So What is an AI-Native ART?

Two weeks ago, I wrote about session two of Scaled Agile's AI-Native SAFe series and the shift from outputs to outcomes. Session three moved to the people. Rebecca Davis, framework methodologist, took us through what's changing for ARTs, teams and roles.

Rebecca opened with the big question: Do we still need ARTs in this era? Her answer was, of course, yes. Her rationale was that in the AI era, we still have system-level problems that no single team can solve alone, and so there's still a need for cross-team collaboration. This makes a lot of sense to me. I still remember the early versions of SAFe, where an ART made sense anywhere a large group of (software) people needed to work together to solve a problem. I think that still holds true today, with or without software. (This is the same logic we applied when introducing the Agile Release Tram.) 

Pump Up the Volume (Any 1990 Christian Slater fans out there?!)

Six characteristics define an AI-Native ART: 

1. Organised around products, 
2. Outcome-driven, 
3. Powered by human expertise and judgment, 
4. Balancing rapid innovation with cadence-based learning, 
5. Optimising shared AI workflows and platforms, and
6. Ensuring AI governance and ethics.

For the most part, these characteristics are not foreign to SAFe. What has changed is the amplification, and that's a positive move.

Cadence is the clearest example. "AI makes ART cadence more important than ever," Rebecca said, because teams working faster accumulate local decisions and dependencies faster, and new learning pulls the product in different directions. The importance of cadence definitely isn't new. Scaled Agile just pumped up the volume. ;-)

Organising around products is the one that caught my attention first. I've spent almost a decade teaching "organise around value". So I had a look at the AI-Native ART article for context. It contrasts organising around products against organising an ART around a system, a function, or a project. Which is exactly what organise around value has always argued: don't build a train around a platform, a department, or a piece of work with an end date. Same same but different.

New to the ART

Number six is the new one. Though AI governance and ethics aren't new to SAFe. What is new is that this is now an explicit expectation of an AI-Native ART.

This responsibility cannot be delegated to the teams. (I think Deming would approve!) It manifests as a centralised responsibility for setting standards and policies that teams work within. The ART gets two responsibilities: ensuring education and adherence to the portfolio's AI governance, ethical guidelines and guardrails, and defining any extra its own products need.

Four practices sit underneath ensuring AI governance and ethics:

1. Defining policies around acceptable risk
2. Integrating AI governance into shared platforms
3. Establishing standards for consistency
4. Operationalising human accountability

These practices make a lot of sense for anyone deploying AI agents, regardless of what operating model they're running: Core SAFe, AI-Native SAFe, or something else completely different. As Rebecca said, "Agents require management. People require leadership."

The Bottleneck Moved

Two weeks ago, Andrew told us the barrier to building solutions had evaporated, crediting Mik Kersten. (Kersten's Output to Outcome is out now. It's downloaded on my Kindle, but that's as far as I've got!) Rebecca picked up the same thread and took it somewhere new. Teams now produce significantly more output than before, she said, “in some organisations, pull requests alone have increased up to 100x.” Her conclusion: “that moves the bottleneck from creation to validation.”

In some ways, I'm not sure creation was ever the bottleneck, at least in the world of software development. I feel like every value stream map I've facilitated illustrates that the bigger delays are upstream and downstream of coding. But what does hold true, whether or not creation was ever the constraint, is that validation is certainly one now. We are producing more output than ever before, at a faster rate than ever before, but humans still need to review all of it.

Is the future more Kanban than Scrum?

The AI-Native Teams article describes four capabilities rather than roles: product, builder, domain expert, and AI. They use a flow-based approach built on three activities: align, sense and respond. Some events are scheduled, others are pulled when they're needed - the team decides what works for them. 

What's gone is the iteration boundary. Nothing happens because it's the end of the sprint. When I wrote Is it SAFe to Kanban? I quoted David Anderson: “Kanban dispenses with the time-boxed iteration and instead decouples the activities of prioritisation, development, and delivery. The cadence of each is allowed to adjust to its own natural level.” 

It sounds like a SAFe Kanban team to me but I guess that is up to the team to decide. ;-)

In my view, the concept of a team having capabilities rather than roles is a real strength. I like how it “demands collective ownership” of the outcomes. I’ve always said Agile is a team sport. This guidance should help organisations get better at the game. 

No Playbook and No Training Wheels

I keep coming back to this section of the AI-Native team article:

Importantly, the AI-Native Team does not have a set blueprint for the events it implements to execute these activities. The team must determine its own interaction patterns based on the context in which it operates.

I feel like it is asking a lot of a team that is already working in a very new and highly dynamic context. (Remember that validation bottleneck?)

Helping the team design and improve the way the work works has historically been the domain of the Scrum Masters (or Team Coach if you prefer). 

As I wrote in Is it SAFe to Kanban?, Scrum comes with a playbook, the Scrum Guide, which I see as providing training wheels for new agile teams. Kanban has no equivalent. Teams have to be disciplined about WIP limits, build a system with clear policies, and then work to reduce the WIP limits over time. That was my conclusion then and I'd still argue it: Kanban isn't a good choice for teams that lack discipline.

There's no Scrum Master in an AI-Native team. The word doesn't appear in the ART article or the Teams article. It appears in the AI Value Architect article as a candidate pool, in which Scrum Masters and Team Coaches are natural candidates for the new AI Value Architect role.  

While the AI Value Architect is described as coaching multiple teams and solutions, this coaching focuses on AI Adoption rather than team-level agility or flow. For me, this raises the question of who will help teams determine their “own interaction patterns”?

Part of me says a high-performing Agile team will work this out for itself as it moves to becoming an AI-native team; however, every organisation that has ever invited me to help them uplift their Agile practice has one thing in common: no Scrum Masters. Perhaps this will be addressed later in the series.

An AI Value Architect by any other name (Thanks, Juliet)

I hadn't seen a role called AI Value Architect before it appeared in AI-Native SAFe, but I have seen organisations create new roles or hats with similar aspirations, around coaching AI adoption and unlocking value from AI workflows and tools. With the creation of the AI Value Architect role and accompanying responsibilities, Scaled Agile has provided a solution to a gap that many organisations have been trying to fill.

Overproduce on Purpose AND Own Every Word

These two expectations of the AI-Native Team will cause some tension. They quite possibly already are. We know that output is now unconstrained, and this creates the opportunity to overproduce in the name of exploration and innovation. However, we also know that Generative AI can be unpredictable (that’s the generative part!). So its work needs to be validated, and keeping a human in the loop is critical here. 

Sustainable pace is going to become an imperative for team and ART success in AI-Native SAFe. I am feeling this acutely myself at the moment. For the last few weeks, I've been enjoying the access to Claude Fable included in my Claude subscription, and I have generated more content than I can consume. Nobody made me do it. I just wanted to maximise my output! Ironic, I know. I have been kicking off chats in quick succession so they run in parallel, and desperately trying to keep up with the response rate. (It's more like whack-a-mole than a sensible use of AI!) Fifteen years of teaching WIP limits destroyed by one limited-access frontier AI model!

I suspect I’m not unique in this regard, so I find myself wondering how a team with no blueprint and no Scrum Master will apply responsible AI practices and maintain a sustainable pace in this new world. (And if they work it out, maybe they can shoot me some tips - thanks in advance.) 

The focus of session four will be SAFe events in the age of AI. Perhaps the answers will be there.

What’s next

Session four went to SAFe events in the age of AI: What Happened to PI Planning in AI-Native SAFe? Session five follows two weeks after that: Building AI-Empowered Products with continuous innovation and governance. Session six, Adopting AI-Native SAFe, is on 25 August, aligning with the release of the AI-Native upgrade path for active SAFe certification holders and the launch of the new AI-Native Value Architect course (an evolution of the current AI-Native Change Agent offering). See you in a couple of weeks.