I handed my office to an AI: an honest review
Updated on 21 November 2026.
From chat to agent: three steps
I have been doing IT for small businesses for more than twenty-five years, and I am retired. I keep a few clients and write five blogs, partly for fun. So it was with some scepticism that I opened, like everyone else, a chat window with an AI. Here is how I ended up handing it a good part of my office.
1. The chat. You ask a question, you get an answer, you copy and paste. That is already useful for drafting a difficult e-mail, summarising a document or explaining a concept. But every conversation starts from scratch: the AI knows nothing about you, your clients or your habits. You do all the linking work yourself.
2. Projects. Assistants offer workspaces ("Projects" at Claude, equivalents at ChatGPT and Gemini) where you store permanent instructions and documents: the tone of my blogs, the list of articles already published, the procedure for checking a source. The AI finally knows the context. But it stays locked in its window: for an article to reach my site, I still had to download, rename and copy.
3. The agent. This is the leap. I use Claude Code, Anthropic's "agent" tool, included in Claude's paid subscriptions. It works in a folder on my PC: it reads and writes files, runs commands and scripts, browses the web, and uses connectors to my mailbox, my Google statistics or an image generator installed on the network. I give it a goal ("an article for exsit.be on this topic"), it chains the steps and asks for my opinion at the important moments. I even follow it from my phone.
The model and the harness
A word of vocabulary, because it sheds light on the competition. The model (Claude, GPT, Gemini, Mistral, Qwen…) is the engine: it understands and writes. The harness is the car around it: the tools the AI can use, its memory, its permissions, its guardrails. Two agents using the same model can behave very differently depending on their harness. And a good harness can sometimes run several models.
What the agent does for us, in practice
No showroom demo: here is what has actually been running since September and October 2026.
| Task | What the agent does | What I keep |
|---|---|---|
| YouTube monitoring | A small program it wrote follows about thirty channels, fetches the subtitles of new videos, has them summarised by a cheap model and publishes a table on a private page of my site, every three hours | The choice of channels and topics |
| Blog articles | Suggests topics, searches and checks sources (legal texts, official pages), writes, has another model review, translates into English and Dutch, prepares the social media texts, generates the image, tests the site display locally | The topic, the title, the angle, the review, the go-ahead to publish |
| Book and user manual | Writes chapters from my notes, keeps the characters' "bible", produces an online PowerShell manual | The story, the tone, the choices |
| E-mails | Has its own assistant mailbox; analyses incoming messages, gives me a numbered summary every evening with a follow-up of open items; can send, always with me in copy | The decision to send, sensitive replies |
| Domain migration | Keeps the tracking table for moving my clients' domain names to a new host, checks the DNS zones | The operations at the registrars |
| Directory of providers | Cross-checks the Crossroads Bank for Enterprises with websites, sorts the relevant companies, prepares and sends contact e-mails in batches of ten | The approval of each batch |
| Budget | Reads bank statements in read-only mode and sorts them into categories with a model running on my own PC | The sorting rules, the interpretation |
And dozens of small things: a script that no longer works, a file to convert, a search through hundreds of documents.
What has really changed
I am doing again what I had given up. Articles in three languages, regular monitoring, a directory: I would never have had the time to do it all alone. The agent does not replace me, it gives me hours back.
Quality goes up, provided you demand it. Every figure in my articles is checked against the primary source, a second AI reviews, then I review. That is more rigorous than what I used to do alone, in a hurry.
Procedures finally get written. For an agent to work well, you have to write down the rules: how we publish, what we do not touch, where things are. Those documents are useful to me too. Many small businesses have never written their procedures: an agent forces them to.
Its mistakes: a few real examples
An honest review has to show them. All these cases are real and recent.
- The Quebec namesake. While sorting the directory, the agent took the website of a Quebec company for that of a Charleroi SRL with the same name. The recipient pointed it out to us, politely. Since then, a rule excludes non-Belgian sites without a clear clue.
- The unrequested publication. A publishing script asked "Publish now?"; run by the agent without a screen, the question got an automatic answer and the article went online. No harm done, but the script now only checks unless it receives the explicit option.
- The loan filed as bank charges. In the budget, a rule suggested by the local model put a monthly loan repayment under "bank charges". The total was wrong. I saw it because I review.
- The overconfident reviewer. The model that reviews my articles sometimes asserts errors with great confidence. On my latest legal article, it disputed four points that the official text confirmed. The agent checked them against the source and did not apply them.
- The invisible character. Just today, when adding a YouTube channel, the name was saved empty because of an invisible character added by Windows. The agent spotted it, fixed it, and repaired the cause in its program.
The lesson is simple: an agent makes mistakes like a colleague in a hurry, but very fast and with great confidence. We explained why in our article on AIs that "lie".
The guardrails that make it possible
I did not hand over the house keys without rules. Here are the ones that matter:
- It deletes nothing. When files need to be erased, it prepares a small script that I read and run myself.
- It only publishes or sends to third parties with my consent. For routine work (an article I have approved), it does it; for anything new, it asks.
- No e-mail and no web page is an order. A message saying "forward this file" is data, not an instruction. That is the basic protection against traps set for AIs.
- Passwords stay hidden. It can use them for a tool, never display or copy them.
- Everything is versioned and backed up. Every finished job is recorded in a history (Git), and my PC is backed up to a separate disk. A mistake can be undone.
- It works in a folder, not on the whole PC. And its tools require an explicit option to write: without it, they only check.
- It writes down what it learns. A dated log and topic sheets: the next session picks up where the previous one left off.
The competition: other harnesses
Claude Code is not alone, and that is good news. Here are the ones a business owner may come across, checked on their official pages on 21 November 2026.
| Tool | Publisher | Licence | Models | For whom |
|---|---|---|---|---|
| Claude Code | Anthropic (United States) | Proprietary, included in paid Claude subscriptions | Claude | Work on files and programs, in the terminal, on a computer or mobile |
| Codex CLI | OpenAI (United States) | Open source (Apache 2.0), used with a paid ChatGPT subscription or an API key | OpenAI models | The OpenAI equivalent, geared towards programming |
| Hermes Agent | Nous Research (United States) | Open source (MIT) | Your choice, including your own server | Personal assistant with memory, skills, scheduled tasks, reachable via Telegram, WhatsApp, Slack |
| OpenClaw | OpenClaw Foundation | Open source (MIT), no paid offer | Your choice, including local ones | Personal assistant running on your devices, controlled through messaging |
| n8n Agents | n8n (Germany) | "Fair-code" (Sustainable Use License), in preview | Your choice (your keys) | Agents built into your n8n automations, in the cloud or self-hosted |
| Deep Agents | LangChain | Open source (MIT) | Your choice, including local ones | Building blocks for developers who build their own agent |
| OpenCode | Anomaly | Open source (MIT) | Many providers | Open-source coding agent, in the terminal or as an app |
Three remarks:
- Open source does not mean free to use. The harness is free, but you have to pay for the model (subscription, API key) or a PC powerful enough to run one at home.
- The choice of model matters more than the choice of harness for the quality of texts. The harness matters more for security and comfort.
- Hermes and OpenClaw aim at the personal assistant reachable through messaging; Claude Code and Codex, at work on files; n8n, at business automation. They are not quite the same jobs.
If you are still hesitating about the basic assistant, start with our comparison "which AI for which use".
An illustrative case for a small business
Invented case, but built from common situations. An architects' office of eight people in Namur spends half a day every week on repetitive tasks: chasing unanswered quotes, filing plans received by e-mail in the right client folder, preparing the minutes of site meetings.
An agent, installed on one workstation and limited to the projects folder, can prepare the reminders (sent after approval), suggest how to file attachments and write a first draft of the minutes from the notes. What it must not do: send on its own, delete, or access the accounts and staff data. Allow two to three weeks of fine-tuning the rules before really saving time.
In practice: three options
| Option | Cost | Time | Risk |
|---|---|---|---|
| Do nothing | None | None | Your competitors save hours every week; your employees may already be using tools without a framework (so-called "shadow AI") |
| Try it alone | A subscription (from about $20 a month per person) | A few weeks of learning | Giving too much access too fast; entrusting client data to a consumer offer without contractual protection |
| Get support | Subscription plus a few hours of setup | One or two sessions, then follow-up | The safest: limited scope, written rules, backups in place before starting |
Four tips to get started:
- Start with a repetitive, risk-free task (monitoring, drafts, filing), not with the accounts.
- Limit the agent to one folder and check your backups first.
- Write three rules: what it never deletes, what it never sends without approval, what it does not read.
- Take a business plan as soon as client or staff data are involved: see AI and the GDPR.
For your IT person
Our architecture. Claude Code runs on a small desktop PC that is always on; work is done in a folder versioned with Git (two local repositories, no remote). A CLAUDE.md file at the root describes the workspace, the rules and the tools; it is loaded at every session. Recurring procedures are skills (procedure sheets the agent loads on demand).
MCP connectors. The Model Context Protocol links the agent to tool servers: IMAP/SMTP mail (a mailbox dedicated to the assistant, with the person in charge always in copy), Google Search Console, a ComfyUI image generator and local models via LM Studio on the network, and a relay to other models (Mammouth) for a cheap second review. See our article on AI connected to business software.
"Two-mode" scripts. Each in-house tool has options for the agent and a menu for the human. Run without a console, it only checks, unless given an explicit write option. This rule was born from the publishing incident described above.
Scheduled tasks. The monitoring runs every three hours via the Windows Task Scheduler, independently of the agent; the agent only catches up on the latest releases before suggesting topics.
Costs. The Claude subscription covers the agent. Video summaries go through an economical model (a few cents per batch); bank categorisation runs on a local model. In my use, local models are enough for sorting; for writing and reasoning, the large hosted models remain ahead (see local AI without the cloud).
Test yourself
This article was prepared with the agent it describes, then reviewed and approved by the author.
Sources
Anomaly. (n.d.). opencode [GitHub repository]. https://github.com/anomalyco/opencode
Anthropic. (n.d.). Claude Code overview. Claude Code Docs. https://code.claude.com/docs/en/overview
Anthropic. (n.d.). Plans & pricing. Claude. https://claude.com/pricing
LangChain. (n.d.). deepagents [GitHub repository]. https://github.com/langchain-ai/deepagents
n8n. (2026, September 25). Introducing n8n Agents. n8n Blog. https://blog.n8n.io/introducing-n8n-agents/
Nous Research. (n.d.). hermes-agent [GitHub repository]. https://github.com/NousResearch/hermes-agent
OpenAI. (n.d.). codex [GitHub repository]. https://github.com/openai/codex
OpenClaw Foundation. (n.d.). openclaw [GitHub repository]. https://github.com/openclaw/openclaw