Do You Still Need Scaling Frameworks With AI?

Organizations didn’t adopt heavy agile scaling frameworks because they loved complexity. They did so to solve coordination problems caused by massive dependency networks, handoff-heavy pipelines, and specialized silos. But in an AI-native development environment, the nature of coordination shifts. While internal team coordination (the overhead of breaking down, tracking, and handing off stories) is rapidly descaled by AI, real systemic constraints like legacy architecture, security governance, compliance gates, and acquisition sprawl remain sticky. The work of leadership is no longer to scale, but to continuously diagnose the difference and dismantle apparatus wherever the need for coordination has vanished.

This shift doesn’t make Lean-Agile principles obsolete; if anything, it makes them more critical. When story-level tasking is automated, the only remaining flow bottlenecks exist at the feature and portfolio level. Kanban boards shift from tracking dependencies to managing strategy, topology, and outcomes. Without the discipline of managing batch size, limiting work-in-progress, and enforcing double-loop learning, organizations risk falling into the AI feature factory: generating code and features faster than the business can comprehend or absorb. Continuous descaling is the new practice: asking how much coordination is still necessary, and organizing for the actual constraints you have rather than the coordination habits of the past.

Zooming out from “Your Scrum Team Didn’t Get Obsolete”: what happens to the organization when team-level complexity collapses.

If you lead engineering or transformation in a product organization right now, you’re probably living in a contradiction. AI is making your teams faster than your operating model knows how to handle, and some part of you suspects half your coordination machinery is now ceremony. But you came up through agile. You know what happens when teams trade discipline for “move fast.” You’re not about to throw out flow, alignment, and evidence-based steering to chase a demo.

Here’s the resolution I’d offer: you don’t have to choose. The mechanics of your scaling framework are largely becoming optional. The principles underneath them matter more than ever. The work, and the advantage, is in telling the two apart.

SAFe, LeSS, Nexus, Spotify: none of them got complex because anyone enjoyed complexity. They got complex because the organizations were. Dense dependency networks. Handoff-heavy pipelines. Specialists siloed across teams who couldn’t share context fast enough to move without coordination. The apparatus was a response to reality. So the real question about AI-native development isn’t “does SAFe still apply.” It’s: what happens to the complexity that made the apparatus necessary?

AI doesn’t just accelerate delivery. It descales it in a pattern. A person backed by agents can take on what used to need a team. A team can take on what used to need an ART. An ART can take on what used to need a Solution Train.

Two mechanisms drive this descaling fractal: smaller stream-aligned teams, because one human now spans more disciplines and thins the dependency web; and spec-driven development, where the story-and-task overhead that needed a board and a ceremony becomes a conversation between a person and their agents, and one you don’t need to manage with the same coordination overhead you used to when each of these stories were often owned by a different person, and sometimes even distribued across teams.

Here is where the shift breaks down, and naming this honestly is the whole point. Descaling works brilliantly on self-inflicted coordination overhead: the handoffs, meetings, and story decomposition we invented to coordinate large groups of people. It stalls completely on real systemic and architectural constraints: legacy architecture, acquisition sprawl, regulatory load, integration history, security and data-governance gates, and live customers. AI compresses the team-internal dependency web fastest. The external web is stickier. Conway’s law didn’t get repealed by a model release.

That distinction is the core diagnosis: real architectural and organizational constraints are densest in exactly the mid-market organizations that need help most. Which means step one is never “just descale.” It’s diagnosing self-inflicted coordination from real systemic constraints, because that’s the difference between an operating-model breakthrough and an expensive demo. If you can’t name the real problem a meeting or structure solves, dismantle it. If the constraint is real, the apparatus stays, concentrated exactly where the irreducible coordination lives.

And one frame I’ll be blunt about: descaling is the same people taking on ambition that was out of reach: bigger strategic surface area, fewer coordination tolls. Descaling capacity that gets quietly banked as headcount cuts is just the feature factory with a smaller payroll. If your only move is cost takeout, you’ve missed the entire opportunity.

Scrum was never about story points. It was about alignment, coordination, and evidence-based steering under uncertainty, and the faster the loop, the more the steering matters. A team running five specs through a sprint needs that rhythm more, not less.

Kanban was never about moving cards to Done. It was about visualizing flow, surfacing bottlenecks, and managing batch size. I’ve argued for fifteen years that the highest-leverage Kanban surface is the feature and initiative level, not the story level. AI-native development finally makes that obvious: when story work collapses into workbench work, the only interesting flow questions left live at the spec and portfolio level.

ART-level Kanban starts to look like portfolio Kanban: potential initiatives, not dependency boards. The ART’s job shifts from managing story dependencies to managing outcomes, strategy, and topology, who’s working on what, are we shaped right, where do we redirect.

AI makes it dangerously easy to go fast in a straight line: vibe to a demo, ship, repeat. That’s the old feature factory with a faster engine: lots of motion, shallow learning, compounding debt. And here’s the part the speed hides: generation got cheap; comprehension didn’t. Understanding, reviewing, and trusting what was generated is now the bottleneck, and AI makes errors subtler and more confident. So you need more deliberate judgment per unit of output, not less.

That’s exactly why the human-in-the-loop matters more, not less. The antidote is the discipline we’ve always had: double-loop learning, compounding what you learn from one pass through the lifecycle into the next.

This is where “compound engineering” (the term Every and others put on the map) gets interesting. Their version is the system getting smarter each task. I’d add the part they undersell: compounding doesn’t happen automatically. It happens when someone deliberately applies decentralized-control-with-alignment, intrinsic motivation, and batch-size-and-feedback discipline to choose which loops are worth running. That’s not a tooling property. It’s a leadership one.

I’ve said for over a decade: continuously simplify, and use the leanest version of any framework that still serves a real coordination need. AI doesn’t change that argument. It accelerates it, and turns steady simplification into a window for disruptive, structural descaling on every front at once.

The question stops being “which parts of SAFe do we keep.” It becomes: given what AI now enables, what coordination problems do we actually still have, and what’s the simplest structure that solves them?

That’s the journey. Not a one-time transformation. A standing practice of asking: can one person do what the team did? Can the team take on what the ART did? Are we organized for the complexity we have, or the complexity we used to have?

If you’re the leader caught between “AI is outrunning my operating model” and “I refuse to trade flow for chaos,” that tension is the right place to be standing. It’s most of what I work on right now. I’m always up for comparing notes with people serious about disrupting how they work without throwing out the principles that made it work in the first place.