Agentic AI for the Enterprise: One Practitioner’s Field Notes
By Dr. Harish Kotadia, Ph.D.
I have spent close to two decades building enterprise systems for Fortune 100 clients. I have watched three real shifts in that time, and I am convinced the one happening now is the largest of them. Over the past few weeks I have written it down in pieces. This is the through-line that connects all of them, written for the people who have to decide what to do about it: the CIO, the CTO, and the executives sitting in the room when the budget is approved.
Start with the word itself, because most of the confusion lives there. Ask ten experts what agentic AI is and you will get twelve answers. I am not guessing at that number. I went and read the field, which is why I published Agentic AI: 50 Authoritative Definitions — The Complete Citation Compendium (2026), pulling from peer-reviewed journals, the major advisory firms, and the platform vendors. Then I sat with all fifty and asked what they actually agreed on, which became Review of 50 Agentic AI Definitions — Comprehensive Analysis. The honest finding: even Gartner and MIT Sloan concede there is no settled definition. That sounds like a problem. It is closer to the most useful fact in the whole space.
Because once you stop waiting for a tidy definition, you can write one that helps the people doing the work. That is what I Read 50 Definitions of Agentic AI. Here’s the One the Builders Actually Need is for, and the cleaner statement of it sits in What is Agentic AI? Definition of Agentic AI. Here is the line I keep coming back to.
Instructions in, results out — that was IT. Intent in, outcomes out — that’s agentic AI.
That shift is not a feature you bolt onto an application. It is a new layer in the stack. We went from physical servers to virtual machines, from VMs to containers, from containers to serverless. Each step raised the level of abstraction the business builds on. Agentic AI is the next rung: a governed, goal-driven software layer where language-model agents, given memory, tools, and a way to coordinate, plan and carry out multi-step work across your cloud, your data, and your applications. You stop describing every step. You describe the outcome you want and the boundaries you will allow.
The part everyone gets wrong: it is the loop, not the model
Most enterprise AI is being built on the wrong layer, and I mean that literally. Teams pick a model, wire it into a feature, and call the project finished. The model was never the product. The loop is. An agent observes, decides, acts through a tool, sees the result, and decides again. That cycle is where the value lives and where the risk lives. I made that case three times over because it is the single idea most organizations are still missing: in Agentic AI: It Is Not the Model. It Is the Loop, in Agentic AI Era: Model is No Longer Product. The Loop Is., and in the executive framing of CIO Guide to Agentic AI.
If you take one thing from this section, take this. When you procure a model you are buying a component. When you stand up a loop you are buying an operating model. The governance, the audit trail, the question of what an agent is allowed to touch and what gets logged when it acts — all of that attaches to the loop, not to the model you happened to choose. Treat it as a model decision and you will be surprised later, in the expensive way.
There is a cost to getting this wrong, and it is not hypothetical. I laid it out in Agentic AI: The Hidden Cost of NOT Doing It RIGHT. Organizations running real agentic workflows are reporting efficiency gains in the order of a third, with materially faster execution. Gartner expects a large share of enterprise applications to embed task-specific agents before long. Meanwhile plenty of companies are running linear prompt chains, calling it transformation, and quietly falling behind competitors who built the loop properly. Standing still is a decision with a price tag.
Where the loop runs is a governance decision, not a preference
Once you accept that the loop is the product, the next question is where it runs. This is the moment people reach for habit and pick the cloud they already use. I would slow that down. In Agentic AI: Why Cloud Platform Choice Matters More Than You Think the argument is that this is not really an AI question at all. It is an infrastructure, risk, and compliance question, and it lands on the desks of the teams who have to let an agent act inside regulated systems.
The practical comparison is in Agentic AI: Two Clouds, One Claude. You can run the same Claude model on AWS Bedrock or on Google Vertex AI and end up with two genuinely different architectures around it. Bedrock gives you serverless, pay-per-token access with almost no infrastructure to manage. Vertex pulls you toward a different operational shape. Same model, different controls, different bill. The model is portable. The decision is not.
Two systems I actually built
None of this is theory for me, and I have no patience for AI writing that pretends otherwise. The auto loan platform I built around Claude did not treat the model as a feature on the side. It put the model in the middle, as the operating model. The honest account of how it came together, including what was harder than expected, is in Agentic AI Case Study: Auto Loan Processing on Claude.
The second one started with a question a Fortune 100 financial services CIO put to his team, the kind that sounds naive until you sit with it. Why does it still take days to issue a credit card when we already know everything about the applicant in seconds? Answering it properly meant rebuilding origination as a multi-agent system rather than a faster version of the old workflow. That work is written up in Agentic AI Case Study: Credit Card Origination on Claude and AWS. Both projects taught the same lesson from different angles. The technology was rarely the bottleneck. The operating model around it was.
The consequence nobody on your vendor’s side wants to say out loud
Here is where this stops being an architecture conversation and becomes a business one. The whole IT services industry has run on the same model for fifty years: bill for hours, staff a pyramid of people, grow revenue by growing headcount. Agentic AI breaks the link between hours worked and value delivered, and your suppliers know it.
The crack first showed up in a procurement meeting, not a lab. A client’s procurement lead looked at a managed services renewal and asked, almost in passing, why the price was the same now that the delivery team was using AI. I have not been able to unhear it. That is The Question Every IT Services Firm Is About to Be Asked. The follow-on, on what firms can actually do about it, is Agentic AI: How to Re-Invent the IT Services Industry. The structural picture — the old staffing pyramid collapsing into something closer to a diamond, with fewer junior hands and more senior judgment — is in IT Services Pyramid Had a Fifty-Year Run. Here’s What Replaces It.
This is not abstract anymore. The market has started to price it in, which I traced in How Agentic AI Is Restructuring the IT Services Model after watching nearly every listed services name sell off on the same day. And the practical question that lands back on the CIO — if the pyramid becomes a diamond, where should each kind of work actually sit — is what I work through in IT Services Delivery Models in the Age of Agentic AI. The short version: some work moves to agents, some stays with senior people, and the contracts you sign in the next two years should already assume the shift.
Where the real work is happening
One more thing for anyone trying to read the signal through the noise. The announcements worth a CIO’s attention right now are not the flashy demos. They are the unglamorous mechanics of running agents in production: who is allowed to act, what an agent can touch, how its actions are logged, where the boundaries hold. I track that week by week, and the most recent roundup is Agentic AI News: Top 10 Enterprise Stories for CIOs and CTOs. When the headlines move toward permissions, logging, and control rather than raw capability, that is the market growing up. Plan accordingly.
What I would do on Monday
Pull the thread back together. Agentic AI is a new abstraction layer, not a smarter chatbot. The value and the risk both live in the loop, not the model. Where that loop runs is a governance call your risk teams need to own. The systems that work treat AI as the operating model, not a feature. And the business model your suppliers have relied on for fifty years is being repriced in real time, which changes how you should be signing contracts this year.
If you read the pieces above in order, the case builds from definition to consequence to action. If you only have time for the one decision that matters most this quarter, make it this one: stop buying models and start designing loops, with the controls attached from day one. The firms that figure that out will spend the next few years compounding the advantage. The ones still chaining prompts will spend them explaining to their boards why the competition pulled away.
© Dr. Harish Kotadia, 2026. All Rights Reserved.
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI. He holds a Ph.D. in Marketing Management with doctoral research in marketing analytics. Follow him at @agenticaiarch on X and at AgenticAIArch.com.