In our previous post, we argued that uncontrolled copilot platforms represent End User Compute (EUC) 2.0. That is the same class of operational risk that banks spent a decade remediating, such as the London Whale incident (JPMorgan’s 2012 trading losses attributed in part to spreadsheet-based risk models). We proposed a tiered framework: personal productivity and team workflows belong on managed copilot platforms with appropriate governance wrappers. Tier 3 and 4 workloads (business-critical and mission-critical) require something more.

This post addresses that “something more.” Managed copilot platforms are genuinely powerful, and for the majority of enterprise AI workloads they are the right answer. But for workloads where the consequences of failure are reportable, financial institutions need to own and operate the governance controls, not delegate them to a platform provider’s managed infrastructure. This distinction comes down to accountability infrastructure.

The governance gap is a design gap

Consider a hypothetical scenario where a Head of Financial Crime Operations builds a working alert triage prototype on a managed copilot platform. The productivity gain is real and immediate. The solution has been running for six months. It pulls data, enriches context, applies typology checks, and produces a recommended disposition. Analysts report meaningful throughput improvements.

The problem surfaces six months later, when the Money Laundering Reporting Officer (MLRO) receives a supervisory request: “Provide the evidence chain for the decision not to file a Suspicious Activity Report (SAR) on case reference 47291. Include the model version, data inputs, enrichment steps, and the basis for the analyst’s disposition.”

The solution can’t answer that question. This isn’t because it made the wrong decision. The decision might have been entirely correct. It’s because the evidence chain sits outside the platform’s design scope. The conversation log exists. The output exists. But the tamper-evident, reproducible chain of evidence that connects a specific model version, a specific prompt version, specific data inputs, and a specific human action into a single auditable record doesn’t. Regulators can ask for this evidence years later. Anti-Money Laundering (AML) record-keeping rules typically require at least five years after the customer relationship ends.

This isn’t an indictment of any platform. It’s an indictment of absent governance design. The same failure occurs when a bank or insurer builds a custom application without evidence infrastructure. What matters is whether they designed for accountability from the outset. Tier 3 isn’t monolithic. Some sub-tasks within a Tier 3 workflow carry lower consequence severity and might be adequately governed on a copilot platform, provided compensating controls are in place.

AWS Quick provides AWS CloudTrail attribution of user activity, Amazon CloudWatch vended logs exportable to the institution’s security information and event management (SIEM) solution, and audit trails for all API calls and configuration changes. What the institution still has to design for Tier 3+ workloads is the step-level record behind each decision: every model invocation and tool call, joined to the case and the human action, in tamper-evident form.

What governed actually means for regulated workloads

Regulators ask whether the AI is accurate, and whether you can demonstrate it is governed. Accuracy is a performance metric. Governance is an accountability infrastructure. For workloads at Tier 3 and above, governance means primarily five things. We focus on these five because they represent the most common governance gaps when a productivity copilot prototype is promoted into production without deliberate governance design. We note that data quality, classification, and consent governance are equally critical but are addressed by the institution’s existing data governance framework.

1. Model risk governance. Any AI component materially influencing a regulated decision might fall inside the institution’s model risk perimeter. These requirements exist across major jurisdictions (for example Prudential Regulatory Authority (PRA) SS1/23 in the UK, SR 26-2 in the US for non-generative AI models, and the Monetary Authority of Singapore (MAS) Guidelines on Artificial Intelligence Risk Management), though the definitional boundary varies. The regulated entity must make that determination for each agent, document the rationale, and apply proportionate controls. The institution must know which model version produced which output and must be able to restrict, validate, and retire models under formal change control.

2. Operational resilience. In the UK, firms must set impact tolerances for their important business services and remain within them in “severe but plausible” scenarios (PRA SS1/21, Financial Conduct Authority (FCA) SYSC 15A). In the EU, Digital Operations Resilience Act (DORA) (Regulation (EU) 2022/2554) sets information and communications technology (ICT) risk management and resilience testing requirements. The financial institution must own the service level objectives, the dependency mapping, the failover configuration, and the incident response. These obligations can be met through direct ownership or through appropriately governed outsourcing arrangements. The critical factor is whether the institution can demonstrate it understands its operational dependencies and maintains meaningful oversight.

3. Audit lineage. Basel Committee on Banking Supervision (BCBS) 239 set the precedent: risk data feeding regulatory reporting must carry end-to-end lineage from source to report. For AI-assisted decisions, this means capturing every model invocation, every data enrichment step, every prompt version, and every human action in a form that is tamper-evident, reproducible, and queryable for the applicable retention period. Conversation logs are necessary but not sufficient.

4. Compute boundary ownership. For workloads processing sensitive data under strict residency requirements, the institution must control its data perimeter: network egress rules, encryption key management, data classification boundaries, and verification that sensitive data doesn’t leave the controlled perimeter. This can be achieved by running the AI inside the institution’s own virtual private cloud (VPC) or through architectural controls such as AWS PrivateLink endpoints and VPC-restricted access policies. The regulatory requirement is demonstrable perimeter control, not a specific hosting model.

5. Change control. Every component in the decision chain (prompts, tool definitions, orchestration logic, and model selections) must be versioned under formal change control, with each production version mapping to its commit, its review record, and its approver. A prompt managed under systems development life cycle (SDLC) governance (version control, peer review, approval gates, deployment pipelines) is a governed change regardless of whether the runtime is a managed platform or custom infrastructure. The practical issue is whether the platform supports integration with existing change management toolchains and whether there is sufficient discipline to mandate its use.

Building the governed stack

Amazon Bedrock AgentCore provides three capabilities as a managed runtime for workloads where the institution must own the governance controls end-to-end. It isn’t the only possible architecture, but it removes the undifferentiated infrastructure work and provides a supported, auditable foundation appropriate for regulated financial services workloads at Tier 3 (business-critical decisions) and above.

The organization owns the compute boundary. AgentCore connects into the organization’s own VPC and runs under its own AWS Identity and Access Management (IAM) controls. The organization defines network egress rules, encryption key management, and data classification boundaries. Sensitive data never leaves the controlled perimeter. The organization can verify this independently. It’s architectural, not contractual.

Policy controls define agent boundaries. With Policy in Amazon Bedrock AgentCore, you can write rules in natural language as Policy and then convert them to Cedar (an open source authorization policy language). These rules control which tools an agent can access, what actions it can perform, and under what conditions. AgentCore then enforces these boundaries, and blocks and logs an agent that attempts to exceed its defined authority. For regulated workloads, this means the organization can implement the principle of least privilege at the agent level, not only the human level.

Tracing and observability provide the evidence chain. When you instrument your agent code, AgentCore captures spans for agent invocations, tool calls, and intermediate steps. Exported to write-once storage such as Amazon S3 Object Lock, those traces become tamper-evident evidence. Test instrumentation coverage is part of validation. The evidence chain is only as good as that coverage. Control functions can trace a specific decision back through the entire chain: which model version, which prompt, which data inputs, which tools were called, in what sequence, and what the agent produced at each step. This infrastructure can help answer regulatory questions five years after the event.

These capabilities provide the infrastructure for governance. The quality of governance still depends on the organization’s implementation maturity, its operational investment, and its willingness to mandate disciplined development practices. Infrastructure supports governance. It doesn’t guarantee it.

Building solutions to specific requirements

Building on AgentCore requires software engineering capability. A governed agent build typically takes on the order of 8–12 weeks, not the hours-to-days of a copilot prototype. The organization must maintain what it deploys. Ongoing operational investment isn’t optional. Formal engagement with Model Risk, Operational Resilience, and Internal Audit is required during the build, not as a sign-off after the fact.

This isn’t a disadvantage, however, if it’s also seen as part of the governance mechanism.

Building a governed agent requires engineering discipline, including spec-driven development, versioned artifacts, peer-reviewed changes, and a formal deployment pipeline. That discipline produces the documentation and evidence infrastructure that the second line of defense (risk management and compliance) will ask for. The governance pack is assembled during development as a by-product of disciplined engineering, not retrofitted onto a prototype six months later at a much higher cost.

The business that builds on AgentCore is making a deliberate choice. It accepts higher upfront cost and longer time to production. In exchange, it gets a system it can defend to regulators, audit to the transaction level, and operate under formal service level objectives for the life of the workload. The business that promotes a copilot prototype into a regulated workflow is making a different choice. It accepts faster time-to-value in exchange for a governance evidence gap that widens and compounds with every decision the agent makes.

Copilots and governed systems coexist

The right architecture for most regulated institutions isn’t “all copilot” or “all custom build.” It’s a hybrid where each component sits at the appropriate tier.

Consider an AML alert triage workflow at a large UK clearing bank. The enrichment agent that pulls data and maintains audit lineage runs on AgentCore because it feeds a regulated decision and must produce evidence packs. The risk scoring model runs on AgentCore because it falls inside the model risk perimeter under SS1/23. For some Tier 3 sub-tasks where compensating controls exist and the platform provides sufficient audit logging, a governed copilot deployment might suffice. The drafting agent that helps analysts write structured narratives sits at Tier 2. The management information dashboard reads from the governed data store in AgentCore but runs on a standard business intelligence platform.

The copilot delivers the productivity gains that justify the program’s business case. AgentCore provides the evidence infrastructure for regulated decisions. They coexist because they serve different purposes, and the governance framework defines the boundary between them. A single workflow typically contains advisory, assistive, and autonomous components. The architecture should reflect that heterogeneity, not force everything into a single tier.

Organizational engineering maturity

Not every business is ready to build on AgentCore today, and these three questions might help determine where an organization sits in the maturity model:

  1. Does it have an engineering team with AI and machine learning (ML) operations experience sufficient to build and maintain governed agents? AgentCore eliminates infrastructure management, but it doesn’t remove the need for engineering judgment about agent design, prompt engineering, tool integration, and production monitoring.
  2. Can it sustain that maintenance over the agent’s lifecycle? Model monitoring, evaluation updates, prompt management, dependency upgrades, and incident response are ongoing commitments, not one-time build costs. An agent deployed without a maintenance commitment degrades silently.
  3. Is the incremental governance control from AgentCore worth the delivery delay and ongoing cost versus a governed copilot deployment? For some Tier 2–3 boundary workloads, a copilot with a governance wrapper (named owner, sampling regime, defined tolerances, human-in-the-loop enforcement) can be the faster and more cost-effective path to production while the organization builds AgentCore capability in parallel.

The framework matches governance infrastructure to consequence severity, not complexity for its own sake. If the consequences of failure are reportable, the governance infrastructure must be defensible. If the organization cannot yet build that infrastructure, the workload shouldn’t yet be automated at that tier.

Act now, but act appropriately

For Chief Risk Officers: Commission the tiering assessment. Identify which AI workloads on managed copilot platforms have drifted into Tier 3 territory without the governance infrastructure to support them. These are your highest-priority remediation targets.

For Chief Technology Officers: Invest in the engineering capability to build governed agents before the demand arrives. The business that waits until a regulator asks “show me the evidence” before building evidence infrastructure is already too late. AgentCore capability is a strategic investment, not a project cost.

For MLROs and Heads of Financial Crime: Your alert triage prototype works. Don’t discard it. Use it as the requirements specification for a governed build on AgentCore that provides the evidence infrastructure to make it defensible when the National Crime Agency (NCA), Financial Crimes Enforcement Network (FinCEN), or MAS asks questions three years from now.

For Chief Executive Officers and Board Members: The choice between speed and governance is a false dichotomy at the portfolio level. Use managed copilot platforms such as AWS Quick broadly for Tier 1 and 2 workloads, and for Tier 3 tasks only where you can evidence the controls. Build governed systems on AgentCore for Tier 3 and 4 workloads, because the regulatory exposure is too significant to ignore.

The time to build that infrastructure is before you need to defend it.

Contact your AWS account manager to find out how we can help you build governed AI systems appropriate for your regulatory environment.

 


About the authors