The Next Value Stream

What Flow Can Learn from the Internet

You might remember a paper I published with John Rauser back a few years now, centering on the learnings and patterns we can leverage from the example of global, software-defined networks.

I keep returning to these ideas as work becomes more intertwined with artificial workers. Software factories gaining increasing attention, programmatic workflows, shifting left, the changing role of management and engineering, and increasing scale and consequence have all brought these ideas back to mind with increasing relevance. As relevant and timely as I think these comparisons are, they’re no less challenging. Back when we wrote the initial paper, a control plane was a more generic concept - now we have agent context, auto-research, Ralph Wiggum loops, MCPs, and token telemetry. It was uncomfortable to consider whether nodes in a network might be people or teams - now we have agents and subagents. The management plane was dominated by OKRs, strategy documents, and portfolio flow - now it’s policy-as-code, taste development, and engineering shifted left. This is just the beginning.

This week, I want to share a sneak peek of an article I’ve written for the Lean Enterprise Institute’s LeanTech Journal - diving deeper into some of these concepts, and what they mean for an agent-supported ecosystem. I hope you enjoy it, and I hope it sparks some conversation.

---

Most organizations are still trying to apply AI to existing workflows: faster coding, faster document generation, faster analysis, faster search. This risks paving the cow path — making the existing process faster without asking whether the process still makes sense.

The deeper AI opportunity is organizational redesign.

Lean thinking has always embraced the concept of value streams — the flow of material and information to go from a customer need, to delivered customer outcomes. Mapping the flow will always be necessary and not disappear. There is no future where value streams stop existing. Inputs still need to be refined into outputs. Outcomes still need to be observed. Feedback still needs to inform and improve the next cycle. AI can change how value streams are designed, routed, measured, and improved, especially in organizations that have narrowly considered value streams as static flows.

The best value streams already operate more like the internet, and the application of AI in organizations should be more like the internet flow of information.

From token ring to dynamic routing

Traditional organizations often behave like early network architectures.

In token ring networking, information moves sequentially through every node. You do not get to the destination unless you pass through each point in the chain. Many software and business processes still work this way: intake, analysis, design, approval, implementation, review, release. Everything moves at the speed of the slowest constraint. This waterfall logic also applies, unfortunately, to many current value streams.

The internet routes dynamically. Instead of sequential management it uses simple protocols, distributed capacity, error correction, telemetry, and loose coupling to move packets reliably across an enormous, constantly changing network. The path matters less than the ability to reach the destination efficiently and resiliently, although the flow still must be governed by clear protocols, constraints, escalation rules, and feedback. Complex systems (human or otherwise) need this.

A dynamic pattern is increasingly relevant to organizations, which are, essentially, a network of capabilities. In organizations pursuing digitization, teams are nodes. Platforms are shared infrastructure. Policies, objectives, service levels, and constraints form the management plane. Information, data, and work coordination form the control plane. Agents increasingly operate in the data plane, executing work, routing requests, surfacing exceptions, and helping teams adapt.

Shifting to a dynamic value-stream network isn’t just a matter of changing a workflow diagram; it requires upgrading the organizational operating system. To successfully facilitate and support dynamic routing, an enterprise must establish and support enabling capabilities that allow for adaptable workflow, such as:

1. Technical & Architectural Enablement

● Loosely Coupled Architecture: Business capabilities and software must be modular (e.g., microservices, bounded contexts). If your systems are tightly coupled, your teams will be, too, forcing a return to static, monolithic coordination.

● A Unified Semantic Data Layer: For work to route dynamically, systems must speak the same language. This requires integrating your toolchain (from ideation to production) so metadata (status, priority, ownership) updates automatically across the network via APIs or a data mesh.

● Value Stream Management (VSM) Platforms: You need real-time visualization of flow metrics (lead time, cycle time, work item age, flow efficiency). If you can’t see the bottlenecks dynamically, you can’t address them or route around them.

2. Governance & Funding Evolution

● Capacity-Based Funding: Shift from project-based funding (which creates temporary, fixed structures) to funding persistent value streams. This allows teams to pivot their focus without needing a new business case approved by a steering committee.

● Automated Guardrails (Policy-as-Code): Trust requires verification. To eliminate manual approval gates, embed security, compliance, and quality checks directly into the operational pipelines. If a work item passes the automated tests, it has the “authority” to route forward.

3. Organizational Design & Culture

● Stream-Aligned Teams (Team Topologies): Organize people around the flow of value, not functional specialties. Teams should possess all the cross-functional skills required to take an idea from concept to cash.

● Capability Focus: In a modern enterprise, two roles matter more than ever: Product roles that deeply understand customer problems, and engineering roles that understand problem solving. Domain knowledge remains critical, but with shared context and small batches of work, capacity becomes far more flexible.

● Explicit Work-in-Progress (WIP) Limits: Dynamic routing fails if every pathway is clogged. Enforcing WIP limits across the network forces the system to signal when it’s at capacity, triggering the need to reroute or reprioritize work.

● Psychological Safety and a Blameless Culture: Dynamic routing inherently involves experimentation and local decision-making. If failures are punished, individuals will default back to the safety of centralized, bureaucratic cover-your-back processes.

With these conditions in place, the flow pattern shifts from: “How do we optimize this fixed sequence?” toward “How do we design a system that can dynamically route from customer need to customer outcome based on capability, capacity, reliability, and context?”

Value streams still matter — but their design evolves

AI does not eliminate value streams but makes them far more important.

A value stream still connects a trigger to an outcome. But the path through the system cannot be singular, fixed, or linear, and, in true lean organizations such as Toyota, this has never been the case for the complete flow of all material and information.

Different customers, domains, products, and services may need different combinations of capabilities at different times. Some needs may be satisfied through self-service. Some may require specialized teams. Some may be handled by agents. Some may require humans in the loop. Network-based value streams offer:

● Multiple concurrent paths to outcomes

● Modular capabilities serving different customer needs

● Dynamic routing based on capacity and capability

● Teams that can be bypassed when unreliable

● Shared services consumed through supply-and-demand dynamics

● Humans moving increasingly toward oversight, policy, design, and exception handling

In stable environments, effective paths may still become paved. A goat path can become a trail. A trail can become a road, and demand prompts a highway. Stability is valuable when demand is known and the path is clear.

But the old highways are not automatically the right highways to modernize. In times of disruption, organizations may need to build new routes rather than repaint the lines on existing ones.

The bottleneck will continue to move

For years, organizations complained that software was too slow, too expensive, and too low quality. But many did not actually address the underlying constraints. They did not invest enough in platforms, automated testing, deployment automation, observability, or flow measurement. They did not understand their value streams or how to improve them.

Now AI is accelerating code generation, but that does not automatically improve the system or the value stream. In many cases, the constraint was not writing code. The constraint was merge queues, review, governance, unclear requirements, brittle architecture, poor test coverage, deployment friction, and weak feedback loops.

AI can make this worse if organizations double down on a non-constraint.

The same pattern will repeat outside software. Purchasing, HR, administration, compliance, and other knowledge-work domains have decades-old processes, undocumented tribal knowledge, and weak process visibility. Many have not successfully implemented lean administration and examined their value streams, let alone applied AI-enabled flow.

If the human is the bottleneck because the knowledge lives only in someone’s head, AI will not magically fix the system. The work must first be made visible, structured, codified, and measurable.

Before smart homes, fix the plumbing

Many organizations want a smart home, but they have not fixed the plumbing.

AI depends on information flow (i.e., plumbing). If the underlying data, context, and process knowledge are fragmented, outdated, or inaccessible, AI has little reliable material to work with. The old rule still applies: garbage in, garbage out. Legacy data is the critical bottleneck, in existing value streams or AI-enabled ones.

Before the current AI wave, data mesh — assigning data to specific domains — was one of the major enterprise topics. Then AI captured attention, and many organizations moved on, but the underlying need for decentralized data did not disappear. If anything, AI makes it more urgent and context more important. The quality, structure, ownership, and retrievability of information determine what AI systems can safely and usefully do.

Consider the risk of AI micro-optimizations atop legacy flows. A project involving multiple enterprise and legacy platforms was plugging in an AI solution over the top. The system had high usage but poor architecture: disconnected tools, weak retrieval, and difficulty getting useful answers from the available material. It was easy to use but ultimately delivered no value. We can’t simply sprinkle AI on top of existing systems and expect magic. To reach high leverage, we have to rebuild the information architecture so knowledge can actually be retrieved, used, and cultivated.

To transform with AI, many organizations face an information rebuild: plumbing, architecture, context, ownership, and flow.

The Cloud revisited

There is a recent precedent for this transition: cloud infrastructure.

Before infrastructure as code, many organizations depended on engineers who knew the manual steps. Those steps lived in heads, runbooks, tickets, and wiki pages. Over time, the work was codified into playbooks, automation, and tools such as Terraform, Ansible, and Kubernetes. That made infrastructure more repeatable, scalable, and transferable.

The same pattern is likely with agents: At first, everyone builds their own. Then duplication becomes obvious. Common skills, reusable agents, shared patterns, and agent-as-code approaches emerge. Eventually, teams converge around standard capabilities that can be understood, improved, governed, and reused.

The AI maturity curve is familiar to cloud:

1. Tribal knowledge

2. Documented steps

3. Repeatable playbooks

4. Automation

5. Shared platforms

6. Dynamic, observable, self-service capability

AI does not skip those stages. It compresses the timeline and raises the cost of ignoring them.

Questions for lean practitioners

Lean practitioners must get beyond the generic “Where can we add AI?” to questions specific to strategy, process, and measurements that clearly identify the highest potential for AI to deliver value:

Strategy Level

Focuses on long-term direction, value definition, macro-constraints, and overarching governance. It defines what success looks like and sets the boundaries for the entire system.

● Which customer outcomes matter most?

● Which value streams are most constrained?

● Where has the bottleneck moved?

● Which old assumptions about cost, capacity, or sequencing are no longer true?

● What management-plane policies, objectives, and service levels should guide the system?

● What lagging indicators define success?

Tactics Level

Bridges the gap between strategy and execution. It focuses on systemic design, cross-team coordination, infrastructure, tool integration, and structural optimization.

● Which capabilities should be modular, reusable, or self-service?

● Where do we need dynamic routing instead of fixed handoffs?

● Where are teams manually coordinating what software or agents could route?

● Where is information fragmented across tools?

● What middle-loop metrics show whether the system is improving?

● What questions should the data answer?

Operations Level

Deals with ground-level execution, immediate workflow visibility, task-specific automation, and real-time monitoring to keep the system flowing smoothly.

● Where is work still trapped in human memory?

● Which tasks are repeatable enough to codify?

● Which decisions require human judgment?

● Which agents need to be visible on the process map?

● What data is required for flow and continuous improvement?

● What leading indicators reveal problems early?

● What telemetry is needed to detect flow breakdowns?

● How does each measurement help people ask better questions?

A useful frame is outer-loop, middle-loop, and inner-loop measurement. The outer loop focuses on lagging outcomes. This is your strategic feedback loop of continuous improvement at organization scale. The inner loop focuses on leading signals closest to the work. This is your operational feedback loop of continuous improvement at team and individual scale. The middle loop connects local learning to broader system improvement. This is your tactical feedback loop of continuous improvement at cross-team or portfolio scale. By intentionally building these loops you create the “control plane” that sophisticated networks use to self-manage, adapt, and operate at peak performance and resilience.

The lean opportunity

AI will reward organizations that understand flow (i.e., lean organizations that already understand and optimize their value streams). It will penalize organizations that do not.

The companies that benefit most will not simply attach AI to existing processes, but use AI as a forcing function to make work visible, codify knowledge, clarify value streams, improve data quality, redesign constraints, and build dynamic systems that can sense, route, learn, and adapt. Beyond the basics, AI should prompt you to reevaluate not only the value stream, but the value-stream network. With the rise of software factories and engineering “shifting left,” we have new opportunities to imagine how work can flow across organizations, just like at the advent of electrification.

Lean has a powerful role to play here. But the conversation needs to move beyond “AI use cases” and toward AI-enabled operating models. Beyond task automation and toward value-stream redesign. Beyond individual productivity and toward networked capability.

The future organization may look less like a hierarchy of functions and more like the internet: simple rules, dynamic routing, observable flow, resilient nodes, modular capabilities, and continuous feedback.

The value stream is not going away. It is becoming more and more like a dynamic, adaptable, reliable network.