The unit of software delivery just changed
Agentic AI shifts software delivery from applications to automated workflows. The business process and the code become the same thing — which means governance needs to shift too.
AWS Enterprise Strategist Mark Schwartz makes a compelling argument that agentic AI is reshaping software architecture at a fundamental level. The shift isn't incremental. It moves the basic unit of delivery from applications and microservices to automated workflows.
That distinction matters more than it sounds.
For years, software delivery organized around two poles. Monolithic applications bundled everything together, creating massive overhead for any change. Microservices broke things apart, but introduced their own complexity — coordination costs, alignment challenges, and governance headaches that scaled with every new service.
Workflows cut through both problems. A workflow represents an end-to-end business process with clear, bounded value. An agent automates that process. The workflow becomes the unit of specification, testing, validation, and investment. Business cases shift from "does building this entire application justify the cost" to "does automating this specific workflow create value." That's a fundamentally different conversation.
What this means in practice
Tools like Amazon Quick Suite already let end users create agentic workflows directly. This isn't about building complete ERP systems from scratch. It's about improving discrete business processes — the kind that employees currently execute manually, step by step, across multiple systems.
Security teams can threat model individual workflows instead of sprawling applications. Testing teams can validate bounded processes rather than regression-testing entire platforms. The scope of analysis shrinks to something manageable, which means it actually gets done.
The governance implications are significant
Here's what matters for legal and product teams: workflows are interchangeable with business processes themselves. The business process and the code that executes it become the same thing — just different representations of the same system.
So governance shifts from overseeing monolithic systems to governing discrete, testable units of work. For product counsel, that translates to a few concrete changes:
Risk assessment becomes granular. Instead of evaluating an entire AI-powered platform, you're assessing whether a specific automated workflow — say, contract intake triage or vendor risk scoring — meets your compliance requirements. The boundaries are clearer. The failure modes are more predictable.
Accountability maps to process owners. When the unit of delivery is a workflow, you can tie accountability to the business owner of that process. That's a significant improvement over the current state, where responsibility for AI system outputs gets diffused across engineering, product, and operations teams with no clear owner.
Testing and validation become tractable. A bounded workflow can be validated against defined inputs and expected outputs. You can document what the agent does, what decisions it makes, and where human oversight applies. That's the foundation for any defensible governance framework.
Investment decisions align with risk decisions. When the business case focuses on a specific workflow, the legal review can focus there too. You're not trying to anticipate every possible use case of a general-purpose system. You're evaluating a defined scope of automation.
Building on distributed accountability
This connects directly to everything I've written about multi-agent systems and distributed accountability. When agents coordinate across workflows, you need governance frameworks that track decision-making across those boundaries. Schwartz focuses on organizational structure — how teams align around workflows instead of applications. I've addressed the governance challenges these architectures create — how you maintain accountability when multiple agents interact, hand off tasks, and make decisions that compound.
Both angles matter when you're building agent-based products. The organizational shift Schwartz describes actually makes governance more achievable, because workflows create natural boundaries for oversight. But it only works if legal and product teams are involved in defining those boundaries from the start, not retrofitting governance after the workflows are already in production.
The unit of delivery changed. The unit of governance needs to change with it.
Source:Mark Schwartz, AWS Enterprise Strategy
Mark Schwartz
