Agentic AI Model Never Decides the Outcome. The Controls Do.

The best agentic AI case studies I have published in the last year share one finding, and it is not about models. Across ten posts covering more than sixty enterprise agentic AI deployments, not one outcome, good or bad, was decided by which model the company picked. Every one was decided by what the agent was allowed to do, who signed off before it acted, and whether the number in the headline was ever measured against a baseline.

This is the summer send-off post. I went back through everything on this site since it launched and pulled the ten posts I would stand behind as the best of the best: two that set the vocabulary, five that document deployments that delivered, and three that put a success next to a failure in the same industry. Read them in that order and they make one argument. Governance is the question every case raises, and architecture is the answer every surviving case gives.

Two posts set the terms

Nothing below makes sense without a definition, so start with my working definition of agentic AI: a governed, goal-driven software layer in which LLM-powered agents plan and execute multi-step business processes. The word governed comes before goal-driven on purpose. If the agent is not governed, what you are reading is a demo, not a case study.

The second post is the ruler I hold every case against. The Five-Stage Agentic AI Roadmap runs from Prompted through Piloted, Governed and Assured to Autonomous, and each stage is earned by institutionalizing the one below it. I place every deployment by the stage its evidence supports, never the stage its press release claims. Most of the market sits at Piloted and does not know it.

Five posts document agents that delivered

The widest lens is the fifty enterprise case studies I collated and mapped in June from vendor libraries, press releases and trade press, grouped by business function and plotted on the roadmap. The chart is the finding. The mass sits at Piloted, thins toward Assured, and none evidence Autonomous. Most of the figures are self-reported, so I read them as direction rather than benchmark, and I say so in the post.

The narrowest lens is the post where only three case studies passed my ROI test. I set four bars: genuinely agentic behavior, hard dollars, a named executive or an independent business journal behind the number, and production scale. Three cleared all four: a global software provider saving $100 million a year on support, a retail bank that cut fraud losses more than 20 percent with an agent that drafts detection rules a human approves one by one, and a packaged-food manufacturer that took $20 million out of its supply chain by cutting truckload optimization from 18 hours to under 30 minutes. All three are narrow, high-volume, rules-bound workflows with an escalation lane designed in on day one. Nobody handed an agent open-ended autonomy and hoped.

Between those two sit three posts I still send to architects. Eleven Claude customer stories, put through the same what-they-built, stack and outcome table, show most had already crossed into Governed before the agent touched anything that mattered. Five financial-services deployments put agents to work against roughly $2.5 trillion in assets under supervision, with client onboarding 30 percent faster and developer productivity up 20 percent or more; I placed the lead case at Assured because its numbers were measured against a real baseline with an audit trail, and nobody claimed Autonomous. And my own auto loan origination case, written as an honest account rather than a brochure, is the one I hand to people who want to see the seams. The model sat in the middle as the operating model, not on the side as a feature, and the hard edges stayed with a person.

Three posts put a success next to a failure

The pairs are where the pattern stops being a hypothesis. In the healthcare versus fintech comparison, a large United States health system’s prior-authorization agent routes every determination through a human before it reaches a payer, and I place it at Governed moving to Assured. A European fintech claimed Autonomous in its first month, could not tell which conversations needed a human, dropped the context on handoff so customers had to start over, and then reversed an AI-driven hiring freeze. The split is where the human sits: before the decision, or after the agent decided it did not need one.

Two 2025 coding-agent deployments make the same point in a different industry. A Wall Street bank gave its agent no standing production access, routed every task through a ticket and every output through a supervising engineer, and reported a three-to-four-times productivity gain with no uncontrolled incident. A consumer coding platform left its agent with write access to production behind nothing but a natural-language instruction; the agent deleted a live database during a declared code freeze and then fabricated the record of what it had done. Those are not different degrees of the same control. They are different categories.

The third pair is two banks that ran the same play and landed in opposite places. A 240-year-old custody bank runs 134 digital employees, each with its own login, a narrow scope and a human manager, and cut contract review time 75 percent. Australia’s largest bank made 45 customer-service roles redundant on a claimed reduction of 2,000 calls a week that its voice bot could not sustain, and reversed the cuts within weeks. That was Piloted, sold as Autonomous.

What the failures share

Put the three pairs together and the failures carry a signature. The human was placed after the agent’s decision. The metric was asserted on a slide rather than measured against a baseline. The permission was a sentence in a prompt rather than a boundary in the architecture. I have not yet found a public failure that breaks this pattern, and I keep looking, because the day I find one is the day the roadmap needs a sixth stage.

So before I believe a case study, I run three checks. I place the deployment on the roadmap by its weakest production agent, not its best pilot, because every reversal above claimed a stage it had not earned. I write down where the human sits, before the decision on every case or after the agent decided it did not need one; both failures in the coding and fintech pairs had it after. And I ask whether the number was measured or asserted; the bank that cut headcount did so on an asserted number, while the bank that approves every rule its agent drafts still cut fraud losses 20 percent.

What I take into a risk committee

Would this corpus pass a risk committee? The seven successes would, for a boring reason. Each one drew the autonomy boundary before the first agent ran. In my work in regulated loan origination that is the whole difference between a deployment that survives audit and one that survives until the first incident. Instructions in, results out was IT. Intent in, outcomes out is agentic AI, and it only holds when the controls exist.

Which of the ten would you argue with, and which case have I missed?

My books go deeper on both: Intent In, Outcomes Out and Earned Autonomy.

Go deeper

Agentic AI Architect: control design for enterprise agents.
Agentic AI Case Studies: deployment evidence, one teardown at a time.
Agentic AI P&L: cost, payback, and risk in a CFO’s voice.
Agentic AI Governance: who owns the harness.
Substack first access: agenticaiarch.substack.com/subscribe

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.

Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.

Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.