Book a demo
Topics · AI-CONTROL-PLANE

Asset Mgmt AI Gateway How Does It Work | Moring

An AI gateway works by standing between agent frameworks and the models and tools they call. Each request passes through identity checks, routing policy, budget caps, and approval gates; the gateway forwards permitted calls to providers, blocks the rest, and logs every decision. For an asset management firm, that enforcement point is where SR 11-7 model risk documentation and EU AI Act oversight duties become operational.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

An AI gateway works by standing between agent frameworks and the models and tools they call. Each request passes through identity checks, routing policy, budget caps, and approval gates; the gateway forwards permitted calls to providers, blocks the rest, and logs every decision. For an asset management firm, that enforcement point is where SR 11-7 model risk documentation and EU AI Act oversight duties become operational.

How it works

An AI gateway is a reverse proxy with policy. Agent frameworks — the code orchestrating research summarization, commentary drafting, or reconciliation workflows at an asset management firm — send their model and tool calls to the gateway instead of to vendors and internal systems directly. The gateway evaluates each call against the calling agent's identity and the applicable policies, then forwards, transforms, or blocks it. Because every call passes through one place, decisions that would otherwise be scattered across application code become centralized and inspectable.

The evaluation sequence per call is short. First, identity: the gateway resolves which agent is calling, using credentials or an authenticated service identity bound to a registry entry with a named owner and declared scope. Calls from unknown or unregistered agents fail closed by default, which prevents shadow usage from bypassing governance. Second, routing policy: is this model or tool permitted for this agent? An agent scoped to research summarization has no business calling the model reserved for client-facing work. Third, budget policy: token and spend counters per workflow decide whether the call fits within caps, so a runaway loop is a block rather than an invoice surprise.

Fourth, approval gates. For workflows tiered as consequential — client-facing output, investment-relevant analysis — the gateway holds the action until a named reviewer confirms it, recording the confirmation against the agent identity. Fifth, transformation and forwarding: the gateway may rewrite model targets, inject vendor credentials the agents never see, and normalize request formats. Sixth, response handling: outputs are returned to the agent, and streaming is preserved where the workflow needs it. Throughout, the gateway logs the request, the response metadata, the policy decision, and the outcome.

The log is not an afterthought; it is the governance product. Because the gateway sits at the enforcement point, its records are complete by construction: no call exists that did not pass through it. That completeness is what turns model risk documentation from a narrative exercise into an evidence-backed one — under SR 11-7, monitoring and validation of AI capabilities can cite actual call histories, policy decisions, and reviewer confirmations. SR 26-2 expectations in the investment management context assume traceable systems, and the gateway's identity-stamped log delivers the traceability.

Where the EU AI Act applies to a firm's systems, the gateway maps directly onto its duties: record-keeping via the log, human oversight via the approval gates, and transparency via the registry's declared scopes. The gateway also handles practical concerns — provider failover, rate limits, and key management — so that operational resilience is part of the same control point rather than a separate stack.

What the gateway deliberately does not do is understand the substance of the asset management task. It enforces who may call what, at what cost, with which approvals; it does not judge whether a research summary is good. That judgment belongs to the human review workflow and the model risk process under SR 11-7. The gateway's value is making those downstream judgments possible by guaranteeing that the facts — who called what, when, with which result — are captured uniformly.

One further property is worth understanding because it shapes everything else: the gateway's decisions are per call, not per system. Two agents in the same workflow tier can have different scopes; the same agent can be allowed one model today and blocked tomorrow after a policy review; an approval gate can apply to a workflow only while it remains tiered high-stakes. That granularity is what allows the firm to express nuanced positions — research agents read broadly but write narrowly, commentary agents touch client-facing models only through review — without a different integration for each rule. Governance stops being a document describing intentions and becomes a set of evaluated decisions with a record behind each one.

Frequently asked questions

In plain terms, how does an AI gateway work?

It is a reverse proxy with policy: agents send their model and tool calls to the gateway rather than to vendors directly, and the gateway checks each call against the agent's identity and applicable rules. Routing policy decides whether the model or tool is permitted, budget caps decide whether the call fits its allowance, and approval gates hold consequential actions until a named human confirms. Permitted calls are forwarded — often with the gateway injecting vendor credentials — and every decision is logged against the agent identity, producing complete records by construction.

What happens when an unregistered agent calls through the gateway?

By default the call fails closed. The gateway resolves each caller to a registry entry with a named owner and declared scope; a call from an unknown agent has no policy to apply and no legitimate identity to log, so it is blocked and recorded. This is the mechanism that keeps shadow usage from bypassing governance at an asset management firm. It also supports the traceability that SR 26-2 expectations in the investment management context presume and that SR 11-7 model risk documentation relies on when describing which systems perform which actions.

How does the gateway support EU AI Act obligations?

Where the firm's systems are in scope, the gateway operationalizes the Act's duties directly: the registry provides declared scopes and transparency about which agents do what, approval gates implement human oversight for consequential workflows, and the tamper-evident log satisfies record-keeping. Because every call passes through the gateway, the firm can demonstrate compliance from actual records rather than narrative assurance. Asset managers should map each in-scope system to its gateway controls during adoption and keep that mapping current as policies and tiers change.

Does the gateway replace human judgment about agent output quality?

No, and it should not try. The gateway enforces who may call what, within which budget, with which approvals; it does not evaluate whether a research summary is accurate or a commentary draft is compliant. That judgment belongs to the human review workflow the gateway routes work into, and to the model risk lifecycle under SR 11-7, which uses the gateway's evidence — call histories, approvals, monitored outputs — as its raw material. The gateway's contribution is guaranteeing the facts exist uniformly, making downstream human judgment meaningful rather than a signature over an opaque draft.

Sources

  1. paloaltonetworks.com
  2. konghq.com
  3. truefoundry.com
  4. cequence.ai
  5. databricks.com
  6. jfrog.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu