From Chat to Agent: The Real AI Adoption Gap

Nearly half of US businesses now pay for AI tools, up from single digits two years ago. That number sounds like the adoption problem is solved. It isn’t.

I’ve been on this curve myself. Tabnine as a VS Code plugin back in 2022, before chat interfaces even existed. I knew OpenAI mainly from their Gym reinforcement learning library, not as a consumer product. Then ChatGPT launched, then GitHub Copilot, Cursor, and now Claude Code. Along the way, running local inference on consumer GPUs, experimenting with RAG pipelines. With each step I learned more about how to connect the dots.

At some point I started treating AI as potential glue between topics I already understood. Not as a replacement for thinking, but as a way to act on ideas faster than I could wire things up myself. I stopped writing code line by line and started directing work like a PM: describing what I wanted, reviewing what came back, iterating fast. That’s how I ended up vibe coding an eBPF monitor in Go without being a Go developer.

The Chat Ceiling

BCG found that 85% of employees are stuck at basic chat and task assistance. Anthropic’s labor market research shows a gap between what AI can theoretically do (94% of tech worker tasks) and what people actually use it for (33%). And Nielsen Norman Group research explains why: frequent AI users are actually less likely to push interaction limits than new users. The chat box trains people to think small, and that ceiling hardens with use.

I noticed this in myself. Months of daily usage, but my prompts stayed in the same lane: “explain this error,” “rewrite this paragraph.” Useful, not transformative.

For a lot of people, the eye-opener was ClawdBot. A Telegram bot that scrapes the internet, acts as your personal trainer, gives food advice, manages tasks. No terminal, no setup. Just message it and it does things. But none of this is new. Developers have been automating these tasks for decades. A scheduled task is just a cron job. Browser automation used to be Selenium, now in a fancier way with a natural language interface. What changed is that now anyone can do it. The barrier between “I wish this was automated” and “it is automated” collapsed overnight.

The Pyramid

I think about AI adoption as four levels. Not a ranking, but a map.

The jump between levels isn’t about learning new software. It’s about what you believe the tool can do:

• L0 to L1: “AI can live inside my editor.” From a separate tab to embedded in your workflow.
• L1 to L2: “AI understands my project.” From suggesting lines to understanding context across files.
• L2 to L3: “AI can act without me watching.” You stop supervising and start delegating outcomes.

At L2, tool connections start to matter. Through protocols like MCP, your editor connects to your ticketing system, security platform, or cloud provider to fetch context automatically. At L3, the agent doesn’t even need to run on your machine. It can live on a server, triggered by a webhook or a schedule, processing work while you sleep.

Most companies calling themselves “AI adopted” are at L1. HBR observed that the best AI users treat it as a reasoning partner, not a productivity shortcut. That’s L3 behavior, and it’s learnable.

How I Got Here Without a Plan

One of the reasons I started this blog was to keep track of things as I learn them. Looking back at the past three months, it turns out I accidentally documented the L0-to-L3 transition in real time.

L1: I set up Claude Desktop to search hotels on booking.com and schedule daily price checks. It took actions on my behalf, but I was watching every step. Still felt like chat with extra features.

L2: I gave Gemini CLI access to git, k3s, and helm on a VM and said: “set up a fun homelab environment.” It decided what to deploy and I came back to a running cluster. I steered the intent, the agent handled execution. That validated what I had been suspecting: there’s so much more to this than just chatting.

L2-L3: I revived a 4-year-old 3D printer entirely through Claude Code. It SSH’d into the Pi, diagnosed a WPA3 WiFi issue, migrated to MainsailOS, flashed firmware, configured a touchscreen. It hit errors, recovered, and kept going.

L3: I built a multi-agent pipeline that takes a text description of a 3D object and generates CAD code using build123d. A second agent renders the model and checks proportions. A third evaluates printability: overhangs, wall thickness, support requirements. No human in the loop between steps. Text in, validated physical object out.

“Design a low-poly bunny for Easter.” The agent asks questions, generates the CAD code, validates geometry, and the Ender 3 prints it.

Each step required trusting the agent with more. And every time I did, my idea of what “using AI” means shifted permanently.

The Enterprise Gap

Most companies are proud of their L1 rollout. Sequoia’s AI 50 report called it out: “AI graduated from an answer engine to an action engine”, but most organizations haven’t made that graduation.

Market spending data reflects the shift. The fastest-growing vendor in 2026 (nearly 40% growth) isn’t the one with the most users. It’s the one going deepest into agentic workflows. People don’t switch because of benchmarks. They switch because someone showed them what L2 or L3 looks like.

That social trigger matters. Microsoft’s Work Trend Index found that power users are 61% more likely to have heard from their CEO about AI’s importance. The aha moment is socially triggered. I’ve seen this firsthand talking to mechanical engineering friends. The moment I showed them an agent generating CAD code from a text description, the conversation changed completely.

The Uncomfortable Part

More access means more attack surface. Companies pushing agentic AI are giving employees one-click access to dozens of connected systems. I spend my day job evaluating exactly this kind of exposure. I’ve written about AI operating at machine speed without guardrails and about monitoring what CI/CD pipelines actually access. Every integration that makes the aha moment easier also expands what a compromised agent can reach.

Not a reason to avoid agentic AI. A reason to be deliberate about it.

What Changes When the Cost Drops

None of these technologies are fundamentally new. Developers have had browser automation, structured data generation, and scheduled jobs for decades. The reason most problems never got solved wasn’t technical. It was economic. The cost simply didn’t justify the effort. I’ve been thinking about this since inference started approaching zero cost.

That barrier is collapsing. Getting a professionally engineered 3D model used to be expensive. I built a rough prototype by describing a shape in plain text and having an agent generate the CAD code. Not production-ready yet. But projects like Blueprint, which lets you design hardware prototypes by chatting with AI, show where this is heading. The models will get better. Today is the worst they will ever be.

The long tail of software, the problems too niche for anyone to build a product around, is about to get automated by the people who have the problems.

Pick one thing you do repeatedly. Describe the outcome to an agent. Let go of the keyboard. Your mental model will be different afterward.