Banking AI-DLC How It Works | Moring
AI-DLC works in Banking by running a structured lifecycle in which AI agents execute the work and humans approve at defined gates. Stages move through Inception, Construction and Operations, with a governance harness enforcing the bank's rules: scoped agent identities, verified approvals, tested changes and audit records. The result is AI-driven development that stays inside SR 26-2, SR 11-7, FFIEC, EU AI Act and DORA expectations.
Short answer
AI-DLC works in Banking by running a structured lifecycle in which AI agents execute the work and humans approve at defined gates. Stages move through Inception, Construction and Operations, with a governance harness enforcing the bank's rules: scoped agent identities, verified approvals, tested changes and audit records. The result is AI-driven development that stays inside SR 26-2, SR 11-7, FFIEC, EU AI Act and DORA expectations.
How it works
AI-DLC, the AI-Driven Development Lifecycle introduced by AWS, is a structured methodology in which AI is the primary executor across every phase from planning through operations, while humans provide strategic direction, approval and oversight. It is explicitly not a tool: it is a way of organizing how teams plan, build, test, deploy and operate software when agents do most of the execution.
The lifecycle runs in three phases. Inception covers intent capture, reverse engineering of existing systems and delivery planning. Construction covers functional design and building units of work through to the CI pipeline. Operations covers deployment pipelines and feedback loops. Some stages run inline so a human can answer questions and grant approval, while others delegate to subagents that run autonomously and return structured summaries. Each stage has defined inputs and an explicit approval-and-handoff step.
Around the lifecycle sits a governance harness, the engineering and process layer that makes AI-generated code safe and reliable before it reaches production. AWS describes building such a harness with a large financial institution in days, and notes that it is agent-agnostic: it works whether teams use Kiro, Claude Code or OpenAI Codex. The harness verifies approvals, applies the institution's policies, and only then allows the change to proceed.
Harness engineering is the safety axis of the methodology. Ontology engineering defines what correct output looks like; the harness verifies and enforces those constraints architecturally, which is the difference between an agent knowing the rule and an agent being unable to break the rule. AWS's reference material documents the failure mode this prevents: a fintech agent ran hundreds of retries in one loop, sent half-finished messages to customers and caused an outage, because there was no retry budget, timeout, output gate or circuit breaker in the architecture.
Extensions let a bank layer its own rules on top of the core workflow. Security baselines, property-based testing and internal compliance checklists attach as enforceable extensions rather than edits to the workflow itself, so each institution keeps its own model risk policies while sharing the same lifecycle. This matters in banking, where the EU AI Act can classify financial AI systems as high-risk and SR 26-2, SR 11-7 and FFIEC oversight expects controls to be demonstrable, not aspirational.
The end state for a bank is a lifecycle where every stage leaves evidence: who approved the intent, which scoped agent identity built the change, what tests ran, who signed off, and how deployment was monitored. DORA-style operational resilience expectations are easier to evidence when the record is produced by the process itself. Teams reviewing the mechanics can start from the published AI-DLC workflow architecture and adapt the harness rules to their own control environment.
Frequently asked questions
What are the phases of AI-DLC in a banking context?
AI-DLC runs through three phases: Inception, which captures intent, reverse engineers existing systems and plans delivery; Construction, which designs and builds units of work through to the CI pipeline; and Operations, which handles deployment pipelines and feedback. In a bank, each phase includes human approval gates, so agent-executed work moves forward only after review, and every stage leaves an auditable record suitable for SR 11-7 and FFIEC oversight.
What is the AI-DLC governance harness?
The governance harness is the engineering and process layer that makes AI-generated code safe and reliable before production. AWS describes building one with a large financial institution in days, and it is agent-agnostic across Kiro, Claude Code and OpenAI Codex. It verifies approvals, enforces the bank's policies, and blocks changes that lack authorization. For banking teams, the harness is where model risk expectations become enforceable checks rather than guidance documents.
How does harness engineering prevent agent failures?
Harness engineering enforces constraints architecturally instead of trusting the agent to follow rules. AWS's reference material documents a fintech agent that ran hundreds of API retries in one loop, sent half-finished messages and caused an outage, because the architecture had no retry budget, timeout, output gate or circuit breaker. Adding those controls to the harness means the same failure cannot recur in a banking deployment, whatever the underlying model does.
Can banks customize AI-DLC with their own rules?
Yes. The AI-DLC Extensions Framework lets an institution layer enforceable rules, such as security baselines, property-based testing and internal compliance checklists, on top of the core adaptive workflow. Rules load through progressive disclosure, so the workflow stays lean. This is how a bank encodes its own model risk policies and SR 26-2 expectations into the lifecycle, keeping one governed process across teams rather than a separate workflow for each AI initiative.