
ARGENT FORCOPILOTS AND CHAT
Master Control System · by Nyedis
Everyone gets a copilot, and every prompt leaves the building
A large company rolls out a chat assistant to thousands of employees. People paste in what they are working on: a customer email, a contract, a claim file. Each prompt goes to an outside AI model, and nobody checks it on the way out.
- In one study of employee prompts to AI tools during 2024, 8.5% contained sensitive information. In the same study, 63.8% of ChatGPT users were on the free tier, and 53.5% of the sensitive prompts were entered there.¹
- In IBM's 2026 study of 602 organizations that suffered a breach, incidents involving unapproved ("shadow") AI more than doubled, to 43% from 20%.²
- OWASP's Top 10 for LLM Applications lists sensitive information disclosure, and says users must understand the risks of unintentionally providing sensitive data.³
- A copilot that answers from company documents can hand one person another team's files. OWASP advises checking each person's permissions before a company document is used to answer them (fine-grained access controls and permission-aware vector stores).⁴
The rollout team is asked three questions on day one: what did the copilot send, who pays for it, and which models is it allowed to use. Without a checkpoint, nobody can answer.
One checkpoint for every copilot
Argent Strategic Intelligence sits between your copilots and the AI models behind them. Every request routed through it is checked before it leaves and again when the answer comes back.
| Check | What Argent does |
|---|---|
| Who | Each copilot is registered as an agent with a named human owner. A request from an agent with no active owner is refused. |
| Data | Social Security, card and account numbers are swapped for a redaction marker before the request leaves. The rest goes through and the copilot still answers. |
| Models | Only models on your allowlist answer. A request for another model is refused, or swapped to an approved one, with no change to the copilot. |
| Spend | Each call is priced before it is sent and checked against the budget that pays for it. |
| Threats | Text the copilot's tools bring back is marked as data, not orders, and prompts that Argent's attack detector flags as an injection or jailbreak attempt are refused. |
| Replies | The model's answer is screened for sensitive data before the person sees it. |
| Record | Every decision lands on a signed, tamper-evident record. |
Once a control is enforcing, a request it cannot check is refused, not sent.
A walk-through: the first week of a company copilot
Picture a company that puts Open WebUI in front of Azure OpenAI for its staff and routes it through Argent. Here is what happens in its first week.
- Registered with an owner. IT registers the copilot in Argent as an agent owned by the head of digital workplace, with its own budget under the operations department. Staff sign in to the copilot as usual.
- A rep pastes a customer email. Dana in account services asks the copilot to draft a reply. The email holds the customer's card number.
- The card number stops at the gate. Argent finds it by its shape and checksum and swaps it for a redaction marker. Azure OpenAI receives the redacted request and drafts the reply. With restore switched on, Argent puts the original number back into the answer, so Dana gets a usable draft and the model never saw the number.
- The reply is screened. If the model's answer carries sensitive data, Argent redacts it, or withholds the whole reply, before it reaches the screen.
- A tester picks an unapproved model. Someone points the copilot at a model nobody approved. Argent refuses it with a plain reason, or swaps it to the approved model you chose.
- The budget speaks up. Alerts fire at 70, 90 and 100 percent of the copilot's budget. When it runs out, the owner can request more, and nothing is granted until an approver decides.
- The record answers the audit. The signed record names the kind of data and how many, never the value. An authorized administrator can open the conversation in Session Replay.
What the rollout team sees
One console for the security team, the platform team and finance, each fenced to what their role allows.
| Where | What it shows |
|---|---|
| Discovery | The AI services your people already reach, from your proxy logs and managed machines, so you can steer them to the approved copilot. |
| AI Clients | A catalog of 83 AI tools that shows which ones route through Argent, with setup steps and a verification check for each, including Open WebUI, LibreChat and Claude Desktop on third-party inference. |
| DLP | Scan scope, what happens on detect, and which kinds of data are on. Every change is a signed event. |
| Replay | A conversation turn by turn, with what the copilot sent beside what Argent forwarded. Reading the raw copy is itself recorded under the reader's name. |
| Budgets | A spreadsheet-style sheet by company, org, department, person and agent, with spend, approvals and caps. |
| Evidence | A signed evidence package an auditor can check on their own machine. |
Branch admins see only their own org or department. See Session Replay and One-Click Audit Evidence.
The capabilities and integrations involved
Capabilities
- Sensitive Data Redaction: customer numbers stop at the gate. The film above shows it end to end.
- Response Screening: the model's answer is checked before anyone sees it.
- Model Control: only approved models answer.
- AI Spend Caps: budgets that refuse what they cannot cover.
- Prompt Injection Defense: planted orders in documents are treated as data.
- Shadow AI Discovery: find the AI tools people already use.
- Identity Tethering and Owner Offboarding: every copilot has an active owner.
- Keyless Agents: the company's model key stays in the vault.
- Watch Before You Enforce: see what each control would do first.
Integrations
- Chat front ends: Open WebUI, LibreChat, Claude Desktop on third-party inference.
- Model routers: LiteLLM.
- Models: Azure OpenAI, Amazon Bedrock, OpenAI, Anthropic, Google Gemini, and self-hosted Ollama or vLLM.
- People: Okta, Microsoft Entra ID, SCIM 2.0, OpenID Connect sign-in.
- Rollout: Microsoft Intune, Jamf, Kandji.
- Data and documents: Microsoft Presidio, and OpenFGA to decide which documents may reach the model.
How a large rollout starts
Watch first. Nothing changes for staff until you decide.
- Find. Feed Discovery your proxy logs, or push the Argent proxy settings to managed machines with Intune, Jamf or Kandji, and see which AI services people already reach.
- Route. Point the copilot's model connection at the Argent gateway and register it with an owner. The model key moves into your vault.
- Watch. Each control records what it would have redacted, refused or capped while requests pass unchanged.
- Enforce. Switch on one control at a time, starting with data protection, then models and budgets.
- Inside your walls. Argent runs in your own Kubernetes cluster, as a single-node bundle, or air-gapped on premises. Compare deployment options
See the whole platform, take the walkthrough, or book a demo.
Where it fits in your program
The same checkpoint governs every AI agent you route through it, not only the copilot.
- Goals: See all AI in your company, Data loss prevention, Budget and token spend, Model control and routing, Threat defense, Audit and accountability, Compliance evidence. All goals
- Use cases: Governing customer-facing AI assistants, Securing AI coding agents, Controlling agent tool access: MCP and agent-to-agent, AI in regulated and air-gapped environments.
- Industries: Financial Services, Insurance, Healthcare, Public Sector, Retail. All industries
- More: All capabilities, data sheets, contact sales.