Book a demo
Topics · AI-DLC

Asset Mgmt AI-DLC Audit | Moring

Auditing AI-DLC in an asset manager means testing whether agent-driven development actually operated inside institutional controls, using evidence from the harness rather than assertions from documents. The audit scope covers agent inventories, access enforcement, code provenance, review gates, and lifecycle records. The evidentiary standard comes from the firm's model risk tradition: controls must be shown operating, with design and operating effectiveness tested separately.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

Auditing AI-DLC in an asset manager means testing whether agent-driven development actually operated inside institutional controls, using evidence from the harness rather than assertions from documents. The audit scope covers agent inventories, access enforcement, code provenance, review gates, and lifecycle records. The evidentiary standard comes from the firm's model risk tradition: controls must be shown operating, with design and operating effectiveness tested separately.

How it works

AI-DLC changes what an audit at an asset manager must examine. When agents plan, implement, test, and operate software, the audit object expands from the code that shipped to the process that produced it: which agents participated, under what scopes, with which human approvals, and what evidence ties each production change to a governed path. The methodology itself, coined by AWS and now adopted industry-wide, positions AI as the primary executor with humans governing and validating, so the audit's centre of gravity is the governance harness around the agents.

The audit's first test is inventory. A complete AI inventory, reconciled against procurement, code, identity, and endpoint records, is what exposes shadow agents and embedded vendor features, and audit guidance treats this reconciliation as the foundation of everything else. For an asset manager, the inventory should cover coding agents, pipeline agents, and any vendor tool with embedded agentic behaviour, because an agent nobody registered is an agent nobody scoped, and its output may already be in production systems the firm reports on.

Access enforcement is the second test, and the distinction auditors draw matters: design effectiveness and operating effectiveness are separate tests, because a well-written policy can still fail in production. The auditor should verify that agent access to repositories, pipelines, and data is technically enforced at task scope, that credentials are unique and traceable, and that the logs attribute every access event to a specific agent identity. The standard from AI governance audit practice is explicit: evidence that controls were technically enforced and that the log is the record of operation, not a post-hoc narrative.

Provenance and review gates form the third test. For each sampled production change, the audit traces the path: which agent generated the code, which human approved the plan, which checks ran before merge, and which approvals authorised deployment. AI-DLC's structure supports this, since the methodology builds human oversight at defined checkpoints, but the audit verifies the checkpoints actually fired. The review record should show meaningful human judgement, not rubber-stamped approvals, because regulators and courts increasingly examine whether oversight was real.

Lifecycle and monitoring records form the fourth test. Agents and the models they use drift, and the audit should confirm the firm monitors AI-generated code quality after deployment, retests on defined triggers, and maintains the version history linking each system state to its production process. IBM's breach research sharpens the stakes: 13 percent of organisations reported breaches of AI models or applications, and 97 percent of those lacked proper AI access controls, with 63 percent of breached organisations having no AI governance policy or one still in development. The audit exists to find those gaps before an incident does.

Findings then close the loop with the firm's existing model risk discipline: severity ratings, accountable owners, due dates, and independent validation before closure, as the audit checklist guidance prescribes. Asset managers have run this cadence for model validation for years under SR 11-7; extending it to the AI-DLC harness is an extension of practice, not a new discipline, and firms that frame it that way will find their audit committees and supervisors already understand the shape of the conversation.

Sources:

Frequently asked questions

What should an audit of AI-DLC at an asset manager cover?

Five test areas, each with evidence. Inventory first: a complete agent inventory reconciled against procurement, code, identity, and endpoint records, exposing shadow agents and embedded vendor features. Access enforcement second, verified technically at task scope with logs attributing every event to a specific agent identity. Provenance and review gates third: sampled production changes traced from generating agent through human plan approval, automated checks, and deployment authorisation. Lifecycle monitoring fourth: post-deployment quality tracking, retesting triggers, and version history linking system states to production processes. Findings management fifth, with severity, owners, dates, and independent validation. Design and operating effectiveness are tested separately throughout.

How does AI-DLC auditing connect to an asset manager's model risk programme?

As an extension of familiar practice. The firm's model risk discipline under SR 11-7 already requires documented governance, effective challenge, independent validation, and evidence that controls operate. AI-DLC auditing applies the same structure to a new object: the agents and harnesses that produce the firm's software. The vocabulary transfers cleanly, inventory as model catalogue, access scoping as use limitations, review gates as effective challenge, post-deployment monitoring as ongoing performance tracking. Firms that frame the audit this way find their governance committees and supervisors already understand the shape, and the AI-DLC harness becomes one more managed asset within an established risk framework rather than an ungoverned novelty.

What evidence do auditors want for AI-DLC governance at asset managers?

Operation, not documentation. The decisive evidence is that access controls were technically enforced, that every data and code access event was attributed to a responsible individual and a specific agent identity, and that the audit log is the record of actual operation rather than a post-hoc narrative assembled for review. Beyond access: sampled change provenance showing the human approvals that authorised each production path, inventory reconciliation results demonstrating coverage, and the quarterly reconciliation showing grants removed as usage lapsed. Audit practice is explicit that the gap between what organisations believe their documentation covers and what auditors accept is an evidence gap, and AI-DLC harnesses should be built to produce the evidence natively.

How should AI-DLC audit findings be managed in an asset management firm?

Through the existing model risk remediation cadence. Each finding receives a severity rating, an accountable owner, a due date, and independent validation before closure, per established audit practice, and findings feed the next harness configuration rather than sitting in a register. Common early findings include shadow agents discovered by inventory reconciliation, access scopes wider than observed task needs, and review gates that fired without meaningful human engagement. The finding pattern matters less than the closure discipline: asset managers are accustomed to tracked remediation with validation, and applying that rigour to the AI-DLC harness is what turns the first audit from a reckoning into a baseline.

Sources

  1. kiteworks.com
  2. aws.amazon.com
  3. adaptivesecurity.com
  4. ranthebuilder.cloud
  5. wiz.io
  6. agility-at-scale.com
  7. - federalreserve.gov
  8. - federalreserve.gov
  9. - artificialintelligenceact.eu