Your AI policy is a request,
not a control.

Per-agent permissions, runtime guardrails,
and an audit trail, enforced at the gateway.

In April 2023, Samsung engineers pasted internal source code into ChatGPT to check it for errors. Within weeks, Samsung restricted generative AI tools on company devices and networks. It was a reasonable reaction. It was also an off switch, and an off switch isn’t governance. It just moves the risk to personal phones and unapproved tools, where you can see it even less.

The governance challenge

Since then, the risk has moved from people to agents. In July 2025, an AI coding agent on Replit deleted a live production database during an explicit code freeze, despite being told repeatedly not to make changes. The freeze was an instruction. It was never a permission.

  • Policy without enforcement. Rules live in documents and system prompts. Agents are asked to behave, and nothing stops them when they don’t.
  • Unclear reach. There’s little visibility into what each agent can touch, from sensitive data to external APIs.
  • No evidence. When the audit committee asks what an agent did yesterday, there’s no record to show.
Quote: Every agent is a new attack surface. Who's monitoring what they actually do?

An AI governance use case example

A 3,500-person motor insurer runs an agent that summarizes claim files and drafts customer updates. Then a claims handler pastes a claimant’s medical notes into a public chatbot to tidy up the wording. The first instinct is to ban it. The better instinct is to make the governed route the easy one. So, the insurer gives its teams approved agents behind a single gateway and enforces policy in the request path rather than the prompt. With approved agents that do the job, people have far less reason to reach for unapproved tools such as public chatbots.

  • Every agent gets its own key, so every call traces back to one agent and one owner.

  • The claims agent can call only its approved models. It can read the claims system through MCP, but it can’t write to payments.

  • Prompt guardrails on each agent’s key inspect requests and responses at the gateway, helping prevent personal data and card numbers from leaving the estate.

  • Budgets, set daily, weekly, or monthly, alert at 80% and stop at 100%, team by team; alerts and anomaly rules you define flag spend spikes and unexpected model calls.

In the first month, the gateway blocks just under 3% of requests, mostly rate limits and one agent repeatedly calling a model nobody approved. Guardrails act on personal data in more than 1,400 requests, redacting or blocking it. The claims team keeps its agent.

(Illustrative example. Figures are modeled, not a customer result.)

The VOSS solution value

VOSS AI Agent Accountability enforces control in the request path. Agents point at the VOSS gateway instead of calling providers directly, so every request through the gateway is checked against policy before it reaches a model or tool.

  • LLM access controls: which models each agent or group can use, with request-rate limits per agent or group.
  • Model portability: change an agent’s model on its key in the VOSS console. It keeps calling the same OpenAI-compatible endpoint, with no rebuild.
  • MCP and tool permissions: read, write, or consequential-write, with new tools defaulting to consequential-write until reviewed.
  • Prompt guardrails: set on each key, they inspect requests and responses at the gateway, helping prevent sensitive data leakage.
  • Alerts and anomaly rules: Alerts on budgets, controls, and policy breaches, plus anomaly rules you define for spend spikes and unexpected model calls.
  • Audit trail: a record of agent activity through the gateway, including what the controls did.
  • Emergency block: cut off a single agent key, or a whole part of your hierarchy, the moment something goes wrong.

Connect an agent in minutes by swapping its model endpoint for the VOSS gateway URL. Then set budgets, model and tool access, and guardrails in the VOSS console with no further agent changes.

What you get

Graduated control, not an on/off switch. This agent, these models, these tools, this budget, this data.
Every control and number is scoped from portfolio to business unit to team to agent, or from MSP to customer.
Evidence instead of assurances, to share with auditors, regulators, and governance committees.
Evidence of control for your EU AI Act and ISO 42001 work: an owned agent inventory, enforced controls, and an audit trail of agent activity. VOSS supports your compliance work but does not certify it or give legal advice.
Governance as the accelerator, not the brake. Teams keep their agents, and the governed route is the easy one, so there’s less pull to work around the rules with shadow IT.

Who this matters to

Security and risks teams icon

CIOs / security / risk teams

who own AI risk but rely solely on static policy documents rather than active controls.

CISO icon

CISOs

who want agents in production without a ban that pushes work into shadow tools.

Compliance and audit teams icon

Compliance and audit teams

who need evidence of what each agent did, and what the controls allowed or blocked.

Business unit leaders icon

MSPs and Service Providers

who resell AI agent governance as a managed service.

Where does your AI policy live? If the honest answer is a document and a system prompt, it’s a request, not a control. For any agent in your estate, could you show what it did yesterday, and what the controls did about it?

Next steps

See every AI agent, govern what it does, prove what it’s worth. Read on

Get full accountability for AI agents with VOSS

Request a VOSS demo button
Request a demo

Discover VOSS through a customized demo that showcases its features and advantages, aiding your decision-making process.

VOSS Free trial button
Take a free trial

Explore the entire VOSS portfolio in your own dedicated SaaS trial environment.

Contact VOSS button
Get in touch

Learn more how VOSS can give you a single digital workplace management platform to help you deliver consistency, control, and efficiency.