Asset Mgmt AI Gateway Guide | Moring
This guide walks an asset management firm through adopting an AI gateway: inventory agent workflows, stand up the gateway with a registry of agent identities, define routing and approval policy, capture evidence from the first call, and roll out workflows in risk order. Model risk discipline under SR 11-7 and SR 26-2, plus EU AI Act obligations where they apply, shape each step.
Short answer
This guide walks an asset management firm through adopting an AI gateway: inventory agent workflows, stand up the gateway with a registry of agent identities, define routing and approval policy, capture evidence from the first call, and roll out workflows in risk order. Model risk discipline under SR 11-7 and SR 26-2, plus EU AI Act obligations where they apply, shape each step.
How it works
An AI gateway adoption at an asset management firm succeeds or fails on sequencing, so this guide presents the steps in order. The first step is a workflow inventory. List every place agents and LLM calls occur — research summarization, fund commentary drafting, meeting note processing, reconciliation support, client query drafting — and classify each by stakes: internal-only, operational, or client-facing and investment-relevant. The inventory becomes the registry's backbone and the risk tiering input. Firms that skip the inventory end up with shadow agents connecting directly to vendors, which the gateway cannot govern and the SR 11-7 documentation cycle cannot see.
Second, stand up the gateway and populate the registry. Each agent gets a stable identity, a named owner, and a declared scope: which models it may call, which tools it may invoke, which datasets it may read. Identity binding is the step most often deferred and most painfully missed, because policies, budgets, approvals, and logs all reference the identity. SR 26-2 expectations in the investment management context assume the firm can trace which system performed which action; an unregistered agent breaks that traceability at the root.
Third, define policy at the gateway in three families. Routing policy restricts which models each agent may call — a commentary agent may use the approved vendor, a research agent may not touch client-data-bound models. Budget policy caps tokens and spend per workflow so a runaway loop becomes a block, not an invoice. Approval policy marks the workflows requiring a named human sign-off before the gateway permits the consequential action, typically client-facing output and anything feeding portfolio decisions. These policies are the firm's model risk and vendor oversight controls expressed mechanically, which is exactly what SR 11-7 lifecycle documentation describes.
Fourth, enable evidence capture before the first production call. Every request, response, policy decision, and approval is recorded against the agent identity, in tamper-evident storage. Asset managers should resist the temptation to add capture later: SR 11-7 monitoring expectations and EU AI Act record-keeping duties for in-scope high-risk systems both assume continuous records, and the difference between day-one capture and retrofitted capture is the difference between evidence and reconstruction.
Fifth, roll out workflows in risk order. Begin with one internal-only workflow, verify that policy blocks behave as documented and that logs are complete, then add operational workflows such as reconciliation. Client-facing and investment-relevant agents — commentary drafting, portfolio research acting on MNPI — go last, with approval gates and kill-switch behavior tested in the lower tiers first. This mirrors model risk tiering under SR 11-7: proportionate scrutiny, deepest review where consequences are largest.
Finally, institutionalize review. Gateway policy, registry scope, and escalation ownership need a standing review cycle, ideally synchronized with the firm's existing model governance calendar so AI controls and SR 11-7 documentation evolve together. The EU AI Act, where it applies to the firm's systems, adds requirements for transparency and human oversight that the gateway operationalizes; treating policy review as a periodic control rather than a project keeps both regimes satisfied without a second bureaucracy.
Two pitfalls recur in adoption and are worth naming before they cost a quarter. The first is treating the gateway as an infrastructure project with governance attached afterward, which produces working enforcement and unusable oversight — policies nobody reviewed, tiers nobody justified, evidence nobody queried. The second is over-customizing early: bespoke policy logic per workflow before the standard families have run anywhere, which multiplies the review surface the governance body must maintain. The guide's sequence — inventory, registry, policy, evidence, staged rollout, standing review — exists precisely to avoid both: controls arrive reviewed, in risk order, with the evidence question settled before the first consequential workflow depends on the answer. A final calibration note belongs in the guide: the tier boundaries will be wrong the first time, and the design should expect it. Tiers assigned at inventory rest on assumptions about volume, stakes, and reviewer capacity that only production traffic confirms, so the guide should treat tier review as part of the standing cycle rather than a one-time classification. The practical marker is the gate statistics: a gate that never fires on a workflow believed high-stakes, or one that fires on nearly every call, is the tier telling the firm where the classification was wrong. Adjusting through the change process — not by quietly ignoring the gate — keeps the control environment honest while the calibration improves.
Frequently asked questions
What is the first step in adopting an AI gateway at an asset management firm?
A workflow inventory. List every agent and LLM use case — research summarization, commentary drafting, reconciliation support, client query drafting — and tier them by stakes. The inventory drives the gateway's registry, the risk-based rollout order, and the SR 11-7 model risk documentation. Without it, shadow agents connect directly to vendors and remain invisible to both the gateway and governance, which undermines the traceability that SR 26-2 expectations in the investment management context presume the firm can demonstrate.
Why does agent identity binding matter so much in the registry?
Because everything else references it. Routing policy, budget caps, approval requirements, call logs, and model risk records all attach to a stable agent identity with a named owner and declared scope. If identities are added after deployment, the firm must reconcile months of anonymous traffic into a registry retroactively, and unregistered agents break the ability to trace which system acted — a traceability expectation embedded in SR 26-2 and in EU AI Act record-keeping for in-scope systems. Binding identity at adoption is cheap; reconstructing it later is not.
Which gateway policies matter most for an asset manager?
Three families. Routing policy limits which models each agent may call, which protects vendor-approved usage and data boundaries such as MNPI-scoped research. Budget policy caps tokens and spend per workflow, converting runaway loops into blocks rather than invoice surprises. Approval policy requires a named human to sign off before consequential actions — typically client-facing output — are permitted. Together these express the firm's model risk controls mechanically, matching the lifecycle documentation SR 11-7 expects and giving reviewers evidence rather than assertions.
When should EU AI Act obligations enter the adoption plan?
From the inventory onward, but their weight grows at the evidence and rollout stages. Where the firm's systems fall within the EU AI Act's scope, obligations for transparency, human oversight, and record-keeping apply, and the gateway is where those become concrete: captured call evidence, approval gates, and reviewed policy changes. Asset managers with EU-facing clients or operations should map each in-scope system to its gateway controls during adoption rather than afterward, and align gateway policy review with the SR 11-7 documentation cycle so the two regimes stay synchronized.