Claude Code for business: 7 non-coding jobs we ran, and what still needed a person

Can non-developers use Claude Code?

Yes. Claude Code is Anthropic's agent that reads files, edits them and runs tasks. Developers use it in a terminal, but it also runs in the Claude desktop app (the Code tab), so there's nothing to install separately and no commands to learn. It is included in every paid Claude plan, from Pro at $20 a month (Anthropic's pricing page, 1–2 November 2026).

What we see people actually use it for is less dramatic than the demos. One in-house team that handed it to marketers described the useful set-up as a folder per client with a short context file, then narrow asks like search terms, tag changes and ad lines. Their words: You don't need to vibe-code anything for this.

How we ran the test

Claude Code in the desktop app on Windows, 2 November 2026. All files were made up for the test or public. No client or company data. Times are rough working times, not benchmarks. For each job we wrote down the request, what came back, and what a person had to do.

The 7 jobs

1. Clean up a messy lead list

Request: "Clean this lead list: fix names, emails, phone numbers and dates into one format, remove duplicates, and flag anything you're unsure about instead of guessing."

Result: seven rows in, six out. It removed one duplicate (the same person typed two ways), put five different date formats into one, standardised phone numbers, and flagged three rows for a person: one invalid email, one missing phone, one missing name. Sample data

What needed a person: a name written as "Smith, Ana" slipped through the first pass and needed one extra rule. And someone has to decide what to do with the flagged rows.

2. Turn a weekly export into a manager summary

Request: "Summarise last week for a manager in five bullet points, flag anything unusual, and show the totals you used."

Result: five bullets plus a small table by channel. The totals (69 leads, 16 meetings, 5 deals, $21,000) matched the correct answer exactly. It also flagged that one channel brought eight leads on a day with no meetings.

What needed a person: deciding what counts as "unusual" for your business. Why the numbers were right: Claude Code calculated them with a small script and showed them, rather than "remembering" them. Asking for the numbers used is the habit that makes AI reports trustworthy.

3. Research a company before a call

Request: "Research HubSpot for a sales call tomorrow: what they sell, who to, size, one recent change, two good questions. Link your sources."

Result: a one-page brief with links and two questions. Honest about conflicts: sources disagreed on the customer count, so the brief showed both figures instead of picking one.

What needed a person, and the biggest miss of the test: the brief stopped at news from June. When we gave the same request to Perplexity on the same day, it found changes from October and linked the company's own filings. For "what changed recently", use a search-first tool or check the company's own news page. How four assistants compared on this exact task.

4. Check your website's pages for SEO basics

Request: "Check every page of our website for title length, description length, one H1, image alt text, word count and internal links. Give me a table."

Result: a table for all 23 pages of our own site, with three pages flagged as thin.

What needed a person: judging which flags matter. A "page not found" page is meant to be short; a contact page being short is fine; the insights page being thin was a fair point (this article is part of the fix).

5. Update a set of process documents

Request: "In these three process documents, replace our old client spreadsheet with our new CRM, and add a step: the account lead reviews before the client email goes out."

Result: all three documents updated in seconds, with the new step in the right place.

What needed a person, and a real mistake: a blind find-and-replace left the line "Tool: Google Sheet 'Client Hub (HubSpot)'", which no longer made sense because the tool isn't a spreadsheet any more. A quick read caught it. Bulk edits are where a review step earns its keep.

6. Turn a call transcript into posts

Request: "Turn this call transcript into five LinkedIn post drafts. Every draft must trace to a line in the transcript. No invented numbers."

Result: five drafts, each with the line it came from. For example, from the customer's if it showed me the numbers it used, so I can check them in ten seconds came a post about the ten-second trust test for any AI output. Illustrative transcript

What needed a person: tone, and, with a real customer, permission to quote them.

7. Build a small internal tool

Request: "Make a simple page where I enter team size and billing type, and it shows what Claude Team, ChatGPT Business and Google Workspace would cost per month and per year."

Result: a working one-page calculator in a few minutes. We tested it in a browser: 10 seats on annual billing at $20 each came to $200 a month and $2,400 a year, as it should.

What needed a person: the prices are typed in by hand and dated, so someone owns keeping them current.

What the test tells you

• It's fast and literal. Every job took minutes from a plain-English request.
• Numbers are reliable when it calculates and shows them. Always ask for the figures it used.
• Research is the weakest job. Treat it as a first draft with links, and check recent news yourself.
• Two in seven had a real error a quick read caught. That's not a reason to avoid it; it's the reason a person stays in the loop.

The set-up that makes it work

• One folder per client or job, with only the files that task needs.
• A short context file (Claude Code reads one called CLAUDE.md automatically): who you are, what the folder is for, the output format, what not to touch.
• Narrow requests, one job at a time, rather than "run my business".
• A person checks anything that goes to a customer or changes records.

For a team, add a shared folder and a hand-off file. Here's the full small-team set-up.

Keeping company files safe

The AI assistants disagree on this when you ask them. One told us a non-developer shouldn't use it on sensitive files on their own; another said it generally asks before acting. Our approach: keep approval prompts on, allow specific folders only, start with sample or non-sensitive files, and decide in advance which folders it must never touch. Anthropic's own documentation warns that its agents can take real actions on your behalf.

When Claude Code is overkill

If the job is "write five posts" with no files, data or repeat schedule, plain Claude chat or Cowork is simpler. Claude Code earns its place on repeatable jobs with files and numbers: weekly reports, clean-ups, bulk document updates and small internal tools.

Common questions