
AI is already at work in your company
People use AI to draft, summarize and write code because it makes their work easier. Much of that AI was never registered with anyone.
- Microsoft found 78% of AI users bring their own AI tools to work.¹
- IBM found incidents involving unapproved ("shadow") AI rose to 43% of organizations from 20%, averaging $5.39 million each.²
You cannot govern AI you cannot see. And a control switched on blind can refuse real work on day one.
Three ways it goes wrong
Shadow AI leaves a security team choosing between blind spots and outages.
- No owner, no record. An assistant in someone's browser, a script that sends files to an AI service every night: nobody registered them, so nobody answers for them.
- A ban pushes it underground. People route around a ban because AI makes their work easier, and the blind spot gets bigger.
- Controls switched on blind. A control turned on before anyone knows what it will hit refuses real work, and teams push back.
The one idea: find it first, in three levels
Argent's Discovery page starts with the proxy you already run, then adds managed machines, then developer registration. Each level adds detail.
| Level | What Argent uses | What you see |
|---|---|---|
| Level 1: Proxy log feed | A copy of the access log your web proxy already writes, in Squid or NDJSON format. Nothing is installed and no traffic is rerouted. | Which AI services people reach, and the login the proxy named. Tagged "Seen via proxy log feed". |
| Level 2: Managed machines | Your device management (Jamf, Kandji, Intune-style or shell/PowerShell scripts) installs the Argent certificate and proxy settings once. Only machines from ranges an administrator approved are admitted. | Which AI services each approved machine reached, by its network address. Tagged "Seen via managed network". |
| Level 3: Developer registration | A developer registers the agent with a named owner and installs its certificate. | The full picture: the agent's identity, its owner and every request it makes. |
- Discovery keeps who reached which AI service, never what was said. Argent keeps only visits to AI services it recognizes and drops every other log line before anything is stored.
- At Level 2, traffic is tunnelled and never decrypted.
- For people who go straight to AI websites there is Browser Guard, a small add-on for Chrome and Edge that your device management installs. It shows which AI sites people open, blocks the ones you choose, and on the AI sites you approve it can check a message for sensitive data before it is sent.
How it works, step by step
Discover, register, then roll controls out. Discovery only records; controls apply only to agents you register, and only after you have watched what they would do.
- Switch Discovery on. It is off until an admin turns it on for the tenant, under Policy > Settings.
- Add a log feed. Under Sources, add a Log feed. Argent shows its token once and keeps only a hash; rotate the token to get a new one.
- Point your proxy's log shipper at Argent. Your proxy keeps running exactly as it does today, and Level 1 only records what it already did.
- Claim your managed networks. A range reads "Awaiting SuperAdmin approval" and admits nothing until it is approved. A range that overlaps another tenant's is refused.
- Review what turned up. Every client the proxy log feed or a managed network sees, but nobody registered, appears in the Agent inventory (open it from the Discovery page) tagged Discovered, with how it was seen.
- Register what you keep. Developers register agents with a named owner, so each gets an identity and a record of every request. Argent also learns how each registered agent normally behaves and flags one that starts acting differently.
- Watch, then enforce. In a new tenant every control starts in watch: requests pass unchanged and what each control would have refused or changed goes on the signed record. Then switch controls to enforce, one at a time. See Watch Before You Enforce.
What the operator sees
Plain console words, not a wall of charts.
| Where | What it shows |
|---|---|
| Discovery page | Each level's state: Off, Awaiting approval, No data received or Receiving data. A feed that goes quiet reads "No data received", never "no AI usage": silence is not safety. |
| Agent inventory | Rows such as "Discovered agent 7f3a9c", tagged Discovered and "Seen via proxy log feed" or "Seen via managed network". Filter by Source; tiles for Active, Dormant, Agent types and New. It stores which client, version and AI model, never what was said. |
| Rollout page | For registered agents in watch: WOULD REFUSE and WOULD CHANGE per control, SAMPLE EVENTS with BEFORE and AFTER, and a RIGHT or WRONG label on each. Before a switch, it states how many requests that control would have refused in the last 7 days. |
For registered agents whose requests run through Argent, watch mode shows six things:
- All the AI in use, and where it comes from.
- What sensitive data is going out, by kind.
- What it costs, by agent, model and provider.
- What Argent would have stopped, and why.
- Which models each agent calls, and how many tokens.
- The prompts, responses and tool calls behind each request.
Every switch, in both directions, needs a fresh passkey check and is a signed event. Some protections never switch to watch: console logins, revoked or expired keys, quarantine and kill switches still stop.
What frameworks already ask for
Frameworks and regulators already ask for an AI inventory.
- NIST AI Risk Management Framework, GOVERN 1.6: "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities."³
- For banks and insurers supervised by the New York Department of Financial Services (NYDFS): its AI guidance adds no new requirements, and says entities that rely on AI "should maintain an inventory of all such systems."⁴
The Agent inventory, with its Discovered and Registered rows, is where that inventory starts. Discovery rows and every record of what a control would have done are signed, ready for one-click audit evidence.
Who it serves
Any team that has to answer "what AI are we using, and who owns it".
- Goals: See all AI in your company, data loss prevention, agent identity and ownership, audit and accountability, compliance evidence and budget and token spend.
- Industries: Financial Services, Insurance, Healthcare, Public Sector, Retail and all industries.
- Use cases: Rolling out copilots and chat safely, Securing AI coding agents and AI in regulated and air-gapped environments.
- Next capabilities: Identity Tethering gives each registered agent an owner, Agent Behavior Watch flags drift, Sensitive Data Redaction and Model Control enforce, and AI Spend Caps put a ceiling on cost. All capabilities.
- Where it runs: inside your own environment. Compare deployment options.