Asset Mgmt AI Gateway Permissions | Moring
Permissions at an asset management AI gateway express who — which agent — may call which model and tool, under which conditions. Scopes live in the agent registry, are enforced at the gateway on every call, and change only through reviewed process. Fine-grained permissions prevent scope creep, support least privilege for agents handling MNPI and client data, and produce the traceability SR 26-2-style expectations presume.
Short answer
Permissions at an asset management AI gateway express who — which agent — may call which model and tool, under which conditions. Scopes live in the agent registry, are enforced at the gateway on every call, and change only through reviewed process. Fine-grained permissions prevent scope creep, support least privilege for agents handling MNPI and client data, and produce the traceability SR 26-2-style expectations presume.
How it works
Permissions are the quiet core of AI gateway governance at an asset management firm. Every other control — budgets, approval gates, evidence — operates within the boundaries that permissions set: which agents may call which models and tools, under which conditions. Getting permissions right is mostly a matter of where they live and how they change. They live in the agent registry, attached to identity with an owner and a declared scope; they are enforced at the gateway on every call; and they change only through reviewed process rather than ad-hoc grants.
Scopes should be as narrow as the workflow tolerates, and narrower than intuition suggests. A research-summarization agent needs the research model and the document retrieval tool; it does not need the client-data-restricted model, the market data write path, or the vendor reserved for client-facing work. Least privilege for agents follows the same logic the firm applies to people: access is a risk position, not a convenience. The reason it matters more for agents is that agents do not get tired or confused — they will use every permission they hold, exactly as often as the workflow prompts, so a wide scope is not latent risk but active usage.
The materiality of the data sharpens the stakes. Agents at an asset manager routinely touch material non-public information — research, earnings data, portfolio positions — and client data with its own handling obligations. Permissions are how information barriers extend to agents: a scope that excludes client-data-restricted models from research workflows is the enforcement of the barrier, not a policy statement about it. SR 26-2-style oversight expectations in the investment management context presume the firm can show which system accessed which data; scoped, identity-bound calls produce that record mechanically.
Change process is where permissions either stay governed or decay. Scope expansions arrive as legitimate needs — a new workflow, a new dataset — and each one is a control change: proposed by the agent's owner, justified against the workflow, reviewed, applied at the gateway, and logged. Periodic recertification catches the decay that no process catches: agents retired without deprovisioning, scopes widened for a project that ended months ago, owners who left. Recertification on the same cycle as gateway policy review keeps the permission estate aligned with the model governance calendar.
Permission failure has a specific signature the gateway surfaces: calls blocked by scope. Block patterns are diagnostic — repeated blocked attempts by one agent at a scope boundary may be a mis-scoped workflow needing correction, or a malfunction, or misuse. Reviewing blocks periodically is a permission health check that most firms skip; the blocked call log is where scope creep shows up before it becomes accepted practice. Where the EU AI Act applies to a firm's systems, the permission estate also feeds transparency duties: what each system can do is defined by what it is permitted to do, and the registry states that precisely.
The discipline to maintain is singularity of enforcement. Permissions that exist in a document but not at the gateway are not controls; permissions enforced by convention rather than identity are not permissions at all. When every call is checked against a registry scope, and every scope change is reviewed and logged, the firm gets what oversight frameworks assume: a permission state that is knowable, attributable, and current.
A worked example makes the permission model concrete. A research-summarization agent: identity bound to the research team's credential, scope limited to the approved research model and the document retrieval tool, tier internal — it may read broadly within licensed research but has no client-data or client-facing model access. A commentary agent: identity bound to the client reporting function, scope limited to the commentary-approved vendor and no research ingestion tool, tier high-stakes — every draft release passes a gate. The two agents share one gateway and one policy engine, and the difference between them is entirely in the registry entries. That is the point: the permission model is the governance, and everything else — blocks, gates, evidence — is its enforcement.
Frequently asked questions
How should permissions be structured for agents at an asset manager?
In the registry, attached to each agent's identity: a declared scope of permitted models, tools, and datasets, with a named owner. Enforcement happens at the gateway on every call, and changes go through reviewed process — proposed, justified, applied, logged. Scopes should be as narrow as the workflow tolerates, because agents use every permission they hold exactly as often as prompted; a wide scope is active usage, not latent risk. This structure extends information barriers to agents handling MNPI and client data and produces the access traceability SR 26-2-style expectations in the investment management context presume.
Why is least privilege more urgent for agents than for people?
Because agents execute mechanically. A human with excessive access may rarely exercise it; an agent with a wide scope will use everything the workflow calls for, every time, at machine speed. Permissions are therefore the primary risk position, and the gateway's per-call enforcement makes them real rather than documented. Scope boundaries also carry data obligations — client-data-restricted models excluded from research workflows are the enforcement of information barriers, and identity-scoped call records show which system accessed which data.
How do permission changes stay governed over time?
Through the same discipline as other control changes: each scope expansion is proposed by the agent's owner, justified against the workflow, reviewed, applied at the gateway, and logged. Periodic recertification catches decay that process misses — retired agents not deprovisioned, project-scoped grants lingering, owners who departed. Reviewing the blocked-call log is a permission health check, since repeated blocks at scope boundaries reveal mis-scoping, malfunction, or misuse. Aligning recertification with the model governance calendar keeps the permission estate aligned with SR 11-7 documentation rather than drifting from it.
How do permissions support regulatory expectations?
They define what each system can do, which is what oversight frameworks evaluate. The registry's precise scopes give SR 11-7 model risk documentation a concrete system boundary, and the identity-scoped call records give the traceability SR 26-2-style expectations in the investment management context assume. Where the EU AI Act applies, permission definitions feed transparency duties — a system's capabilities are its permitted actions — and the gateway's enforcement record demonstrates that permissions bind in practice. A permission that exists only in a document, enforced nowhere, is not a control.