Asset Mgmt AI Gateway Model Calls | Moring
Model calls are the unit of AI gateway enforcement at an asset management firm: every request from an agent to a model passes through identity checks, routing policy, budget counters, and approval gates, and every outcome is logged. Governing calls — rather than workflows in the abstract — gives SR 11-7 model risk records concrete evidence and makes cost, vendor use, and oversight measurable.
Short answer
Model calls are the unit of AI gateway enforcement at an asset management firm: every request from an agent to a model passes through identity checks, routing policy, budget counters, and approval gates, and every outcome is logged. Governing calls — rather than workflows in the abstract — gives SR 11-7 model risk records concrete evidence and makes cost, vendor use, and oversight measurable.
How it works
The model call is the atomic unit of agent behavior, which is why the AI gateway governs there. A workflow — summarizing research, drafting commentary, reconciling positions — decomposes into individual calls to individual models, and each call is where identity, policy, and evidence can be applied without ambiguity. Asset management firms that try to govern at the application layer instead discover the governance dissolves: one codebase makes dozens of calls through several code paths, and no policy statement can attach to all of them reliably.
Per call, the gateway runs a fixed sequence. Identity resolution maps the caller to a registry entry with an owner and declared scope. Routing policy decides whether the requested model is permitted for that agent — the commentary agent uses the approved vendor, the research agent cannot reach models bound to client-data-restricted datasets. Budget counters check the call against the workflow's token and spend allowance. Approval gates hold calls tiered as consequential until a named reviewer confirms. Only then does the gateway forward the request, injecting provider credentials the agent never holds, and normalizing the request format for the target provider.
Responses get the same treatment. The gateway records the outcome — success, failure, block — with response metadata, and streams the result back where the workflow needs streaming. Where evaluation or content checks are configured, they apply to responses at the same point, so monitoring sees the full exchange. The record for each call is therefore complete: which agent, which model, what policy applied, what was approved or blocked, what came back.
That per-call record is what makes model risk management concrete. Under SR 11-7, documentation and monitoring of AI capabilities need evidence across development, validation, and operation; call histories tied to agent identities are the operational evidence, and they accumulate without extra effort because every call passes through one point. SR 26-2-style vendor oversight in the investment management context also resolves at the call level: which provider served which agent, for which workflow, at what volume and cost. Budget counters turn that into spend records, so the vendor relationship is measured, not estimated.
Treating calls as the unit also clarifies change management. When a model version changes — provider upgrade, vendor swap, capability shift — the change is a routing policy edit at the gateway, reviewed and applied like any control change, rather than a code change replicated across every agent codebase. The EU AI Act, where the firm's systems are in scope, benefits similarly: record-keeping and human oversight attach to calls, so the duties are demonstrable from the enforcement point rather than reconstructed per application.
The discipline is to keep the call path singular. Direct vendor connections from production agents, keys embedded in application code, or side channels to internal models all break the invariant that every call passes through identity, policy, and evidence. Where legacy systems cannot route through the gateway, the exception should be documented, time-bounded, and monitored — because an ungoverned call is not just a missing log entry; it is a workflow whose risk profile the firm can no longer describe.
The call-level view also clarifies what to measure. Useful metrics are per agent-workflow, not firm-wide aggregates: calls per completed task, tokens per task, blocks per hundred calls, approvals per hundred consequential calls, and cost per task including retries. Aggregates hide exactly what governance needs to see — the workflow whose token use doubled after a prompt change, the agent whose block rate jumped after a model swap. Because every call carries identity and tier, these metrics are one query rather than a data project, and they connect directly to the review cycles: a metric that moves is either a finding for the agent's owner or a question for the model risk process, and the record shows which.
Frequently asked questions
Why does the AI gateway govern at the level of individual model calls?
Because calls are where identity, policy, and evidence can attach without ambiguity. A workflow decomposes into many calls across several code paths, and application-layer governance cannot reliably apply to all of them. Per call, the gateway resolves the agent's registry identity, checks routing policy against declared scope, counts the call against budget caps, and holds approval-tiered calls for named review — then forwards the request with injected provider credentials and logs the complete outcome. That granularity is what makes SR 11-7 model risk evidence and vendor oversight concrete rather than narrative.
How do model calls connect to vendor oversight?
Every call names its provider, and the gateway's records aggregate them: which vendor served which agent, for which workflow, at what volume and cost. Budget counters turn usage into spend records, so the firm's relationship with each AI service provider is measured rather than estimated — which is what vendor oversight under SR 26-2-style expectations in the investment management context assumes. Central credential injection also means vendors see one managed relationship, and switching or adding providers is a routing policy change rather than a code change across every agent codebase.
What happens to a model call that violates policy?
It is blocked before reaching the provider, and the block is logged with the agent identity, the policy that fired, and the attempted scope. The workflow receives a structured error rather than an opaque failure, and the block routes to the escalation path — the agent's owner for policy blocks, the review queue for confidence flags, security for anomalies. Blocked calls are not exceptional noise; they are the enforcement working, and their rate and pattern are monitoring inputs for SR 11-7 model risk records, showing where scopes and budgets actually bind.
How should firms handle agents that cannot route through the gateway?
Treat direct vendor connections, embedded API keys, and side channels to internal models as exceptions to document, not variants to tolerate. Where a legacy system genuinely cannot adopt the gateway yet, the exception should carry an owner, a time bound, and compensating monitoring — because an ungoverned call is a workflow whose risk profile the firm cannot describe. The goal is the invariant that every model call from a production agent passes through identity, policy, and evidence; the exception process exists to shrink the set that violates it.