Basic, Agentic, Advanced: the 3 levels of AI for software teams
Most software teams already use AI. The real question is which level. A field guide to Basic, Agentic, and Advanced AI, and what each one actually changes.
Almost every engineering team uses AI now. Very few have changed how they ship because of it. That gap, between “we use AI” and “AI changed how we work”, is the whole story, and it has three distinct levels.
Here is the field guide we use, and how to tell which one you are on.
Level 1: Basic AI (chatbot-assisted)
A developer pastes a function into a chatbot, gets a suggestion, pastes it back. A PM asks a model to tidy a spec. This is where most teams live, and it is genuinely useful: it speeds up isolated subtasks.
What it does not touch is the part of delivery that actually eats the week: triaging reports, writing tickets, reviewing PRs, chasing approvals, shipping. All of that stays manual. The ceiling is low for a simple reason. A person is still the bus carrying every piece of context from chat to tracker to codebase to CI to review. The model made one seat on the bus faster. The route is unchanged.
About 70% of organisations now use generative AI in at least one function (Stanford HAI, AI Index 2026). Most of them stop here.
Level 2: Agentic AI (a coding agent you drive)
Now a coding agent (Claude Code, Codex, Cursor) does real work. It reads the repo, writes across several files, runs commands. This is a genuine step up, and the demos are impressive.
The catch is in the word “you drive.” You still write each prompt, read every diff, run the tests, open the pull request, merge it, and trigger the deploy. The agent got dramatically faster at typing code. You became the bottleneck at everything around the typing: the hand-offs between tools, which is where the hours were hiding in the first place.
Roughly 23% of companies are scaling agentic AI somewhere in the business, but adoption stays in the single digits within any one function (McKinsey, State of AI; Deloitte’s 2026 survey lands on the same figure). Babysitting does not scale.
Level 3: Advanced AI (the hand-offs automated)
At the third level the connective tissue itself is automated. A request in chat becomes its own branch, its own PR, an automated review, and a guarded merge, with a person approving, not operating. The work that used to be five manual hops between five tools becomes one chat command and one human decision.
This is the level almost nobody reaches: only about 1% of companies describe their AI rollout as “mature”, the point where AI fundamentally changes how work is done (McKinsey, 2025). It is rare because it is hard, and it is hard in a specific way: it only works with real guardrails. Every check green with no bypasses, a human who owns the merge button, and shared state that cannot silently corrupt itself under parallel load. Skip those and automating the hand-offs just automates the mistakes faster.
We run this level on our own systems: 452 AI-built pull requests and counting, measured, in production.
Where do most teams actually sit?
It is a steep funnel:
| Level | Roughly how many companies |
|---|---|
| Basic (gen-AI assist) | ~70% |
| Agentic (scaling AI agents) | ~23% |
| Advanced (“mature”) | ~1% |
(Sources: Stanford HAI 2026, McKinsey 2025–26, Deloitte 2026, the same figures cited above.) The honest read: this is not a question of AI versus no AI. Nearly everyone is on the ladder. The question is how many rungs of recovered time you are leaving above you.
How do you move up a level?
The jump that matters is from Level 2 to Level 3, and it is not about a better model. It is about removing yourself from the hand-offs:
- Measure where the week goes. Time one real workflow (bug intake, review, release) end to end. You cannot recover hours you have not counted.
- Pick one workflow, not ten. Depth beats breadth. One workflow fully automated is worth more than ten half-automated.
- Automate the hand-off, not just the typing. The win is the agent opening the PR, the review running itself, the merge being one guarded click, not a faster autocomplete.
- Put the guardrails in first. The reason Level 3 is rare is that teams try to skip this. Don’t.
You can see the same workflow at all three levels (and the hours each one leaves behind) in our interactive comparison. Those numbers are illustrative; the proof from our own systems is measured.
If you want to find which level you are on and what moving up a rung is worth, book a strategy call.