Where AI needs to be for a small business to use it
AI agents are impressive. But your business information isn't on a computer. And that changes everything.
If you've used ChatGPT, you know how it works: you paste in some text, it writes a response. It's useful for a lot of things. But there's an obvious limit: it only knows what you give it, and the only thing it can do is write.
AI agents are a different thing. They differ from ChatGPT in two fundamental ways: they have access to your data and tools to work with it. You don't paste text — they read your files, run programs, check results, and correct. They work in a loop: do, check, adjust. For as long as it takes.
Claude Code (Anthropic), Codex (OpenAI), OpenClaw (open source) — the latter reached 247,000 GitHub stars in 60 days and works through WhatsApp or Telegram: you message it like a coworker and it operates on your computer. You talk to them in natural language, like a person, and they do the work.
What they do is genuinely impressive. Not demo-impressive: impressive as in "this just did in twenty minutes what would've taken me half a day." We use them every day. A large part of our code, this blog, our product documents — built with agents.
Can I use one for my business?
Yes — with nuances worth understanding.
We have our financial plan, for example, in a code repository: text files with costs, projections, funding sources. We tell the agent "update the projections with these new numbers" and it does — reads the files, understands the structure, modifies what needs changing. Works very well. But to share those documents between partners we use git, a version control system that programmers use. Not exactly the workflow of a carpentry shop or an architecture firm.
An agent like Claude Code or OpenClaw could, in principle, help you draft quotes, organize client data, or prepare documents for your accountant. But it needs access to the information — and that's where things get complicated.
Reading your email. For the agent to process invoices arriving by email, it needs to be connected to your account. There are several ways to do it — give it direct access, connect it through Google, set up forwarding rules. Each option requires configuring permissions, credentials that expire periodically, filters so it doesn't process spam. If you also want it to send emails — you need your own domain, anti-spam configuration, a signature. Even with tutorials available, it's a project.
Reading your WhatsApp. Harder still. Personal WhatsApp doesn't allow external programs to read your messages. The Business version does, but it requires business verification and an authorized provider. OpenClaw solves this differently — by being a messaging app itself — but it still operates on your computer, not connected to your bank or your accountant.
Accessing your bank. In Europe, this is regulated by European law (PSD2). You need a licensed intermediary — you can't just "connect." There are providers that offer this, but setting them up requires contracts, permissions, and maintenance.
Each integration is viable. It can be done, and it's worth exploring if you're up for it. But each one is a project with its own permissions, quirks, and edge cases. And it needs upkeep — services change their terms, credentials expire, providers update.
Why infrastructure matters more than it seems
Because it determines whether the agent works on its own or needs you to feed it.
A supplier sends you an invoice via WhatsApp. If the agent isn't connected to WhatsApp, you have to download the photo, go to the computer, open the tool, give it the image. Then check that it processed it correctly. You've saved some work, but you're still the manual link between your business and the AI. Multiply that by every invoice, every email from your accountant, every bank transaction. The promise of autonomy becomes just another task.
We've spent months building this plumbing. Connecting email required building a system to receive messages automatically, creating dedicated email addresses per company, and making the system group messages from the same conversation together. Connecting WhatsApp required building a system that can look at a photo of an invoice and extract the data. Connecting banking, a regulated provider and automatic transaction syncing. Each one took weeks.
And it still needs adjusting. Our Google Drive integration needed six separate fixes over five days to reach 80% reliability. A single invoice import takes two to three minutes, four rounds of AI reasoning, and a full process of reading data, detecting taxes, and verifying results. That's what "infrastructure" actually means when we talk about AI for small businesses.
It's not the model. The model is extremely powerful — it's the same one behind ChatGPT. It's the plumbing that connects that model to the places where your business information actually lives.
Try an agent. Seriously.
That said: if you can, try one. Try Claude Code, try OpenClaw, spend a few hours exploring. What you can do today with an agent that has access to your machine was science fiction two years ago. If you have someone on your team with a technical profile, have them spend a couple of afternoons on it.
And even if you can't set up integrations, structure your business so an agent can work with it: centralize documents, use digital tools, keep your data accessible. That'll pay off regardless of which tool you end up using.
But being where you work is only the first step. Having AI on your WhatsApp doesn't help much if it knows nothing about your business. That's what the next article is about.