# ColossalX Assistant: governed AI chat for every employee.

> ColossalX Assistant is governed AI chat for a whole workforce, from the same login as the ColossalX console. Personal data is redacted before it reaches the model, each group gets its own approved models and provider key, the knowledge base is screened for injected instructions, and the guardrails that stepped in are shown on the answer.

Your people get capable AI chat from the same login, with personal data redacted before any model sees it, under your security team's policies.

Canonical page: https://colossalx.tech/assistant · Last reviewed: 6 Oct 2026

*Illustration:* Assistant · before the model: employee to ColossalX Assistant (prompt); knowledge base to ColossalX Assistant (cited); ColossalX Assistant with data redacted, to group model (tax ID redacted); ColossalX Assistant to self-hosted model (per group).

## The threat and the control

- **The threat:** People paste customer data into consumer chatbots when the approved tool is slower or missing.
- **The control:** ColossalX Assistant gives them governed chat that redacts personal data before the model sees it.

## The model saw a placeholder, not the number.

Personal data in a prompt or an attached file is redacted before it reaches the model. The person sees a notice saying how many items were removed, and the reply deals with the placeholder.

- **Redacted before sending.** Prompts and attached files are redacted before they reach any model.
- **Shown on the answer.** The guardrails that stepped in, such as redaction or an excluded document, are listed with the answer.
- **One engine, two products.** The console's own assistant redacts with the same engine, under each person's role.

*Illustration:* An illustrative ColossalX Assistant conversation: the tax ID in the message is replaced with a token before it leaves, so the model answers about the token and never sees the identifier, and the answer shows the redaction and the group model that answered. Notes: 1. Tax ID replaced first 2. The redaction, shown on the answer 3. The group model that answered

## Each group gets its own models and key.

Administrators choose which model categories each role or SSO group may use. They also choose which provider key its chat runs on, and the server enforces both, not the browser.

- **Wide model choice.** Hosted providers or self-hosted models such as Ollama and vLLM, on your keys or ours.
- **Chat-only role.** Map one SSO group to the chat-only role, and its members get chat, not the console.
- **Confidential content blocked.** When confidential content is detected, the request can be blocked before any model answers.

*Illustration:* Model access · per group: everyone-else group to Deep Thinking model, "Use the Deep Thinking model for this contract summary.". Checks: Signed in, chat role passed, Model in group bundle failed. Verdict: refused, Not in their bundle.

## A knowledge base screened before it answers.

Each passage retrieved from the documents you add is screened for injected instructions. The model sees it only after that check; answers cite their sources, and an excluded passage is never dropped silently.

- **Excluded, and said so.** When a passage is excluded as suspicious, the answer carries a note saying so.
- **Writers need permission.** Adding documents needs its own permission, separate from chatting with them.
- **Cited answers.** Answers show the documents they drew on, so a person can check the source.

*Illustration:* Knowledge base · one passage out: employee to knowledge base, "What is our travel refund policy?". Checks: Passage screened for injection flagged, Personal data passed, Sources cited passed. Verdict: allowed, Answered, one passage excluded.

## Your brand, your budget, one login.

Give people a standalone chat window in your colours and words. Finance sees usage and cost by model and by conversation, tracked against a budget.

- **Standalone and branded.** A minimal chat window with your name, logo, colour and welcome message.
- **Same policies as the console.** Chat runs through the same AI gateway, guardrails and provider keys as the console.
- **Usage by model and chat.** Tokens and cost per model and per conversation, with the average cost of a conversation.

*Illustration:* Usage · questions it answers: What did each model cost? ends in Cost per model; Which chats cost most? ends in Cost per conversation; Are we within budget? ends in Spend against budget; Who may use what? ends in Group model bundles.

## What ColossalX does not do

- Usage and cost are tracked against a budget; reaching it does not stop chat.
- Redaction removes the personal data it recognises; it cannot vouch for data it does not.
- Answers cite knowledge-base sources; other claims in an answer are not verified.
- Delivered as SaaS only, inside the same workspace as the ColossalX console.

*Illustration:* ColossalX Assistant · stated plainly: Login same as the console; Models per role or SSO group; Personal data redacted before the model; Budget tracked, not a cap; Delivery SaaS only. Stated, not implied.

## Questions

### What is a governed AI assistant?

A governed AI assistant is AI chat a company runs under its own rules: which people may use which models, what happens to personal data before it reaches a model, which documents the assistant may draw on, and what it all costs. ColossalX Assistant applies those rules on the server, from the same login and workspace as the ColossalX console.

### How is it different from a consumer AI chatbot?

A consumer chatbot follows its provider’s rules. ColossalX Assistant follows yours: personal data is redacted before the model sees it, each group reaches only the models and provider key you assign, your knowledge base is screened before it is used, and usage and cost are visible by model and by conversation. It runs inside your ColossalX workspace.

### Which models can employees use?

The ones you approve for their group. ColossalX Assistant connects to a wide range of hosted providers and to self-hosted models such as Ollama and vLLM, on platform-managed keys or your own. Administrators curate model categories per role or SSO group, and the server refuses a model outside a person’s bundle.

### How is personal and confidential data protected in chat?

Personal data in a prompt or an attached file is redacted before it reaches the model, and the person sees a notice saying how many items were removed. Passages retrieved from the knowledge base are screened for injected instructions first, and answers list the guardrails that stepped in. Redaction covers the data it recognises; it cannot vouch for what it does not.

### Can employees use our own documents in chat?

Yes, through the knowledge base. People with permission to add documents upload them; everyone else can chat with them. Each passage retrieved for an answer is screened for injected instructions before the model sees it, an excluded passage is noted on the answer, and answers cite the documents they drew on so a person can check.

### How do we roll it out to a large workforce?

Map an SSO or directory group to the chat-only role, and everyone in it is entitled with one mapping. Chat-only users get chat, not the governance console, and the server enforces that. A standalone chat window with your name, logo, colour and welcome message gives them their own front door.

---

ColossalX is an AI security and governance platform from Quantexra Labs LLP, delivered as SaaS. Book a walkthrough: https://colossalx.tech/demo · client.success@quantexra.tech