How Do I Use Claude Agents to Run Business Operations?

You use Claude agents to run operations one process at a time: each agent runs on a schedule through Claude Code or the Claude Agent SDK, reads your systems over MCP, follows written rules it loads on every run, and sends anything it isn't sure about to a named person's exception queue. Start with one high volume, rule heavy process such as invoice coding or pipeline research, run it in shadow for 4 weeks, and expect 6 to 8 weeks per process before it works unsupervised.

Why most companies get stuck in the chat window

Most companies I walk into bought Claude seats for 30 or 40 people and called it an AI rollout. Six months later the seats are used for emails and meeting notes. Nothing in operations has changed, because a chat window waits for a person to type. An operations agent has to start on its own at 6am, read the bank feed, and finish before anyone logs in.

The second mistake is the opposite. A keen engineer wires Claude into the ERP with full write access in a weekend demo. It works on the happy path. Then it posts 40 invoices to the wrong cost centre in one night because a supplier renamed itself, and the finance director switches the whole thing off. I've seen that kill a programme for a year.

Pick the Claude surface by who runs the agent

Claude comes in four forms that matter for operations, and the choice decides who maintains the agent. I pick by one question: who starts the agent, and where does it live?

The Agent SDK gives you the same tools and agent loop as Claude Code, so an agent that starts as a scheduled Claude Code job can move into your own code later without a rewrite of its rules.

The five steps I use inside a client

1. One process, one owner, one number

I choose the process by volume times rule density. Invoice coding at 800 a month with a written chart of accounts beats contract review at 12 a month with no playbook. The process gets a named owner, usually the person who does it today, and one number we'll judge it on, such as hours per week or error rate. I baseline that number for 4 weeks before the agent touches anything.

2. Connect over MCP, read only for the first month

Claude reaches your systems through MCP, the Model Context Protocol. Each connection is a door you choose: the accounting ledger, the CRM, the shared drive, the ticket queue. For the first 4 weeks every door is read only. The agent proposes, writes its answer to a review sheet, and a person compares it with what they did. I want 95% agreement over 2 straight weeks before any write access is switched on.

3. Write the rules into files, not into prompts

Claude Code and the Agent SDK load skills, commands and memory from a project folder on every run. That folder is the agent's operating manual: the chart of accounts, the approval limits, the supplier aliases, what goes to the exception queue. When the business changes, the owner edits one file and the next run picks it up. A rule that only lives in someone's chat history doesn't exist.

4. Permissions and hooks do the policing

The SDK lets you set which tools run on their own and which need approval, and hooks run your own code at fixed points in each run. I use a hook to block any payment, any email to a customer and any delete. The agent prepares a payment run, a named person releases it. That line doesn't move after 6 good months.

5. A budget gate before every run

Every agent I run checks a ledger before it starts: runs this week and tokens this week, against a cap. Over the cap, it stops and says so. It never raises its own limit. That one check is the difference between a predictable monthly cost and a surprise on the card statement.

From my operating seat

I'm Hayat Amin, an AI-native CFO, AI operator and AI coach for CFOs who teaches finance leaders how to build AI into finance operations. I spent twenty years in the C-suite, with three exits and three FT100 listings, and now I build and run Claude agents inside companies: finance, sales pipeline, onboarding and compliance evidence. I coach CFOs one to one on the same work. The starting point for that is AI for CFOs.

My own business runs this way. The article you're reading was researched, drafted and checked by a scheduled Claude Code run that read its rules from a file, passed a weekly run and token cap before it started, and was blocked from naming anything on my banned list. Inside one client I run, an agent researches the sales pipeline overnight and the team has a ranked account list by the morning. The person who owns it spends 20 to 30 minutes a morning on the exceptions. That's the whole human cost.

Do I need developers to build Claude agents for my business?

For the first agent, no. A scheduled Claude Code run with a rules file and two or three MCP connections needs someone who can read a log and edit a config file, not an engineering team. Once an agent has to sit inside your own product or run at volume on the Agent SDK, you want one engineer, part time, who owns the code and the credentials.

Are Claude agents worth it for a small business?

Yes, if there's a process with enough volume to measure. Below about 200 transactions a month the hours saved won't cover the time to set the agent up and check it. Above that, a 20 to 50 person company can run one or two agents with the founder or finance lead as the owner, about 2 hours a week each.

How do I stop Claude agents running up a large bill?

Put a cap in front of every agent before it starts, not a report after it finishes. I give each agent a weekly limit on runs and on tokens, and the agent checks the ledger and stops itself when it's over. Raising a cap is a decision for the person who owns the budget, never for the agent.

Where I come in

This is what I build and run inside companies. I pick the first process, wire the Claude agent to your systems with read access, write its rulebook with the person who owns the work, and hand it over once it has held its number for a month. Then we do the next one. Which process in your company would you hand to an agent first? See how I work as your AI agent operator, or start at meethayat.com.