AI Agents for Small Business: 12 Use Cases, DIY or Hire

The agents that pay off in a small business do one narrow, repeated job: triage the inbox, draft quotes, chase invoices, summarize calls. Each example below lists the trigger, the tools it needs, and whether you can build it yourself. We also count its features, so you can see roughly what it would cost to have built.

Disclosure: Claudekit sells custom AI agents, priced by feature count. We use that same count below because it’s a handy way to size any agent, whoever builds it. Our tier ranges, read from our Hire page on October 25, 2026: Starter covers 1–3 features ($790, or $490 for the first 10 clients), Business 4–7 ($1,900 or $1,190), Pro 8–12 ($4,500 or $2,900). We make no claims below about time or money any agent will save you; that depends on your volume, and we haven’t measured it for your business.

What counts as an agent here

An AI agent reads something (an email, an order, a transcript), decides what it means, and then acts or drafts an action, using your tools. Anthropic’s article on building effective agents separates workflows (“LLMs and tools are orchestrated through predefined code paths”) from agents (“LLMs dynamically direct their own processes and tool usage”). Most small-business wins are workflows with one judgment step in the middle. That’s fine: simpler is cheaper to run and easier to trust.

A feature below is one concrete job, connection included. “Reads new orders from Shopify” is one feature; connecting to Shopify is part of it.

DIY level means:

• No code: you can do it in the Claude apps, for example with a Claude Project or Claude Desktop with an MCP server.
• Some code: one script or small agent you can build with Claude Code if you can read code. Our step-by-step guide to building an AI agent with Claude shows how.
• Hire: several systems, writes to live data, or needs tests and a handover you can trust.

The 12 use cases at a glance

#	Use case	Trigger	Features	Size	DIY level
1	Inbox triage	New email	3	Starter	Some code
2	Quote drafting	New quote request	4	Business	Hire, or some code
3	Invoice chasing	Daily schedule	3	Starter	Some code
4	Call and meeting summaries	New transcript	3	Starter	No code for one-offs
5	E-commerce order desk	New order, daily schedule	5	Business	Hire
6	Customer email answers	New support email	3	Starter	No code for drafts
7	Morning sales report	Daily schedule	3	Starter	Some code
8	Lead qualification	New lead	5	Business	Hire, or some code
9	Social posting	New product or post	3	Starter	Some code
10	Receipt and bookkeeping prep	New receipt email	4	Business	Some code
11	Ask your own database	You ask Claude	3	Starter	Some code
12	Dev team kit	Commit, pull request, release	3	Starter	Some code

The use cases in detail

1. Inbox triage

• Trigger: a new email in a shared inbox.
• Tools: Gmail or Outlook, Slack.
• Features: (1) reads new emails and labels each by type, (2) drafts a reply from your templates for routine types, (3) posts urgent ones to a Slack channel.
• DIY: some code. Start with labels and drafts only; let it send nothing until you trust the labels.

2. Quote drafting

• Trigger: a quote request from your website form.
• Tools: your form, a price list in Google Sheets, your document tool, email.
• Features: (1) reads the request, (2) looks up prices in the sheet, (3) writes the quote document, (4) emails it to you for approval.
• DIY: some code if your price list is clean. Hire if prices depend on rules nobody has written down, because the build is mostly writing those rules.

3. Invoice chasing

• Trigger: every morning.
• Tools: your accounting tool, email, a log sheet.
• Features: (1) reads unpaid invoices, (2) writes a reminder whose tone depends on how overdue it is, (3) logs each reminder in a sheet.
• DIY: some code. Keep a person approving the first weeks of reminders; tone mistakes with customers are expensive.

4. Call and meeting summaries

• Trigger: a new transcript from your call or meeting tool.
• Tools: the transcript source, your task tool, your CRM.
• Features: (1) summarizes the transcript, (2) adds action items to your task tool, (3) adds a note to the customer’s CRM record.
• DIY: no code for one-off summaries: paste the transcript into a Claude Project that holds your summary template. Automating it across every call needs some code.

5. E-commerce order desk

• Trigger: new orders, plus a daily schedule.
• Tools: Shopify, Google Sheets, email, Facebook.
• Features: (1) reads new orders from Shopify, (2) looks up stock in Google Sheets, (3) answers customer emails from a template, (4) writes a sales report every morning, (5) posts new products to Facebook.
• DIY: hire. Five features across four systems is where testing and error handling start to matter more than the prompt. This is the business automation example on our Hire page.

6. Customer email answers

• Trigger: a new support email.
• Tools: your inbox, a help document or FAQ.
• Features: (1) reads incoming support emails, (2) drafts an answer from your help content, (3) sends anything it isn’t sure about to a person.
• DIY: no code for drafts: a Claude Project with your help docs gives your team consistent answers to paste. Answering customers automatically needs some code and a clear rule for when to hand over.

7. Morning sales report

• Trigger: every morning.
• Tools: your store or payment tool, an ad-spend sheet, email.
• Features: (1) reads yesterday’s sales, (2) reads ad spend from the sheet, (3) emails a short report with what changed.
• DIY: some code, and a good first project: it only reads data, so a mistake costs a wrong email, not a wrong refund.

8. Lead qualification

• Trigger: a new lead from a form or your CRM.
• Tools: the form, the lead’s public website, your CRM, email.
• Features: (1) reads the new lead, (2) reads the company’s public website, (3) scores it against your written criteria, (4) updates the CRM, (5) drafts a first email for a salesperson.
• DIY: some code if your criteria fit on one page; hire if the CRM write-back is complex. Read only public pages a person could visit; scraping that breaks a site’s terms is off the table for any builder.

9. Social posting

• Trigger: a new product or blog post.
• Tools: your site, Facebook, LinkedIn.
• Features: (1) drafts posts from the new item, (2) posts to Facebook, (3) posts to LinkedIn.
• DIY: some code. Keep a person approving posts; a brand voice slip is public.

10. Receipt and bookkeeping prep

• Trigger: a receipt or bill arrives by email.
• Tools: email, a spreadsheet or your accounting tool.
• Features: (1) reads receipt emails and attachments, (2) extracts date, supplier, amount and tax, (3) adds a row to the sheet, (4) flags likely duplicates.
• DIY: some code. It prepares data for your bookkeeper; it doesn’t replace them.

11. Ask your own database

• Trigger: you ask Claude a question in Claude Desktop or Claude Code.
• Tools: an MCP server in front of your database or internal API, read-only.
• Features: (1) looks up an order by number, (2) lists today’s unshipped orders, (3) finds a customer by email.
• DIY: some code. An MCP server is a small program; our guide on what MCP is explains the idea. Keep it read-only until you have a reason not to.

12. Dev team kit

• Trigger: commits, pull requests and releases.
• Tools: Claude Code, your repo.
• Features: (1) reviews code against your style guide, (2) writes release notes from your git log, (3) blocks commits when the tests fail.
• DIY: some code, and you can start free: our tested kits for PR review (reviews your branch diff) and release notes (drafts a changelog from git) are close to the first two.

Signs you don’t need an agent

Before you build or buy, check these. Each one is a reason to wait or to pick something simpler.

• The steps never change. “Form submitted, add a row, send an email” needs an automation tool, not an AI model deciding anything.
• Your process isn’t stable. If the way you quote or answer customers changes every month, you’ll pay to rebuild the agent every month. Run it by hand until it settles.
• The volume is tiny. Five emails a week are faster to answer yourself than to specify, test and maintain.
• One person does it in Claude already. If a Claude Project with good instructions gets the job done in a few minutes, that is your agent.
• Nobody can say what “right” looks like. If you can’t give an example input and the result you expect, no builder can test it, and neither can you.
• It touches regulated data. Health records and card data come with compliance rules that a small agent build isn’t the place to learn. We refuse these jobs, and you should be wary of anyone who doesn’t ask.

How to pick your first agent

1. Write down three repeated jobs that take someone’s time every week.
2. For each, write one example input and the result you expect.
3. Pick the one that only reads data or only writes drafts. Low risk first.
4. Count its features: one concrete job per line, connection included.
5. Estimate the running cost with the API cost calculator: runs per month, tokens per run, model.
6. Decide: build it yourself, hire a freelancer, or use a fixed-price service. Our guide to hiring an AI agent developer has questions to ask any of them.

For build and running costs with sources, read AI agent development cost. If you’d like us to build it, send the feature list on our Hire page.