Everyone wants AI agents that can work on their own. The hard part is knowing what each step toward that looks like, and being honest about which step you are on today.

I built the Agentic AI Roadmap to make that clear. It has five stages, from Prompted to Autonomous. You reach each stage by doing the one before it well. Here is how I put it in the original post:

 

“It does not measure how clever your agents are. It measures how reliably, and how accountably, your organization turns intent into outcomes.

You earn autonomy by first earning assurance.”

— The Agentic AI Roadmap: 5 Levels From Prompted to Autonomous

The easiest way to understand the five stages is to picture a car that drives itself. Tesla Full Self-Driving is the most famous try at this, and it follows our five stages closely.

Start with the honest part. Tesla sells it as “Full Self-Driving.” But the version you can buy is really called Full Self-Driving (Supervised), and it is rated Level 2. That means a person still has to watch the road. The name says autonomous. The rating says assisted. Enterprise AI has the same gap, and the roadmap helps us see it.

Here is what your Tesla looks like at each stage.

Stage 1 — Prompted

No agents run on their own. People use AI tools one at a time. The result depends on who is using the tool, and nothing repeats across the company.

In the car: basic Autopilot. The car keeps your speed and holds the lane. You still do the driving. The car only helps in small moments. Turn it off and the help leaves with the driver.

Stage 2 — Piloted

A few agents run, but only in small tests. A person watches them and can stop them. Each one is built from scratch. Most companies get stuck here, which is why so many agent projects get canceled.

In the car: this is Full Self-Driving (Supervised) as it ships today. The car can drive city streets, change lanes, and park itself. But you must watch it, with your hands ready. If something looks wrong, you take over. Every drive is a test with a human standing by. Tesla has driven 10 billion miles this way, and the car still needs you.

Stage 3 — Governed

Agents become a company asset, not one person’s project. There is one shared design, one set of rules, and one place to register every agent.

In the car: the Robotaxi program. The car is no longer just yours. It is part of a managed fleet. It drives inside a fixed, mapped area, runs on one tested system, and is watched from a control room. As of mid-2026 it runs in Austin, Houston, and Dallas.

Stage 4 — Assured

You measure what the agent does against clear targets. You keep records good enough to show an auditor.

In the car: now the car has to prove itself. Tesla counts how many miles each car drives with no human help. It gathers billions of miles of data. Regulators test it too: in May 2026 the NHTSA said recent Model Ys were the first cars to pass its new safety benchmark. The robotaxis that drive with no safety person sit here, not because Tesla decided to trust them, but because the data made the case.

Stage 5 — Autonomous

The system runs and improves on its own. You trust it the way you trust a skilled professional, with no one standing by.

In the car: true full self-driving. You get in anywhere, there is no wheel to take, no map limits, and no person watching, in any weather. As of mid-2026 this does not exist, for Tesla or for anyone. Stage 5 is empty.

The shape is the same

When I mapped 50 enterprise AI case studies to this roadmap, the picture matched the car. Most projects sat at Piloted. Only a couple reached Assured. Stage 5 was empty. What the market calls “autonomous” is, on a closer look, supervised at best.

You earn autonomy by first earning assurance. Tesla cannot remove the driver until the miles, the numbers, and the regulators say it can. Your agent is no different.

I see this in my own regulated work. In auto loan and credit card systems, no agent runs on its own until the safety layer is solid first: audit trails, human checkpoints, and a way to undo. The order is not negotiable. Intent in, outcomes out, but only once you can prove the outcomes.

© 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. Views expressed are the author’s own and do not represent that of any employer or client. Follow him on LinkedIn at www.linkedin.com/in/hkotadia and at AgenticAIArch.com.