Marco Gaietti brings a perspective forged in the trenches of global management consulting, where the abstract promises of “autonomous” systems often collide with the messy reality of spreadsheets and legacy databases. With decades spent dissecting how enterprises actually function, he has become a leading voice in tempering the hype surrounding AI agents with a dose of operational pragmatism. His work centers on the bridge between what technology can do in a vacuum and what it must do to survive the rigors of corporate governance and security. In this discussion, we explore the shift from the fully autonomous dream to the tangible value of partial automation. We delve into why the ERP must remain the system of record for AI to be trustworthy and how the granularization of workflows—turning 10 human steps into 50 machine-driven actions—is redefining efficiency. Ultimately, the conversation highlights that the real hurdle isn’t just the code, but the operating model required to govern a fleet of agents working across disparate platforms.
While some vendors promise a fully autonomous enterprise, partial automation is often seen as the only achievable reality for most organizations. How do you reconcile the grand vision of an agentic future with the practical constraints businesses face today?
The gap between the glossy vendor slide and the cold reality of a production environment is wide, and frankly, it is likely to stay that way for the foreseeable future. When we talk about an autonomous enterprise, we are describing a world where AI agents and Large Language Models sit at the very center of every process, but that ignores the heavy friction of data quality, compliance, and interoperability. In my experience, the promise of total autonomy is technically possible for small, isolated use cases, yet it becomes a nightmare once you try to scale it across the entire enterprise stack. This is why I advocate for the “partial automation” frame; it acknowledges that most companies are currently in a state where agentic AI can accelerate repetitive work but still requires a foundation of orchestration and contextual awareness that is hard to perfect. We have to move past the demo story and get into the deployment story, where we acknowledge that broader autonomy depends on hard-won foundations like permissions and system integration. It feels much more like building a bridge one brick at a time rather than flipping a switch and watching the lights turn on across the entire organization.
When evaluating these new AI agents, buyers are often dazzled by what they see in a controlled environment. What specific questions should leaders be asking to determine if a demo can actually translate into a secure, production-ready deployment?
A demo is a performance, but a deployment is a commitment that requires a high level of scrutiny. To avoid the trap of “shiny-object syndrome,” buyers need to pull back the curtain and ask exactly what is automated right now versus what is still just a bullet point on a future roadmap. You have to interrogate which parts of the workflow are truly autonomous and which parts still require that essential human review or handoff. I always tell my clients to ask: “What data foundation does this agent depend on?” If the vendor cannot explain how the agent respects identity rules and authorization controls, then you are looking at a security risk waiting to happen. You need to know if the agent can act repeatedly and safely when the workflow touches real business consequences, like moving money or changing supply chain orders. Without that governance, you are essentially letting a powerful tool run loose in a glass shop without any insurance.
You have often mentioned that AI agents need a trusted system of record to be effective. Why is it that platforms like ERP or CRM are the natural starting points for this technology, rather than just building agents on top of general data lakes?
The reason is fundamentally about trust and the integrity of the data being used. An AI agent is only as useful as the foundation it acts from, and in the corporate world, the ERP remains the ultimate system of record for running the business. Whether it is finance, procurement, or supply chain data, these systems already have the governed enterprise data and identity rules that an agent needs to make a valid, compliant decision. If you try to run an agent on unstructured, ungoverned data, it might summarize a meeting beautifully, but it will fail the moment it needs to match an invoice or check supplier risk. We are seeing major players like SAP push toward an autonomous model because they know that AI becomes transformative only when it draws from a context layer that is already verified. It doesn’t mean every single agent has to live in the ERP, but you must know where the data comes from and which system owns the record. When an agent needs to reach across multiple applications, that system of record acts as the North Star, ensuring that permissions are enforced and the logic remains sound.
There is an interesting shift happening where agents are making workflows more granular—essentially breaking a 10-step process into 50 smaller steps. Does this actually simplify the business, or are we just trading one kind of complexity for another?
It is a paradox because the workflow might look simpler to the end-user, but the underlying machinery becomes significantly more complex to manage. When an agent takes over a back-office process that used to be a few human approvals, it might break that down into 50 structured, machine-executable actions. These are tiny, granular steps—extracting data from a PDF, matching it against a record, triggering a compliance check, and then routing it to the next node. While this granularity allows the agent to move faster and more consistently than any human ever could, it creates a new burden of orchestration for the IT and operations teams. Organizations have to ask themselves: “Who owns the agent when it stalls or makes a mistake?” You are essentially building a complex engine where every gear must be perfectly aligned. If you don’t have serious governance over these agent-driven workflows, you are just automating chaos at a higher velocity.
As companies start deploying agents from different vendors—some for HR, some for CRM, and others for IT—what are the biggest risks of this fragmented multi-agent environment?
The real danger is the operating model problem, which is often overlooked in favor of the technical specs. It isn’t enough for each agent to work in its own little silo; you have to worry about what happens when they all start touching the same work or accessing the same records. One vendor’s agent might be updating a record in the ERP while another is summarizing a customer interaction in the CRM, and if they aren’t communicating, you get conflicts. This is where interoperability and orchestration stop being buzzwords and become survival requirements for the enterprise. You have to decide which agents are allowed to do what, when they must hand work back to a human, and how they avoid stepping on each other’s toes. If these agents cannot respect business rules across vendor platforms, your autonomous enterprise will quickly become a collection of digital islands that can’t talk to each other. It takes a very disciplined approach to ensure that as these tools expand from one use case to many, they still behave in a way that is secure and compliant.
We see partial automation handling things like invoice matching and HR case routing today. Is this just a stepping stone, or is this the permanent plateau for most enterprise AI?
Partial automation should never be seen as a failure or a “small idea.” It is the connective tissue of enterprise work. When you look at use cases like supplier risk checks, meeting action items, or IT ticket triage, you are looking at the very points where delays and bad handoffs usually happen in a modern business. These aren’t trivial tasks; they are the gears that keep the company moving forward. By taking these well-defined pieces of work and making them faster and more consistent, you are building the data and governance muscle needed for whatever comes next. Most companies don’t transform overnight; they clean up a process here and connect a data source there. Over time, those changes add up to something significant. For now, focusing on these practical entry points is the most responsible path. It allows you to find immediate value while you figure out the much harder questions of how to govern a truly autonomous system in the future.
What is your forecast for the evolution of the autonomous enterprise over the next five years?
I believe the future will be defined by a shift from “agentic features” to “orchestrated ecosystems,” where the focus moves away from what a single AI can do to how a fleet of agents can be managed as a cohesive workforce. We will see the autonomous enterprise transition from a marketing slogan to a rigorous operating model, where human review remains essential but is strategically placed at high-stakes decision points rather than in every administrative step. Within five years, the organizations that win will be those that have spent today perfecting their data foundations and governance rules, allowing them to finally bridge the gap between impressive demos and reliable, secure production at scale. The enterprise won’t be fully autonomous in the sense that humans disappear, but it will be autonomous in the sense that the “busy work” becomes invisible, leaving us to manage the exceptions and the strategy. We will eventually stop talking about the AI itself and start focusing on the results of these granular workflows.
