Last week Anthropic published economic research on Claude Code, and the interesting part isn't how fast usage is growing; it's one finding: in agentic coding, the winner isn't the person who can write code, it's the person who understands the problem. We deploy AI agents for clients every week, so we read it closely and checked it against what we see on real projects. Here's a short breakdown and our own take, minus the press-release summary.
What the research found
Anthropic analyzed around 400,000 working sessions of Claude Code from about 235,000 people between October 2025 and April 2026. This isn't an opinion survey; it's a look at how people actually use the tool.
The first thing that stands out is how the work splits. People make roughly 70% of the "what to do" decisions, and the model makes about 80% of the "how to build it" ones. A typical session runs around four prompts, each one fanning out into a dozen model actions, sometimes past a hundred. The intent and the architecture stay with the human; the model takes the implementation grind.
So far this reads like the usual "AI assistant" story. The next part is where it gets interesting.
The core finding: expertise beats job title
Anthropic doesn't look at whether you're a software engineer. It looks at how well you understand the specific task in front of you, and on that axis the gap is wide:
- For novices in a task, verified success sits at 15% of sessions. For people who know the domain, it's 28–33%.
- Experts recover from errors at roughly four times the rate of novices.
- When a session goes sideways, novices abandon it 19% of the time; everyone else, 5–7%.
The most telling number for us: on coding tasks, every major occupation performs within 7 percentage points of software engineers. An analyst, a finance person, a marketer solves the task almost like a staff developer, as long as they understand what they're building. A coding background stopped being the entry ticket.
One more detail that lines up exactly with our experience: the better someone understands the task, the more work the model does per instruction. An expert triggers about 12 actions per prompt, a novice 5. The model doesn't guess what you want; it amplifies whoever can state it precisely.
What it means for non-technical teams
The takeaway we're keeping, and passing to clients: value has shifted from "who can write code" to "who understands the problem well enough to frame it right."
It used to be that automating a slice of work needed a developer in the middle: you explain the process, they translate it into code. Now the person who knows the process from the inside, their funnel, their reporting, their inbound leads, can build a working tool almost without that middleman. The bottleneck moved from "write the code" to "describe clearly what should happen."
This is exactly what we wrote about in why AI doesn't replace engineers: the model removes the grunt work, not the need to understand the task. That understanding just became the main asset instead of a nice-to-have.
Our view: what changes in hiring and working with a vendor
For a team, it changes what you pay for. Hire for domain knowledge and clear thinking, not just a stack on a résumé. Someone who deeply understands your field and can state a task cleanly will pull more out of AI tools than a nominal "techie" who doesn't get the business.
For anyone bringing in a vendor, the mirror image holds: the payoff from agentic coding scales directly with how clearly the problem is described. That's why on our projects we spend half the time not on code but on framing: mapping the process, finding where it actually hurts, agreeing on what "done" means. Across 100+ AI consultations we saw the same pattern: the projects that won weren't the ones with the cleverest model, they were the ones where the client knew exactly what result they needed.
In short: the tool amplifies expertise. Where there's none, it amplifies chaos.
Where the research has limits
So this doesn't read as a Claude Code ad, here's the honest part about its limits. Anthropic didn't measure what happened to the code in real life: tests passed, so it counted as success, but "passed the tests" and "delivered business value" aren't the same thing. On top of that, the whole classification rests on a model reading transcripts, which isn't a perfectly accurate method. So treat the findings as direction, not exact figures. The direction matches what we see in the field, which is why we take it seriously.
TL;DR
- Success with AI tools comes from understanding the problem, not a coding background.
- Non-developers trail engineers by only ~7 points; domain beats stack.
- The sharper the task is framed, the more the model does on its own.
- The bottleneck moved to framing and verifying the result.
FAQ
So engineers aren't needed anymore? They are. The research says something else: the model takes implementation, but not architecture, problem understanding, or verifying the result. Engineers matter most where the task is non-trivial and the cost of a mistake is high. More in our post on why AI doesn't replace engineers.
Do you need to know how to code to use AI agents? Per Anthropic, not really: non-developers perform within 7 points of engineers. Understanding your domain and stating tasks clearly matters far more.
Where should a non-technical team start? With a narrow, well-understood process you know cold, and a clear definition of the result, not with "let's put AI everywhere." If you want help with the framing, that's our AI implementation services.
How much time do people actually spend in Claude Code? Per the research, about 20 hours a week on average. That's no longer a weekend experiment; it's a primary tool.
If you want to map out where AI agents will pay off for your business and where they'd just burn money, get in touch.