricted assistant |
The Pending Approvals and Access Requests tabs show a small badge when items are waiting for you.
The Assistants list
The Assistants tab lists the assistants in your organization. Use the controls at the top to find what you need:
- Search assistants... — filter by name or description.
- The status filter — All Status, Active, or Inactive. The page opens on Active, so inactive assistants stay hidden until you choose All Status or Inactive.
Each row shows the assistant's avatar and name, its Model, its Conversations total, and a Status toggle:
| Column | Meaning |
|---|---|
| Assistant | Avatar, name, and short description |
| Model | The model the assistant runs on |
| Conversations | How many conversations it has handled |
| Status | An Active / Inactive switch — flip it to turn the assistant on or off |
The list shows 10 assistants per page. When there are more, page buttons appear under the table together with a line such as "Showing 1 - 10 of 24".
Click any row to open a detail panel showing the assistant's description, notes, base model, conversation total, and system prompt. At the bottom of the panel are Edit Assistant, Delete, and Close — you see only the buttons your role allows.
Turning an assistant Inactive hides it from users without deleting it. Use this to pause an assistant while you refine it.
Delete an assistant
- Click the assistant's row to open the detail panel.
- Select Delete.
- Confirm in the Delete assistant window.
The assistant is archived and removed from the list. If you only want to take it out of use for a while, set it to Inactive instead.
Create an assistant
Select Create Assistant (top right). A form opens with six tabs. You can move between tabs freely; required fields are checked when you save, and a red dot marks the Basics or System Prompt tab if a required field there is still missing.
1. Basics
| Field | Required | What it does |
|---|---|---|
| Name | Yes | The assistant's name, shown to users. Up to 100 characters |
| Description | No | A short summary of what the assistant does. Up to 500 characters |
| Notes | No | Extra text shown to users to introduce the assistant. This does not affect how the assistant behaves — it only guides people. Up to 2,000 characters |
| Starter Questions | No | Up to 4 short questions (50 characters each) shown as suggestion chips on the assistant's welcome screen |
| Suggest follow-up questions | No | When on, clickable follow-up questions are suggested after each answer. On by default |
| Character / Emoji | No | Pick an animated character as the avatar; if you don't choose one, an emoji is used instead |
A new assistant's form already contains four general starter questions, written in English. Edit them, or remove the ones you don't want, before you save.
2. Model Config
- Model — choose which model powers the assistant. Search the list or filter it by provider. Your platform model is always available; external providers appear here only if you've enabled them on the Models & Providers page. A model is required.
- Model Parameters — fine-tune behavior with Temperature, Top P (Nucleus Sampling), Top K, Max Tokens, Frequency Penalty, and Presence Penalty. Tick a parameter to set it; any parameter you leave unchecked uses the model's default.
:::caution External models send data outside your organization If you pick an external provider's model, the assistant shows an External Model warning: conversations are processed on that provider's servers. For sensitive work, prefer your platform model. See Models & Providers for how providers and data handling are set up. :::
3. System Prompt
The system prompt is the assistant's core instructions — who it is, how it should behave, and what it should and shouldn't do. This field is required. Write it in clear, plain language, as if briefing a new team member.
A system prompt can be up to 32,000 characters. The counter in the corner of the editor shows how much you have used. If you go over, the editor shows "System prompt exceeds 32K character limit. Please shorten it." and the assistant can't be saved until you shorten the text.
Template variables
Under the editor, the Template Variables box lists placeholders the assistant fills in for each conversation. Click a variable to insert it where your cursor is, or click the copy icon next to it.
| Variable | Replaced with |
|---|---|
{{tenant_name}} | Your organization's name |
{{user_name}} | The name of the person chatting |
{{user_role}} | That person's role in the organization |
{{department}} | That person's department |
{{date}} | The current date |
{{time}} | The current time |
Draft with AI
If you'd rather not start from an empty box, let the platform write a first draft.
- On the System Prompt tab, select Draft with AI.
- Under What should this assistant do?, describe the assistant in a few sentences — what it does, who it serves, and what it must not do. You have up to 2,000 characters.
- Select Create draft. The panel shows "Writing the draft" with a timer. This can take a few minutes, so keep the page open.
- When the draft is ready it appears in a box marked AI suggestion, with a line that lists the knowledge sources the assistant is connected to. Nothing has changed in your system prompt yet.
- Choose Accept to put the draft into the system prompt field, or Reject to go back to your description and try again.
After you accept, a notice above the editor offers Undo, which brings back the text that was there before. Read the draft, adjust it, and then save the assistant as usual.
Good to know:
- While a draft is being written, the other tabs, History, and closing the form are locked. To stop waiting, select Stop drafting — your description is kept.
- The draft is written in Turkish.
- There is a limit on how many drafts you can create in an hour. When you reach it, a message tells you the count resets at the top of the hour.
- If a draft can't be created, a message explains why. Select Try again to return to your description.
- A draft you save without changing it is labelled AI-assisted in the assistant's version history.
Checks while you write
The editor looks at the system prompt as you type and shows notes under the text box.
| What you see | What it means | What you can do |
|---|---|---|
| "The system prompt contains invisible characters or extra spacing: …" | The text has invisible characters, a space at the start or end, or three or more blank lines in a row — often left over from pasting | Select Clean up to remove them |
| "Unrecognized variable: …. It is sent to the chat as written." | A placeholder in double curly braces isn't one of the template variables, so it won't be replaced | Correct the name or remove it |
| "This paragraph appears 2 times: …" | The same paragraph is in the prompt more than once | Select Show to jump to the repeat |
| "Line 12: text that looks like a key or password." | Part of the prompt looks like a key or password | Remove it from the prompt |
Only the last one interrupts a save. If the prompt still looks like it contains a key or password when you save, the bottom of the form asks you to decide: Back to editing or Save anyway.
Don't put passwords, keys, or other secrets in a system prompt. Every saved version is kept in the assistant's version history and can't be deleted from within the product.
4. Knowledge Base
Decide what documents and websites the assistant can draw on.
- Enable Knowledge Base — let the assistant search and reference your organization's documents. When it's off, the assistant answers from its instructions and model knowledge only.
- Document Access Scope — once Knowledge Base is on, choose which document levels the assistant may search:
| Scope | What it covers |
|---|---|
| User's Personal Documents | Documents the person chatting has uploaded |
| User's Team Documents | Documents from that person's teams |
| User's Department Documents | Documents in that person's department |
| Organization Documents | Company-wide documents everyone can see |
| Confidential Documents | Restricted documents that only admins can view |
- Assistant-Specific Documents — optionally build a document collection just for this assistant. Tick Enable Assistant-Specific KB, then choose whether it enforces each user's permissions (people only see what they're allowed to) or bypasses them so everyone using the assistant can see those documents. Use bypass with care. A document picker then appears: search your documents and tick the ones to attach, or drop files in to upload them.
- Answer from Websites — off by default: a crawled website isn't used in chat until an assistant is turned on for it here. Enable this to let the assistant also answer from your organization's crawled websites. Website pages follow the same rules as documents — the scope levels you picked above apply to them too. See Website Sources to add and manage the websites your organization crawls.
- Enable Web Search — let the assistant look up current information from the web. This is the same setting as the web search tool on the Tools tab: when the assistant has tools selected, ticking one ticks the other.
When you edit an existing assistant, its attached documents are listed here. Remove one with the × on its row, or tick more documents and select Attach selected. Both take effect straight away — you don't need Save Changes for them.
Attaching documents needs its own permission. Without it, the picker isn't shown and you see "You don't have permission to attach documents. Someone who does can add them once the assistant is created."
:::info How document access combines An assistant only ever answers from what it — and the person using it — can see. Selected scopes are combined, users still only reach documents their own permissions allow, and confidential documents are searched only for admins. To learn how document levels work, see Knowledge Base access levels. :::
5. Tools
Tools are the actions an assistant can take — for example searching, working with documents, or creating tasks. They're grouped by category, such as Search, Documents, Organization, Tasks, Service Desk, Blueprints, HR & People, and Utility.
- Select the tools the assistant should have. The line above the list shows how many are selected.
- Use Select All or Clear to set everything at once.
:::caution Leaving the list empty does not turn everything on With no tools selected, the counter reads "No tools selected — limited legacy set". The assistant then falls back to a small set — web search, the knowledge base, and organization information — and it cannot create documents, tasks, or presentations. Select the tools you want the assistant to have. :::
Some tools carry labels so you know what you're granting:
| Label | Meaning |
|---|---|
| WRITE | The action changes data (not just reads it) |
| Admin Only | Only admins can use this action |
| Admin / Exec | Available to admins and executives |
| Manager+ | Available to department managers and above |
Even when a tool is enabled on the assistant, a person can only use it if their own role allows it.
6. Advanced
- Enable Thinking Mode — show the assistant's reasoning steps before it gives its final answer.
- Department Access — shown only when you edit an existing assistant. Tick the departments that may use the assistant, then select Save next to the heading (the button appears once you change the selection). It is saved on its own, separately from Save Changes. See Control who can use an assistant.
When you're done, select Create Assistant to save. (When editing an existing assistant, the button reads Save Changes.)
A new assistant is Active and open to everyone in your organization as soon as it is created. To narrow who can use it, see Control who can use an assistant.
Edit an assistant
Click the assistant's row, then select Edit Assistant in the detail panel. The same six tabs open with the current settings filled in. Change what you need and select Save Changes.
Version history
Each save that changes how the assistant behaves is recorded as a version, so you can see what changed and bring back an earlier system prompt.
- On the System Prompt tab, select History. The Version history view replaces the form; your unsaved edits stay as they are.
- Pick a version from the list. Versions are listed newest first and grouped by day (Today, Yesterday, then by date). Each row shows the version number, how it was made (for example Edit, AI-assisted, or Restore), the time, who saved it, and which settings changed. The newest one carries a Current tag. Use Show older versions to load more.
- Under the list, read the comparison between that version and the saved current state:
- System prompt — a line-by-line comparison. Added lines are marked with +, removed lines with −, and a count of each is shown.
- Other settings — a table of the settings that differ, with one column for the version you picked and one for Current. Select Also show unchanged settings to see the rest.
- To bring back an old prompt, select Load system prompt into editor.
- Select Back to editor to return to the form.
Loading a version only puts its system prompt into the editor. The model and every other setting stay as they are, and nothing is saved until you select Save Changes — which creates a new version. A notice above the editor offers Undo if you change your mind.
If someone else saved while you were editing
If a colleague saves the same assistant while your form is open, your save is held back and a notice appears at the top of the form: "Someone else saved this assistant while you were editing." Your changes are not lost. Choose one of:
| Option | What it does |
|---|---|
| See differences | Shows your unsaved system prompt next to the latest saved one |
| Load latest | After you confirm, replaces everything in the form with the latest saved state. Your unsaved changes are discarded |
| Save anyway | Saves your version over the latest one |
Pending Approvals
A ready-made assistant that is delivered to your organization waits in Pending Approvals before it goes live. Assistants you make yourself with Create Assistant don't pass through here — they are active straight away.
Expand an entry to review it and set two things:
- Knowledge Base Access — tick which document levels it may search: Personal Documents, Team Documents, Department Documents, Organization Documents, and Confidential.
- Access Level — who can reach the assistant:
| Access Level | Who gets it |
|---|---|
| Entire Organization | Everyone in your organization |
| Specific Departments | Only the departments you pick |
| Specific Users | Only people you give the assistant to, one by one |
When you choose Specific Departments, select the Allowed Departments from the list.
When you choose Specific Users, an Allowed People block explains that people are given the assistant one at a time, from each person's panel on the Users page. It also warns: "Until at least one person is given it, nobody will be able to use this assistant." Admins see a Save settings and open Users button there, which saves your choices and takes you to the Users page — see Give an assistant to one person.
Select Save Settings to keep your choices, then:
- Approve — the assistant goes live and becomes active.
- Reject — the assistant is archived. You can add an optional reason.
Access Requests
When a user asks for access to a restricted assistant, the request appears in Access Requests with the requester's name, email, the duration they asked for, and the request date.
- Approve — a dialog opens where you can set an Access Expiry Date. Leave it empty to use the duration the user requested.
- Reject — you can add an optional reason.
Control who can use an assistant
An assistant you make with Create Assistant is open to everyone in your organization. An assistant that came through Pending Approvals has the access level you chose there. Either way, you can change it afterwards.
| You want | Where to set it |
|---|---|
| Everyone in the organization | On the Departments page, select Manage Assistants in the Organization-wide Assistants card and tick the assistant |
| Only certain departments | On the Departments page, untick the assistant in Manage Assistants. Then edit the assistant and tick the departments under Department Access on the Advanced tab |
| Only certain people | Give the assistant to each person from the Users page (below) |
Whatever you choose, people whose system role is Admin or Executive can use every assistant.
See Departments for the department side of this.
Give an assistant to one person
- Open Organization → Users and open the person's details.
- In the Assistant Access section, select Give another assistant.
- Under Access for, choose how long the access lasts: No expiry, or 1, 7, 30, or 90 days. Set this first — the assistant is given the moment you pick it.
- Search for the assistant and click it.
The Assistant Access section lists every assistant the person can reach and why: Everyone, Department, or Given directly. Only Given directly rows can be removed there — select the × on the row and confirm.
Tips
- Keep your default assistant active. Arketic AI is most people's starting point — leaving it on gives everyone a reliable place to begin.
- Scope knowledge deliberately. An assistant only answers from what it and the current user can see. Grant the narrowest document access that still gets the job done.
- Always pick the tools. An assistant with no tools selected can't create documents, tasks, or presentations.
- Prefer the platform model for sensitive work. External models process conversations on the provider's servers.
- Use Notes and Starter Questions. They help users understand what each assistant is for and how to begin.
- Review approvals and access requests promptly so new assistants and waiting users aren't held up.
Related pages
- Models & Providers — enable providers and control how conversation data is handled
- Knowledge Base access levels — how document visibility works
- Website Sources — add and manage the websites your organization crawls
- User Management — manage the people who use your assistants
- Roles & Permissions — what each role can see and do