Asset Management AI Agent Identity Build Vs Buy | Moring
For agent identity governance, most asset managers should buy the enforcement layer and build only the integration layer: the hard problems, such as ephemeral credentials, delegation chains, and behavioral verification, are being solved by dedicated platforms, while the firm's own policies, ownership structures, and model risk mapping must always be built in-house. Buying alone is insufficient; building alone is expensive and late.
Short answer
For agent identity governance, most asset managers should buy the enforcement layer and build only the integration layer: the hard problems, such as ephemeral credentials, delegation chains, and behavioral verification, are being solved by dedicated platforms, while the firm's own policies, ownership structures, and model risk mapping must always be built in-house. Buying alone is insufficient; building alone is expensive and late.
How it works
Build versus buy for agent identity governance should start from an honest inventory of what is genuinely hard. Identity research catalogs the problems: agents authenticate programmatically without sessions, use ephemeral credentials, spawn child agents dynamically, and create delegation chains across sub-agents and APIs that legacy identity models were never designed to track. Solving these from scratch means building credential issuance, machine-speed authorization, chain tracking, behavioral reconciliation, and lifecycle automation, then maintaining them as agent patterns evolve. Few asset managers should want that engineering burden when the same capabilities are the core product of several dedicated vendors.
What must be built in-house regardless is the policy and accountability layer. Workforce governance research is specific that each agent needs a verified identity, a defined role, task-specific permissions, and auditable records, and that access governance should reflect industry risk, organizational role, and task scope. No vendor can decide which of the firm's agents may read portfolio data, which require pre-approval for writes, or who owns each identity; those decisions encode the asset manager's own control philosophy. Buy the machinery that enforces decisions at runtime; build the decision content, the ownership registry, and the classification gradient yourself.
The integration requirement shapes the buying decision more than feature lists do. An asset manager's identity governance must connect to what already exists: entitlement systems for human staff, the model risk framework that governs systems influencing decisions, and the runtime gateway that evaluates agent tool calls. Checklist research emphasizes tying each agent action to an owner, an approved scope, enforceable limits, and verifiable evidence, and that evidence only flows if the identity layer shares it with audit, compliance, and the control plane. Candidates should be scored on integration quality: whether they can consume your owner registry, export evidence your audit teams can use, and enforce decisions inside your existing runtime path.
Evaluate build candidates candidly if the firm considers it. The argument for building is control and fit: an internal team can encode asset-management-specific rules, such as research wall treatment or client communication boundaries, without negotiating with a vendor roadmap. The argument against is pace and completeness: identity research describes this as one of the most consequential security challenges of the current period, with the gap between legacy identity systems and agent requirements widening; a homegrown system built this year may lack delegation-chain tracking or behavioral verification that mature platforms already ship, and the firm owns its maintenance forever. A realistic hybrid, extending existing identity administration with purchased agent-specific capabilities, captures most of the fit with far less risk.
Regulatory alignment belongs to the firm either way. Model risk disciplines under SR 11-7 and continued under SR 26-2 require documented validation, independent review, and clear accountability, and whichever route the firm takes, the identity governance must produce evidence in that shape. Where European operations or vendors bring the EU AI Act into scope, deployer obligations attach to the firm, not the vendor, so contract terms should guarantee evidence export, availability commitments, and exit data portability. Treat the identity platform as a critical third party and review it with the same rigor as a data provider or custodian.
The practical recommendation for most asset managers: buy the agent identity platform, insist on integration with existing identity administration rather than a parallel island, and invest internal effort in the layers only the firm can supply, ownership assignments, permission mandates, classification, and the mapping into model risk evidence. Pilot the combination on one consequential agent before committing firm-wide. The build-versus-buy question, answered this way, is less about ideology than about where each party's comparative advantage genuinely lies: vendors at runtime enforcement, the firm at accountability.
Frequently asked questions
Which parts of agent identity governance should an asset manager build rather than buy?
The policy and accountability layers. Workforce governance research requires each agent to carry a verified identity, a defined role, task-specific permissions, and auditable records calibrated to industry risk and task scope, and no vendor can decide which agents may read portfolio data or who owns each identity. Ownership assignments, permission mandates, risk classification, and the mapping into the firm's model risk evidence must be built in-house. Vendors supply the enforcement machinery; the firm must always supply the decisions it enforces.
What makes agent identity governance risky to build entirely in-house?
The problem set is deep and moving. Identity framework research describes agents authenticating programmatically, using ephemeral credentials, spawning child agents, and forming delegation chains across sub-agents and APIs that legacy identity models cannot track, and calls the gap one of the most consequential security challenges of the period. Replicating credential issuance, machine-speed authorization, chain tracking, and behavioral verification internally consumes scarce engineering capacity, and the firm owns maintenance indefinitely as agent patterns keep evolving, while mature platforms ship those capabilities as their core product.
How should an asset manager evaluate identity governance vendors?
Lead with integration, not feature counts. Score candidates on whether they can consume the firm's owner registry, enforce decisions inside the existing runtime path for tool calls, and export evidence that internal audit and compliance can use directly, since checklist research ties every agent action to an owner, an approved scope, enforceable limits, and verifiable evidence. Then test enforcement live on one consequential agent, checking a blocked action, a staged approval, and the resulting audit record, and confirm contract terms cover availability, evidence export, and exit portability.
Does buying an identity platform change the firm's regulatory obligations?
No, and that is the point to internalize. Model risk expectations under SR 11-7 and continued in SR 26-2 attach to the firm's own systems and require documented validation, independent review, and clear accountability regardless of whose software enforces controls. Where EU AI Act duties arise through European operations or vendors, deployer obligations likewise rest with the firm. Vendor contracts should therefore guarantee the evidence flows the firm needs, and the identity platform should be reviewed as a critical third party with the same rigor applied to data providers or custodians.