Asset Mgmt Model Risk Build vs Buy | Moring
For AI model risk management in asset management, buying suits commodity controls such as inventory tooling and monitoring dashboards, while building suits firm-specific elements like validation standards and investment-model risk taxonomies. Most firms assemble a hybrid: purchased platforms for telemetry and workflow, internally owned policies and risk appetite aligned to SR 11-7 and EU AI Act expectations.
Short answer
For AI model risk management in asset management, buying suits commodity controls such as inventory tooling and monitoring dashboards, while building suits firm-specific elements like validation standards and investment-model risk taxonomies. Most firms assemble a hybrid: purchased platforms for telemetry and workflow, internally owned policies and risk appetite aligned to SR 11-7 and EU AI Act expectations.
How it works
The build versus buy question for AI model risk management divides cleanly along the line between generic control machinery and firm-specific judgment. Asset management firms need an inventory that discovers models across platforms, monitoring that tracks drift, and workflow that routes validation requests, and these are now commodity capabilities that vendors deliver maturely. What no vendor can supply is the firm's own risk taxonomy: which models feed portfolio decisions, how investment performance tolerance is expressed, and what override discipline applies to a research agent. Building those elements internally is not a choice but a definition of ownership.
Buying has clear advantages on telemetry-heavy layers. Model discovery across cloud platforms and SaaS tools, token and cost attribution, and access logging require integrations that vendors maintain continuously; a firm building these would spend engineering capacity reproducing a market that already exists. The NIST AI Risk Management Framework and the Cloud Security Alliance's model risk framework give buyers a neutral checklist for evaluating vendors against governance functions such as mapping, measuring, and managing risk, so procurement can be structured rather than impressionistic. Enterprise AI governance guides recommend exactly this: adopt platforms for runtime signals, keep policy internal.
Building retains control where the models themselves are proprietary. An asset manager whose quantitative edge lives in proprietary signal models cannot hand validation of those models to a platform tuned for generic machine learning; the validation standards, challenger model design, and assumption documentation must encode investment judgment. Supervisory expectations in the United States, from SR 11-7 through the more recent SR 26-2, hold the firm accountable for validation independence regardless of tooling, so purchased workflow around an internally owned validation methodology is the defensible pattern. European obligations under the EU AI Act similarly attach to the deployer, not the vendor.
A practical hybrid emerges in most asset management programs. The firm buys discovery, monitoring, and evidence storage, builds the risk taxonomy, validation standards, and committee processes, and integrates both through an inventory that serves as the shared backbone. Research on AI risk management as a function of AI governance argues for treating these as one holistic system rather than separate programs, which supports the hybrid: the platform records what runs, the firm decides what it means. Databricks guidance on responsible AI programs makes the same point from the engineering side, recommending that governance artifacts live with the systems that produce them.
Vendor models themselves raise the third dimension of the decision: whether to adopt third-party foundation models or build domain models. Robeco's asset management research notes that managers increasingly use external AI within investment processes, and supervisory thematic reviews of AI model risk highlight third-party AI as a distinct control area, requiring the same inventory entry, materiality assessment, and monitoring as internally built models. Buying model capability is legitimate; outsourcing accountability is not. Contract terms should secure the transparency needed to register and validate what was purchased.
The decision framework therefore reduces to accountability rather than technology preference. Buy wherever a vendor maintains an integration better than the firm can, build wherever regulatory accountability or investment judgment cannot transfer, and require that anything bought feeds evidence into the firm's own model inventory. Asset management firms that follow this split avoid the two classic failures: an internally built monitoring stack that decays with staff turnover, and a purchased platform whose certificates of compliance substitute for the firm's own validation conclusions.
Frequently asked questions
Which parts of AI model risk management should an asset manager buy rather than build?
Buy the telemetry and integration layers: model discovery across cloud platforms and SaaS tools, drift and cost monitoring, evidence storage, and validation workflow routing. Vendors maintain these integrations continuously, and a firm building them would spend scarce engineering capacity replicating a mature market. Evaluate candidates against the functions in the NIST AI Risk Management Framework and the Cloud Security Alliance model risk framework so procurement maps to governance needs. Keep internally owned everything that encodes firm-specific judgment: the risk taxonomy, validation standards, and override discipline that SR 11-7 and SR 26-2 hold the firm accountable for.
Why must asset management firms still build elements of model risk management in-house?
Because accountability does not transfer with software. United States supervisory expectations from SR 11-7 through SR 26-2 hold the firm responsible for validation independence, risk materiality assessment, and monitoring regardless of which platform records the evidence. Investment judgment is equally non-transferable: validation standards for proprietary signal models must encode the firm's own performance tolerance and assumption review, which no generic vendor configuration supplies. European obligations under the EU AI Act attach to the deployer as well. The defensible pattern is purchased workflow and telemetry around internally owned methodology and committee processes. For asset management this is structural: model risk accountability for validation conclusions cannot transfer to a platform vendor.
How should asset managers handle third-party AI models in the build versus buy decision?
Treat purchased model capability as an inventory entry with full governance obligations, not a governance exemption. Supervisory thematic reviews of AI model risk management identify third-party AI as a distinct control area, requiring the same materiality assessment, registration, and monitoring as internally built models. Robeco's research observes that asset managers increasingly use external AI inside investment processes, so contract terms should secure the transparency needed to validate what was bought: training data summaries, change notifications, and performance reporting. The firm can buy the model; it cannot buy out of accountability for how the model behaves in its portfolios.
What does a practical hybrid of built and bought model risk controls look like?
A working hybrid assigns the inventory as the shared backbone. The firm buys discovery, monitoring dashboards, and evidence storage, and integrates them so every model, whether proprietary, vendor-supplied, or embedded in a platform, registers automatically. Internally, the firm maintains the risk taxonomy, validation standards, committee processes, and risk appetite statements aligned to SR 11-7 expectations and EU AI Act transparency duties. Research on AI risk management as a function of governance supports treating the combination as one holistic system: the platform records what runs, while the firm decides what it means and owns every conclusion drawn from the record. Asset management firms operating this hybrid satisfy model risk examination questions from one evidence base rather than two registers.