Short answer: Anthropic analyzed about 400,000 Claude Code sessions from roughly 235,000 people and found that domain expertise, not coding ability, predicts who gets real results from AI agents. Non-software professionals reached verified success within a few points of software engineers, 26 versus 30 percent. What separated the people who succeeded from the people who quit was two learnable habits: giving precise instructions, and redirecting the agent instead of giving up when something broke.
Anthropic just published new research on how people actually use Claude Code, its agentic coding tool. The study looked at roughly 400,000 Claude Code sessions from about 235,000 people between October 2025 and April 2026. Most of the coverage will focus on software engineers. The more useful lesson for business owners is simpler: your industry expertise is now the scarce input. The agent can handle a lot of the implementation. It still needs someone who knows what the work is supposed to accomplish.
That matters because many owners are still waiting on the wrong thing. They are waiting to "learn enough code," hire a developer, or write a full project plan before they automate a task they already understand deeply.
Anthropic's research points in the opposite direction. The best starting point is not coding skill. It is a real business process you know cold.
Four findings are worth your attention.
1. Non-software users are only a few points behind software professionals
Anthropic measured "verified success," meaning the session ended with real evidence that the goal was met, such as passing tests, committed work, or explicit confirmation from the user. Software-related occupations reached verified success in about 30 percent of sessions. Other professions reached 26 percent. In code-producing sessions, the gap was similarly small: 34 percent for software-related occupations and 29 percent for everyone else.
On partial success, the gap nearly disappears. Software-related users reached at least partial success in 89 percent of code-producing sessions. Other professions reached 88 percent. Every one of the ten largest occupations in the study landed within seven percentage points of software and math users. Anthropic's own conclusion was blunt: "coding agents are making a coding background less relevant to successful programming."
For a non-technical owner, that is the permission slip. You do not need to become a software engineer before you can start using AI agents to build useful internal tools. The expensive move is waiting until you feel technical enough.
2. You decide what. The agent decides how.
Across the sessions Anthropic studied, people made about 70 percent of the planning decisions and only about 20 percent of the execution decisions. Put differently: users mostly decided what needed to happen, while Claude handled most of how it happened. Anthropic summarized the division of labor this way: "people decide what to build, and the agent decides how to build it."
That is exactly the kind of work split a business owner should want.
You do not need to know how to write a reconciliation script. You need to know which numbers have to match, which exceptions matter, what a clean output looks like, and when the answer is obviously wrong.
You do not need to know how to build a scheduling tool from scratch. You need to know how your crews are assigned, which customers cannot be moved, how travel time affects the day, and what "done right" means in your operation.
The tool is not asking you to become a programmer first. It is asking you to be precise about the work.
3. Expertise is about the task, not the title
This is the reframe most owners miss.
Anthropic notes that expertise is task-specific. A senior engineer asking about a language they have never used may be a beginner at that task. An accountant who has never written Python but can explain the exact reconciliation rules for two ledgers is an expert at that task.
That distinction matters.
You are not merely a "non-technical person." You are an expert in how invoices actually get paid, how jobs actually get scheduled, how customers actually behave, how inventory actually breaks, and where the spreadsheet always goes wrong at month-end.
Those are not side details. They are the instructions the agent needs.
Point the agent at work you understand deeply, and you are not starting from weakness. You are starting with the one thing the model cannot invent: real-world judgment about your business.
4. The gap between frustration and results is two habits, not talent
Anthropic found a clear difference between novice and expert behavior inside a session.
In novice-rated sessions, each prompt triggered about five Claude actions and roughly 600 words of output. In expert-rated sessions, each prompt triggered more than twice as many actions, about 12, and roughly 3,200 words of output. Experts got more work from the agent because they gave it clearer direction.
The bigger difference showed up when something went wrong. In troubled sessions, novice users abandoned the session 19 percent of the time. Everyone else abandoned it only 5 to 7 percent of the time.
That is not a talent gap. It is a behavior gap.
Experienced users do two things better. First, they frame the request with more context. Second, they check the output and redirect the agent when it drifts. We pulled both of these apart in why beginners fail with AI coding agents and the most important habit for getting results.
Both habits are learnable. Your first few attempts may feel clumsy. That is normal. It is not evidence that AI agents are only for technical people.
One more signal points in the same direction. Between October 2025 and April 2026, the share of Claude Code sessions spent fixing broken code fell from 33 percent to 19 percent. At the same time, usage shifted toward operating software, analyzing data, and writing documents, while Anthropic's estimate of the average session value rose by about 27 percent.
People are not just fixing bugs faster. As they get comfortable, they are bringing bigger and more valuable problems to the tool.
What to do next
You do not need a giant transformation plan. You do not need to hire a developer before you begin. You need one real task you understand well and a focused block of time to work through it.
Start with something specific. Not "improve my operations." Try "compare these two exports and flag invoices where the amount, customer name, or payment date does not match." Not "build me a dashboard." Try "turn this weekly sales export into a summary by location, salesperson, and product category, with anything more than 20 percent below average highlighted."
The research points to three practical shifts.
Start using AI agents on real work now
The occupation data is the important signal. People outside software occupations are already getting results close to software professionals. That does not mean every task will work on the first try. It does mean the old barrier is much lower than most owners assume.
The right first project is not the most ambitious one. It is the one where you can immediately tell whether the output is right. Pick a task where you already know the rules: reconcile a report, clean a spreadsheet, draft a follow-up sequence, sort customer notes into categories, or generate a weekly operating summary. Upselling your current clients is a strong first workflow. Your expertise is what makes the task safe to delegate.
Stop using AI only to plan
A lot of owners are still using AI like a brainstorming partner. That is useful, but it is only a small part of the opportunity. If you already live in ChatGPT, here is why the agentic tools are a different gear.
Anthropic's research shows that in typical Claude Code sessions, the agent handled most execution decisions. The user did not just ask for ideas. The user pointed the agent at work and let it take action. That is the shift. Do not stop at "make me a plan." Ask for the thing itself: the spreadsheet, the script, the cleaned file, the draft, the checklist, the report. Then inspect it like an owner.
Watch the output like the owner you are
The people getting results are not blindly trusting the agent. They are steering it. That is the job.
Check the result against a number you already know. Compare it to last week's report. Look for the weird customer case. Test it on the messy file, not the clean one. Tell the agent exactly what failed and run it again. This is where your business judgment becomes leverage. The agent can move fast, but you know what "wrong" looks like.
Five habits to use on your next AI-agent task
- Define "done right" before the agent starts. Name the goal, the output format, and how you will test the result in your first message.
- Lead with the knowledge only you have. Explain the exceptions, edge cases, customer rules, naming conventions, and failure points. The agent cannot guess how your business actually works.
- Verify every result against something real. Do not judge the output by whether it sounds confident. Check it against a known record, report, number, or example.
- When it breaks, redirect instead of walking away. Say what is wrong specifically. For example: "this matched invoice numbers but ignored payment date," or "this grouped cancelled jobs with completed jobs." Then run the next attempt.
- Keep it to one task at a time. Vague prompts produce shallow work. A single well-defined job gives the agent room to execute.
The takeaway from Anthropic's research is not that coding no longer matters. It is that coding is no longer the only gate. For business owners, the scarce skill is knowing the work well enough to direct it, test it, and recognize a useful result.
The expertise you already have is the scarce input. The tools just caught up to it.
Frequently asked questions
Do I need to know how to code to get real results with tools like Claude Code?
No. Anthropic's research shows non-technical users reach verified success rates within a few points of software professionals, 26 percent versus 30 percent, with nearly identical partial-success rates. Domain expertise in your specific task matters far more than coding ability.
What share of the work does the AI actually do?
In a typical session, people made about 70 percent of the planning decisions, the "what should we build and why," while the agent made about 80 percent of the execution decisions, the "how do we actually build it." You stay in control of direction and quality.
Why do some people succeed with these tools while others get frustrated and quit?
The biggest differences are not talent or coding skill. Successful users give precise instructions and verify the output instead of trusting it blindly. Novices abandoned troubled sessions 19 percent of the time; experienced users only 5 to 7 percent. Both habits are learnable. The full breakdown is in why beginners fail with AI coding agents.
What tasks should a non-technical business owner try first?
Start with work you already understand deeply: invoice reconciliation, client onboarding sequences, scheduling logic, compliance checks, or report generation. Any recurring process where you know exactly what "correct" looks like is a safe first project. Save open-ended exploration for later.
Is this only true for Claude Code, or does it apply to other AI coding agents?
The core finding, that domain expertise beats raw coding skill, is structural to how agentic tools work. Users who clearly define goals, exceptions, and success criteria get better results across current agentic systems. The specific numbers come from Claude Code data, but the principle is broader.
How long does it take to get good at using these tools?
Most of the gap closes quickly. Between October 2025 and April 2026, the share of sessions spent fixing broken code fell from 33 percent to 19 percent as people gained experience, and the average session value rose about 27 percent. Many owners see meaningful results from their first few well-structured sessions.
Should I still hire a developer if I can use AI agents myself?
Many routine and mid-complexity tasks no longer require a full-time developer to get started. You can validate ideas, build internal tools, and automate processes much faster. Developers become more valuable for high-stakes architecture, complex integrations, and work where deep technical judgment is still required.
What is the single highest-leverage habit I can adopt today?
Be specific in your first message, and always verify the output against something you already know is correct. Those two habits account for most of the difference between novice frustration and expert-level results. More on the first move in the most important habit for getting results from an AI agent.
Keep reading: why beginners fail with AI coding agents, the most important habit for getting results, turn your SOPs into AI agents, and the Test, Refine method.