PointGuard’s AI Gateway brings model routing, virtual keys and usage budgets together

PointGuard AI introduced AI Gateway on October 29, bringing model routing, usage controls and security‑policy functions into its enterprise platform. The company’s announcement describes a shared place for applications and agents to reach approved models, rather than each team maintaining separate routing and provider credentials.

The reader question is what that central point controls. Choosing a model, attributing a request to a team and deciding whether an agent may use a business tool are related decisions, but they are not the same permission. PointGuard’s release connects those layers while describing distinct mechanisms for each.

PointGuard says the gateway uses virtual API keys to scope access and attribute consumption by tenant, team, application or workload, without distributing the underlying model‑provider credentials. That separates the identifier an application uses through the gateway from the provider credentials behind the service, according to the launch description.

The practical value of that distinction is accountability. A shared provider credential can leave a team looking at an aggregate bill; workload‑level attribution is intended to show where requests originated. It does not, on its own, explain whether a particular request was useful or whether an agent should have taken the next action.

The release also describes hierarchical budgets, rate limits, quotas and calendar‑based resets. Those are different controls around consumption: a spending allocation, the pace of requests, an allowance and its renewal period. PointGuard presents them as ways to apply recurring usage policies across shared AI services. The announcement does not give a complete configuration or billing‑entitlement matrix.

The gateway’s routing covers approved SaaS, open‑source and in‑house models, PointGuard says. Its health‑aware routing includes automatic failover to configured backup providers. A backup therefore belongs to the enterprise’s configured route; the announcement is not a promise that a request can fall back to any model on the market. Source: PointGuard AI

Tool access is a separate layer. PointGuard describes an integration with its MCP Security Gateway for inspecting interactions and controlling access to tools. It also describes enterprise identity and on‑behalf‑of delegation that retain the user context behind a request.

For an enterprise team, the distinction is consequential: permission to send a model request does not itself establish permission to use every connected business tool. Routing and tool authorization need to be understood together. The launch describes how PointGuard connects them, rather than removing that distinction.

For background on those architecture decisions, OpenTools’ Enterprise System Design reference covers authority boundaries, permission‑aware data, routing and rate limits. It is engineering context, not a PointGuard configuration guide or a test of its controls.

PointGuard says integrated guardrails inspect prompts and responses for injection attempts and sensitive data, while the gateway connects with the wider platform’s discovery, observability and runtime‑control functions. It also describes semantic caching intended to reuse suitable responses to similar requests and avoid repeated model calls. These are [the vendor’s product descriptions](https://www.pointguardai.com/blog/scaling‑enterprise‑ai‑requires‑a‑new‑approach‑to‑security‑and‑control), not results of our own security or cost testing. The useful evaluation starts with the workload: which models are approved, which team or application receives the usage attribution, how recurring allowances are configured and which tool permissions remain separate. Those questions identify what an organization needs to establish before treating a gateway as part of its operating policy. The announcement invites enterprises to request a demonstration. It does not establish public self‑service availability, a complete supported‑provider list or prices. It also does not establish that the described controls make an organization’s AI use universally secure or guarantee a reduction in its bill. AI‑generated conceptual illustration of model routing, usage allowances and separate tool permissions. It is not a PointGuard interface, product deployment or numerical result.