How I use AI agents to run my business

I used to spend days each quarter on bookkeeping. Now it takes half an hour a month. Here is what I actually built with AI agents, and why it matters for more than tech companies.

I used to average a few days per quarter on bookkeeping. Invoices sat in a pile, bank transactions stacked up, and I was always missing a receipt or two. That is not why I went independent. But it was the reality.

Today I spend half an hour a month. My AI agents handle the rest.

Here is what I have actually built for myself, what it does for me day to day, and why I think it matters for far more companies than the ones working in tech.

My setup in broad strokes

I have a personal AI agent named Nova. She runs (yes, I call it "she") on a dedicated Mac Mini with her own Google account and Apple account. She has access to my online AI tools. I talk to her through Telegram or WhatsApp.

Nova is the link between me and a set of specialized systems I built in Claude Code. She can look things up in my accounts, pull data from my CRM, and publish content on my website. All through conversation or text.

Accounting that runs itself

The first thing I built was an accounting program. I call it RegnskabsErik. It is basically a database hooked up to software that identifies receipts, matches them against bank transactions, and books them automatically.

The system runs on a probability model. If amount, date, and description match above 85%, the receipt gets booked without me touching anything. That covers most cases. When the bank description doesn't match the invoice, usually because it shows a different company name or a code, the receipt lands on a list I approve manually in the app.

I could turn off the manual approval, but I like having the final check. It takes five minutes.

Nova is connected to RegnskabsErik. If I photograph a receipt and send it to her on Telegram with the message "cash payment, company books", she works out where it belongs and books it.

My agent also goes through my company email. If an invoice is attached, she downloads it and sends it to the database. From there it follows the same rules as everything else. According to Karbon, 46% of American accountants already use AI daily, and AI bookkeeping tools cut manual work by up to 80%. My own numbers look similar. I went from days to minutes.

The system also generates an updated income statement, so I can always see VAT status, salary, revenue, and the bottom line. An automatic annual report is the next step. Realistically it is one prompt away.

A CRM that collects all communication

The next system is a CRM. All clients, contracts, documents, and correspondence in one place, built for me and the way I work.

Twice a day, morning and afternoon, Nova scans my email. If a mail is from an existing client, it gets pulled into the CRM with a reference to that company and contact person. I have the full history in one place without copying anything by hand. It happens 100% automatically.

The CRM also handles the money side. It generates invoices and sends them to the client when the invoice date comes up. Fully automatic.

When someone contacts me through my website, their details go into the CRM too, with a note about where they came from and what they wrote. I could attach an email flow if that made sense. I haven't yet, but the option is there.

I can also ask Nova to research a contact. Who should I talk to at a given company, what is their area of responsibility, what type of person are they? That gives me a head start before the first conversation.

Content and trends

I have an analysis agent that delivers a report every morning on what is moving on Reddit and X. What people are talking about, what is trending, what is new. It can also pull the last thirty days of trends and show me where the attention is.

I use that to pick topics. If something is relevant, I write a blog post or a LinkedIn update about it.

For research I use Perplexity, a deep research agent that finds and verifies data from credible sources. Not blog posts from random websites, but official reports and studies. You can always argue about what counts as credible, but I have set it up to prioritize primary sources.

And then it writes. I get a blog post that is 85-95% done. I adjust the last few percent myself. What used to take hours now takes ten to fifteen minutes.

My blog posts live in an Obsidian vault. When a post is approved, I tell the agent to publish it. It goes up via GitHub and Vercel, and I touch nothing. One message, and it is live.

Email that sorts itself

A simple feature, but it saves a surprising amount of time. Nova goes through my email and structures it. If something needs an answer right now, I get a message on Telegram. Spam gets unsubscribed or deleted. The rest gets sorted.

The barrier is not the technology

According to the U.S. Chamber of Commerce, 58% of small businesses already use generative AI, if you count a ChatGPT subscription. Gartner expects 40% of enterprise applications to have built-in AI agents by the end of 2026. The market for AI in accounting alone is expected to reach 10.87 billion dollars in 2026.

The tools are available now, not at some point in the future.

But the biggest barrier is not technical. Microsoft's CEO has said as much himself. AI agents require a new way of working. Companies have to look at their existing processes and decide to do them differently. Not because the old way was bad, but because there is a path that saves fifty to eighty percent of the time.

Most companies know AI matters. They just don't know how to approach it. People are buried in operations, trying to keep up, and when someone adds "AI insights" on top of all that, nothing happens. One more item on the list.

That approach fails. It takes time and resources set aside to find out what can actually be automated. Not as a side project. As a real priority.

I know because I have tried it myself. Some things failed. Others paid off. The difference was that I sat down and tested properly instead of just talking about it.

What it means in practice

I run a business where the books nearly keep themselves. Where client communication collects itself. Where content gets researched, written, and published with little effort from me. Where my email is sorted before I open the inbox.

It is not magic. It is systems I built over time, tested, and adjusted. Each one saves me hours every week. I doubt I would have had time to build any of it without the agents freeing up hours as I went.

If you are in a company wondering where to start: look at the tasks that take the most time and follow fixed rules. Bookkeeping, invoicing, email handling, data collection. That is where the first win is.

And more important: don't wait for someone else to figure it out for you.