Asset Mgmt AI-DLC Build Vs Buy | Moring
For asset managers adopting AI-DLC, the build-versus-buy question splits: buy the tooling and platform capabilities that are commoditising, and build the governance harness that encodes institutional rules, since no vendor can supply the firm's risk appetite, approval thresholds, or model risk alignment. The evidence from financial services implementations supports this split, with harnesses assembled in days on top of commercial agent platforms.
Short answer
For asset managers adopting AI-DLC, the build-versus-buy question splits: buy the tooling and platform capabilities that are commoditising, and build the governance harness that encodes institutional rules, since no vendor can supply the firm's risk appetite, approval thresholds, or model risk alignment. The evidence from financial services implementations supports this split, with harnesses assembled in days on top of commercial agent platforms.
How it works
The build-versus-buy question in AI-DLC is really two questions, and firms that conflate them make poor decisions on both. The first is which agent tooling to use for development work, where the market has converged: coding agents from the major platforms are commoditising quickly, and financial services guidance explicitly frames the governance harness as tool-agnostic, working regardless of whether teams use Kiro, Claude Code, or Codex. The second is where the firm's institutional rules live, and that is a build decision by nature.
The case for building the governance layer is definitional. The harness encodes what the firm considers acceptable: which repositories agents may touch, which data they may read, which approval thresholds apply to which change types, how generated code is validated before production, and how the whole record aligns with the firm's model risk management framework. AWS's financial services guidance describes exactly this layer as the differentiator, a governance harness configured in days that makes agents operate within institutional rules regardless of tooling, and engineering and GRC teams both gain from it. No vendor can supply another institution's risk appetite.
The case for buying the execution layer is equally definitional. Agent platforms, coding tools, and pipeline integrations are competitive markets with rapid improvement cycles, and an asset manager that builds its own coding agent is spending scarce engineering on a commodity while its governance layer, the part that is actually differentiating, waits. The firm's engineering talent is better spent on the harness components: the policy engine, the evidence pipeline, the review gates wired into CI and CD, and the reporting that connects the programme to risk governance.
A middle path deserves mention because most firms will land there: adopt a commercial platform that exposes governance extension points, and build the configuration that makes the extension points carry institutional meaning. The evaluation criteria follow: can the platform enforce the firm's access scopes at the tool boundary, emit evidence the audit function can consume, support human approval gates at defined checkpoints, and accommodate the firm's model risk documentation? Platforms that treat governance as an add-on rather than an interface will cost more to tame than they save.
For asset managers specifically, the supervisory dimension tips the scale toward owning the governance layer. Under SR 11-7 tradition, the firm answers for its controls; a vendor's assertions about safe AI development do not transfer accountability. The EU AI Act's expectations for high-risk systems similarly require the deployer to demonstrate risk management and oversight. Building the harness, even on bought components, is what keeps the firm in the answering position with evidence in hand.
The sequencing advice from financial services practice is encouraging: the harness was built in days working alongside a large institution, because the components, identity integration, policy configuration, evidence logging, mostly exist in firms with mature control environments. The build-versus-buy decision is therefore less about capacity than about clarity: buy the execution, build the rules, and let the firm's own risk discipline, not a vendor's feature list, define what governed AI development means.
Sources:
Frequently asked questions
Should an asset manager build or buy its AI-DLC governance harness?
Build the governance layer, buy the execution tooling. The harness encodes what only the firm can define: which repositories and data agents may touch, which approval thresholds apply to which change types, how generated code is validated, and how records align with the firm's model risk management framework. Supervisory accountability under the SR 11-7 tradition means a vendor's assertions do not transfer responsibility for those controls. The execution layer, coding agents and pipeline integrations, is a competitive commodity market where building wastes scarce engineering. Financial services implementations show the harness can be assembled in days on commercial platforms when the control components already exist.
Which AI-DLC components should an asset manager buy rather than build?
The execution layer: coding agents, development platforms, pipeline integrations, and model access, where the market is competitive and improving quickly. Coding agents from the major platforms are commoditising, and the governance harness in financial services practice is explicitly tool-agnostic, working across Kiro, Claude Code, Codex, and equivalents. Building an in-house coding agent spends engineering on a commodity while the differentiating work, the policy engine, evidence pipeline, review gates, and risk reporting, waits. The buying discipline is to select platforms that expose governance extension points, since enforcement access scopes, consumable evidence, and approval gate support determine whether the bought layer can be governed at all.
What should an asset manager evaluate in a commercial AI-DLC platform?
Whether governance is an interface or an afterthought. The platform should enforce the firm's access scopes at the tool and repository boundary, emit logs the audit function can consume with agent identity and context on every record, support human approval gates at the methodology's defined checkpoints, and accommodate the firm's model risk documentation rather than replacing it. Additional criteria: deployment options that keep code and data within the firm's boundary, contractual terms on data use in model calls, and a roadmap showing governance capabilities maturing. Platforms that treat these as premium add-ons usually cost more to tame than they save, and the evaluation should price that in.
How long does building an AI-DLC governance harness take for an asset manager?
Days to a working harness, when the firm's control components already exist. Financial services practice reports a harness built in three days working alongside a large institution, because the underlying pieces, identity integration, policy configuration, evidence logging, were largely present in the firm's existing environment. An asset manager with mature identity management, CI and CD pipelines, and a model risk documentation practice starts from the same position. The realistic timeline is a working harness within weeks including testing with the first agents, followed by refinement as quarterly reviews surface calibration needs. The constraint is usually clarity about institutional rules, not engineering capacity, which is why the rules work precedes the build.