Claude Code agents for small business: a worked example

What is a Claude Code agent?

An agent is a job that runs without you at the keyboard. It has written instructions, access to only the data it needs, a schedule, and a report at the end that a person reads. Claude Code is how you build it and, for some agents, what runs it. The goal is that you start the day with a short list of what needs attention instead of digging for it.

What does our Hey Bud inventory agent do?

Hey Bud Skincare is our own skincare brand, sold online through Shopify and through retailers. It is ours, not a client. Its inventory system does three things before the working day starts:

• A scheduled job on a small server refreshes stock data and takes a snapshot.
• The inventory platform compares stock against sales and forecasts, and marks products that need reordering.
• A daily brief, written by Claude, flags critical stock, retail orders waiting to be picked and supplier bills falling due that week.

What Claude Code did. Ollie Watts, who runs marketing across our businesses and is not a developer, built the whole platform with Claude Code: the stock, orders, shipments and forecasting screens, the scheduled jobs and the daily brief. It wasn’t one big build. Features were added one request at a time as the business needed them, and Claude Code is still how it gets changed.

How does a scheduled agent work, step by step?

Every agent we build follows the same eight steps. None of them needs you to write code; they need you to be precise about the job.

1. Write the job down. What to check, what counts as a problem, and who needs to know. “Products with less than two weeks of stock at current sales” beats “check inventory”.
2. Give read-only access to the data. For Shopify, a custom app with read scopes only. See connecting Claude Code to Xero, Shopify and Gmail.
3. Put the instructions in a file. A CLAUDE.md or instruction file with the checks and an example of the report you want back.
4. Choose where it runs. A small cloud server or a computer that stays on. Set its time zone to Australia/Melbourne (or yours) so daylight saving doesn’t shift the run by an hour.
5. Fix its permissions in advance. Nobody is there to approve steps, so list exactly what it may do and deny everything else.
6. Send the result where you’ll read it. Email, a Slack channel or a dashboard.
7. Alert on failure. A job that silently didn’t run looks the same as a quiet morning. Have a separate check that tells you when no report arrived.
8. Watch it for two weeks. Read every report against reality before you rely on it.

Should the agent be a script or Claude Code itself?

There are two ways to build it, and the right one depends on how much judgement each run needs.

	Claude Code writes a scheduled script	Claude Code runs on a schedule
How it works	A normal program runs each morning and calls Claude only for the part that needs judgement, such as writing the brief	A scheduler starts Claude Code in headless mode, and Claude follows the instructions file each run
Best for	Repeatable checks on structured data	Jobs that vary, or that need Claude to investigate
Cost per run	Lower	Higher, because Claude does more reading
Example	Our Hey Bud inventory system	A morning review of new enquiries that drafts replies for a person to send

For the second approach, a run is one command. This illustration runs a morning stock check with tight limits:

claude -p "Run the morning stock check described in CLAUDE.md and save the report to report.md" \
  --permission-mode dontAsk \
  --allowedTools "Read" "Bash(python fetch_stock.py)" "Edit(./report.md)" \
  --max-turns 20 \
  --max-budget-usd 1.00

dontAsk refuses anything not on the allowed list, --max-turns stops a run that goes in circles, and --max-budget-usd caps what one run can spend. A scheduler such as cron then starts it each morning, for example 0 6 * * * for 6am. The flags are documented in Anthropic’s CLI reference.

What does a scheduled agent cost to run?

In our experience the Claude API cost of a typical small-business agent is $3–15 a month, depending on how much it reads and how often it runs. Add a small server if you don’t already have a computer that stays on. Each of our agents has its own API key and a monthly spend limit, so a mistake costs dollars, not hundreds. Our Claude Code cost guide works through a modelled example in AUD.

What can go wrong, and how do you guard against it?

Risk	Guardrail
The job silently doesn’t run	A separate alert when no report arrives
It reads the wrong data	Read-only access, and a person reads every report
It runs up a bill	Spend limit on the key, --max-budget-usd and --max-turns on each run
The report format drifts	An example report in the instructions file
Daylight saving shifts the time	Server time zone set to Australia/Melbourne (or yours)
Hidden instructions inside data, such as a product description	No write access, and allowed actions limited to the job

Our Claude Code safety guide covers permissions and secrets in more detail.

Does every automation need AI?

No. Our DotVA website lead pipeline is a useful contrast. Claude Code built it, including the spam and job-applicant filtering, the lead database and the dashboard, and the core went live in a single day in July 2026. But it runs as ordinary code on Cloudflare, scoring each enquiry with rules, because that job doesn’t need AI on every run. Use AI where the job needs judgement, and plain code everywhere else.

How On Autopilot can help

If you sell on Shopify, AI Inventory Watch is our productised version of this agent: $497 setup plus $99 a month, with the Claude API cost included. A Quick Start ($497 one-off) builds one automation of your choice and hands it over. If you want to learn to build and run agents yourself, the Claude Code Setup Day is $1,500 fixed and ends with one real automation running. If you’d rather not run anything, Managed AI builds and maintains your agents from $1,500 a month. All prices are ex GST; see the pricing page or book a free 30-minute AI audit.