Questions, answered plainly.

Agent authority enforcement limits what an AI agent can do by granting only the permissions its task requires, checked at runtime. Detection-based security tries to recognise and block malicious instructions; enforcement removes the capability, so an injected instruction has nothing to invoke.

Detection works by recognising an attack pattern, which fails against attacks it has not seen. Capability removal does not depend on recognition. If a permission was never granted, no instruction can invoke it, however it is phrased, encoded or split across documents.

Two permissions can each be safe alone and dangerous together. Reading customer records is routine. Calling an external webhook is routine. Combined, they form a path for data to leave. Most access-management systems evaluate permissions one at a time, so the combination goes unflagged.

Five modules are in early access: Runtime Gateway, Agent Access Control Panel, Policy Engine, Agent Observability and Health, and Token and Cost Control. MCP and Computer-Use Guard, Agent Speed Optimization and Agent Playground are coming soon.

Any agent that calls an OpenAI-compatible chat completions API: n8n, LangGraph, CrewAI, Flowise and custom code. MCP interception is coming soon.

Each check takes under 0.05 ms at the 99th percentile in our tests (0.046 ms). The gateway sits in the call path, not the reasoning path, and adds no model round trip. Token verification is offline, with no callback to an issuer.

The gateway runs alongside your agents with nothing blocked. You receive a report of the access each agent holds against the access it used, and every combination that could move data out. You then decide whether to enforce.

A capability enters a policy only if it appears often enough and in enough separate sessions, behaviour seen after the agent read untrusted content counts for less, and unusually busy sessions are discarded. Every generated policy still needs a person to approve it.

The call is paused, not dropped, and appears in the console. The approver can allow it once, add it to the policy, or deny it. Each choice is recorded.

No. The gateway is self-hosted and runs inside your environment. Traces record the shape of each request, such as that a query had a table name and a row limit, never the values.

The honest upper limit on how much of a department’s work can be automated, stated before an engagement begins. It describes the share of work hours automated, not decision authority; decisions stay with named people. Our published figures are estimates from task-level analysis, revised as engagements complete.

Because it is the honest number. Advice and liability are the function of a legal team. Automating the document handling around them does not change that.

Pricing is not published. Book a demo or write to info@startiqosai.com, and the team will share terms for your situation.

The education arm of StartiQOS AI: practical AI courses for students, freshers and working professionals. It runs on its own site, academy.startiqosai.com.