Over the past year, most conversations about AI have focused on what these systems can produce. The more important question is whether they can be trusted to act. That shift changes everything.
There are a handful of books that arrive at exactly the right moment — not because they predict the future, but because they give you the right framework to observe it clearly.
Agentic Artificial Intelligence, by Pascal Bornet and a team of 27 co-authors including Thomas H. Davenport, Jochen Wirtz, and David De Cremer, is one of those books. Published in March 2025, it set out to answer a deceptively simple question: what does it actually mean for AI to move from answering questions to taking action? And how do we make sense of the enormous spectrum between a basic automation script and a fully autonomous system?
I've been working through it carefully over the past year — not as an academic exercise, but as a practitioner trying to make sense of what I'm seeing in the market. What follows are my reflections, one year on. I'm not here to improve on what Bornet and his colleagues wrote. I'm here to share what it looks like when you hold their framework up against a year of real-world evidence.
The book's central contribution is what the authors call the Agentic AI Progression Framework — a five-level taxonomy running from basic automation to full autonomy. They use the self-driving car analogy throughout, and it earns its keep.
At Level 1, you have rule-based automation: fixed scripts, RPA tools, if-then logic. No intelligence, no adaptability. The digital equivalent of cruise control — useful in the right conditions, brittle when anything unexpected happens.
Level 2 introduces genuine cognitive capability: machine learning, natural language processing, computer vision. Systems that can understand unstructured data, make predictions, handle semi-structured inputs. Advanced driver assistance, in the automotive analogy — the system is doing real work, but a human is still supervising.
Level 3 is where true AI agents emerge. Systems that can plan multi-step processes, chain tools together, reason through complexity, and adapt based on context. This is autonomous highway driving — capable in most conditions, but still needing human intervention at the edges. Large language models, memory, and tool-use form the technical foundation here.
Level 4 approaches genuine autonomy: systems that can set their own sub-goals, learn from experience, and adapt their strategies over time. The authors treated this level as largely experimental when they wrote the book — and for good reason. Level 5 — full, unrestricted autonomy — they described as theoretical, requiring capabilities that don't yet exist.
One thing the authors were careful to emphasize, and which I think is underappreciated in the public conversation, is that this isn't a hierarchy where higher is always better. A Level 2 deployment that reliably reduces processing costs by 40% may deliver more real value than a Level 4 experiment that never makes it out of the pilot phase. Matching the right level of capability to the right problem is the actual strategic discipline.
Twelve months on, that observation feels more important, not less.
The most honest thing I can say about 2025 is that it was a year of significant capability progress, and simultaneously, a year of significant reality-checking.
Progress at Level 3 has been faster than most people expected. A year ago, the most capable deployed systems were still primarily answering questions and generating content. Today, the better ones can break down complex tasks, interact with external tools, synthesize research across multiple steps, and maintain context across extended interactions. The shift from "chat" to "task-oriented reasoning" happened quickly. If anything, the book may have slightly underestimated how fast Level 3 would mature.
But Level 4 has arrived in a narrower form than many anticipated. It was easy, twelve months ago, to imagine Level 4 as a near-term leap toward broadly autonomous systems. What has actually emerged is something more constrained. Level 4 capabilities are appearing — but they are domain-specific, heavily supervised, and bounded by carefully designed environments. Agents can now complete sophisticated workflows, interact with enterprise systems, and execute multi-step tasks. But only within guardrails that took significant effort to build and maintain. What Level 4 looks like in practice, so far, is less "autonomy" and more structured, supervised delegation. That's not a failure — it's a more honest description of where the technology actually is.
The most important realization of the past year, though, may be the one that the book pointed toward but that only becomes unmistakably clear when you're watching real deployments: the hard problem isn't intelligence. It's reliability.
The models themselves are genuinely impressive. But deploying agents in real enterprise environments exposes a different set of challenges — edge cases, failure recovery, consistency, trust. The gap between "can this system do the task once, in a controlled setting?" and "can this system do this task reliably, at scale, without breaking things or requiring constant supervision?" turns out to be enormous. Gartner has predicted that over 40% of agentic AI projects will be cancelled by the end of 2027 — not because the technology doesn't work, but because organizations are discovering that production deployment is a fundamentally different problem from proof-of-concept.
Several things I suspected before reading the book have been reinforced by both the framework and the year's evidence.
The durable value in this transition is unlikely to sit at the model layer. It will sit at the intersection of systems, data, workflows, and governance — in other words, where work actually happens. The hyperscalers are investing in infrastructure. The model labs are racing to improve capabilities. But the enterprise platforms that appear to be winning — Salesforce, ServiceNow, Workday, SAP — are focusing their energy on something different: orchestration, control, and the governance layers that allow agentic systems to operate inside real organizational environments. That's not a coincidence.
The companies with the strongest positions are, in many cases, the ones whose competitive advantages used to be considered unglamorous. Deep workflow complexity. Regulatory compliance infrastructure. Audit trails. The "boring" moats. As agents begin to take consequential action inside organizations, the ability to monitor, govern, and prove what happened turns out to matter enormously. The book identified this. The past year has made it a front-page strategic reality.
There is one area where the framework, as I originally understood it, felt slightly too linear.
A year ago, I read the five levels as a progression — a path organizations would travel, broadly in sequence, toward increasing autonomy. What has become clearer is that Level 4 is better understood as a design space than a destination. Different organizations will adopt different levels of autonomy depending on their risk tolerance, regulatory environment, operational complexity, and the specific workflows involved. In many cases, partial autonomy — a well-governed Level 3 or an early, bounded Level 4 — may be the optimal configuration indefinitely. Not because the technology isn't capable of more, but because the governance requirements, accountability structures, and human judgment needs of certain domains make full autonomy neither desirable nor appropriate.
That's a meaningful shift. The question isn't really "how quickly can we get to Level 4 or 5?" The question is "what is the right level of autonomy for this specific workflow, in this specific context, with this specific risk profile?" That's a leadership question as much as a technology question.
If I had to reduce everything I've observed this year to a single idea, it would be this: the industry has largely figured out how to build systems that can think. It is still figuring out how to build systems that can act reliably.
And that gap between thinking and acting — between impressive capability in a controlled environment and dependable, governed performance in a real one — is where most of the genuinely important work is happening right now.
Bornet and his colleagues wrote a framework that helps you see that clearly. A year of reality has done nothing to undermine it. If anything, it has sharpened it.
I've been speaking with colleagues and friends about my observations, genuinely curious what others are seeing. Are the organizations you work with primarily at Level 2? Beginning to explore Level 3? And what's the single biggest barrier you're running into?
#AgenticAI #EnterpriseSoftware #SaaS #AIStrategy #FutureOfWork #DigitalStrategy
