Skip to content
MACH X | September 29–30, Amsterdam | Register Now
Tile hero

What changed (Part 3 of 3): Moving through the shift

Jul 1 2026 - By Alliance

A three-part position paper from the MACH Alliance Agent Adoption & Operations Working Group

Series – What changed: Moving through the shift

Previously in this series — Part 1 set out what changed; Part 2 examined the three mental models that fail, the individual "click," and what an organization looks like once it has made the shift. This final part turns to the barriers in the way, and how to move through them.

Why the shift is hard: specific barriers

The mental model shift is unlikely to happen automatically, even with direct exposure to AI. Naming the specific barriers helps organizations address them rather than waiting for them to resolve on their own.

At the individual level, the most well-documented barrier is automation bias. Research by Parasuraman and Manzey, published in Human Factors in 2010 and now with over a thousand citations, established that people operating alongside automated systems systematically over-rely on those systems, even after gaining experience with them. Critically, this is not overcome by simple practice. It is a structural feature of how human attention works in multi-task environments. [3] The individual click carries a risk: the shift in how problems are perceived can shade into insufficient oversight of what agents actually do.

Other individual barriers include:

  • Personal willingness to adopt AI based on biases about the tech
  • Bandwidth and cognitive capacity to learn the new technology
  • Confusion or overwhelm caused by market messages and news about AI

The second individual barrier is early-failure dismissal. Someone tries AI on a problem, gets a poor result, and concludes the technology is unreliable. This is usually less a judgment about capability than about fit. AI agents perform very differently across problem types, and a bad early experience with one kind of task predicts little about performance on others. Organizations that do not create structured, guided exposure to AI risk allowing early failures to calcify into resistance that is hard to reverse.

At the organizational level, the most significant barrier is framing AI adoption as a technology evaluation rather than an organizational capability build. Technology evaluations have an endpoint: you assess a tool, decide whether to adopt it, then deploy it. Capability development does not have such an endpoint. You build the skills, processes, structures, and operating models that allow a new kind of work to be done well over time. AI adoption structured as a technology evaluation tends to produce pilots. Structured as capability development, it creates the environment to produce sustained value.

A related barrier is leadership-team misalignment. In many organizations, the people most excited about AI's potential are driving faster and broader adoption, while risk, legal, compliance, and operations teams are applying legacy frameworks that were not designed for probabilistic, agentic systems. Neither instinct is wrong. The challenge is that they are operating on different mental models, and the resulting friction produces neither sufficient governance nor sufficient adoption.

Moving through the shift

This working group is researching and developing materials to help organizations move through the shift at scale. Below are some orienting principles.

For individuals, sustained exposure on real problems matters more than training or awareness. Demonstration sessions raise awareness, but they do not produce "the click". People need to use AI agents on work that actually matters to them, long enough to develop genuine intuition. They need more than just access to the tools. They need deliberate practice on meaningful work. Organizations that want to accelerate individual adoption need to create the conditions for that exposure.

Teams need to do this together. Shared exposure creates shared vocabulary, and shared vocabulary is what makes governance conversations concrete rather than abstract. Teams that have worked through real problems with AI can discuss its capabilities and limitations specifically. Teams without experience working on real problems tend to talk past each other at a level of generality that makes it hard to agree on anything actionable with agentic AI.

For organizations, the reframe that matters most is that AI adoption is not a technology evaluation. It is instead an operational capability to be developed. Technology evaluations have an end date. Capability development does not. Organizations that invest in roles, monitoring infrastructure, and shared process alongside agent deployment compound their returns. Those that treat it as a deployment project tend to hit a ceiling where governance failures create enough friction that adoption stalls.

Why frameworks are the organizational mechanism

The individual mental model shift is necessary. It is not enough.

A single practitioner who has internalized the generative, agentic nature of AI cannot, by doing so alone, produce accountable, well-governed AI deployment across an enterprise. The insight needs to be institutionalized. It needs to be encoded into processes, roles, shared vocabulary, and operational structures that persist beyond any individual and function without requiring every person involved to have had "the click" independently.

This is what governance frameworks, operating models, and readiness standards do. They are not bureaucratic overhead imposed on AI. They are the organizational mechanism for scaling the mental model shift. A governance framework that asks "who owns this agent in production, and how is its behavior monitored?" is asking the questions that an individual who has made the shift would ask. And, it asks them systematically, before deployment, for every agent, regardless of which team built it.

The research topics and areas for exploration proposed from this working group reflect this logic directly:

  1. Enterprise Agent Adoption Framework giving organizations a structured path from experimentation to scaled operational capability, preventing the pilot stagnation that results when adoption outpaces operational readiness.
  2. Enterprise Agent Operating Model defining roles, ownership, and accountability across the organization, ensuring that agents in production have owners, not just builders.
  3. Agent Governance Framework encoding the judgment calls that individuals with the right mental model make intuitively: when agents should act autonomously, when they should escalate, and how authority boundaries get defined.
  4. Agent Production Readiness Checklist asking the organizational equivalent of the shift question ("Are we ready to operate this?") before deployment, not after failures.
  5. Agent Operations Playbook providing the operational structures for monitoring, incident response, and lifecycle management that agents require as long-lived production systems.
  6. Agent Observability Model addressing the specific challenge of understanding what a probabilistic, goal-pursuing system actually did and why, which is a different problem from debugging deterministic software.

Together, these give organizations a way to operate in the post-shift world even before everyone has had the individual "click". They distribute the insight.

Conclusion

The mental models that enterprises use to make sense of AI agents are being established now in how organizations structure pilots, in how governance conversations get framed, in what roles get created, in what questions leaders learn to ask. Once those models are set, they are difficult to revise.

Organizations that do this work now will have a different kind of advantage: not better access to AI capabilities, but genuine operational maturity. They will know how to govern agents at scale because they built the structures for it early, not just because it seemed like the right thing to do in retrospect.

Organizations that delay, or that try to manage agentic AI with frameworks designed for different kinds of systems, will encounter a version of the ceiling that already shows up in the data: widespread deployment, limited value, and a growing gap between what AI agents could do and what the organization can safely permit.

The deliverables this working group is building are practical instruments (shared vocabulary, operational frameworks, governance guidance) for helping more organizations make the shift faster and more safely. That work has to start somewhere. It starts with being honest about what changed.

Sources

This position paper is published by the MACH Alliance Agent Adoption & Operations Working Group, part of the MACH Alliance Agent Ecosystem initiative. Working group membership, charter, and deliverables are maintained at github.com/machalliance/wg-agent-adoption-operations.

Members of the MACH Alliance Agent Adoption & Operations Working Group:

Andrew Kumar, Dylan Valade, Everett Zufelt, Ryan Lunka, Jennifer Wright, Nick Tomassetti, Tim Steele, Tomasz Pindel, Ashwin Mudaliar, Lance Mercereau, Husain Khan, Deepak Jaidka, Peter German, Itay Droog, Sezin Cagil, Yann Boisclair-Roy, Dolapo Amusan & Gus Fune.