Does the AI Act actually apply to you?
For your own business, or as a service you charge other businesses like yours for.
Nobody runs this lab yet. Take it, give it your hours and your prices, and it is yours.
| System | What it does | Decides on people | Provider or deployer |
|---|---|---|---|
| CV sorter | Ranks applications before a human reads | Yes, hiring | Deployer, bought tool |
| Support chatbot | Answers customers on the site | No | Deployer, bought tool |
| Invoice reader | Extracts figures from supplier PDFs | No | Deployer, bought tool |
| Internal writing assistant | Drafts emails for staff | No | Deployer, bought tool |
Four systems, one of which decides on people. The column that changes everything is the third, and it is the one nobody fills in before being asked.
Your situation
You run two systems. The CV sorter is high risk under Annex III, point 4. The support chatbot is limited risk: it only owes transparency.
What that means before August 2026
| SystemSystem | ClassClass | What is owedWhat is owed |
|---|---|---|
| CV sorter | High risk | Logging, human review, technical file |
| Support chatbot | Limited risk | Tell the user they are talking to an AI |
Start with the technical file for the sorter. It is the long one, and it is the one an inspection asks for first.
One line per system, kept up to date, produced on request
An inspection asks for this before it asks for anything else. Kept as a living file, not rebuilt the week it is asked for.
CV sorter, high risk
| FieldField | EntryEntry |
|---|---|
| Purpose | Ranks applications for the two open roles |
| Supplier | Named in the contract, version recorded |
| Human review | Head of HR, reads the full pile, records the override |
| Logs kept | 6 months, on the supplier's side, exportable |
| Last review | 11 August, by the HR lead |
Support chatbot, limited risk
| FieldField | EntryEntry |
|---|---|
| Obligation | Tell the user they are talking to an AI |
| Where it is said | First message and footer of the widget |
| Last check | 11 August, wording unchanged since June |
What goes wrong with registers
They are written once and never touched again. Put the review date in it, and review when the supplier ships a version, when the purpose changes, and once a year regardless. A register whose last line is eighteen months old says more against you than no register at all.
The EU AI Act reads badly. It is long, written for lawyers, and most small companies first meet it through a vendor slide with something to sell behind it. The question that matters is simple: does what my company does with AI fall under this text, and if so under which part. This lab would sort that out from the text itself and from an inventory of your own systems, and it would keep the register current as it goes. It does not give legal advice, and it says so.
Put there once, read by every magik.
Put there once by you, read by every magik, corrected when reality moves. Full on day one, unlike the collections below.
Worth more in month six than in week one.
The collections belong to the lab, not to any single magik. Every magik reads them and every run adds to them, which is why this lab is worth more in month six than in week one.
| System | Use | Risk | Obligation |
|---|---|---|---|
| Support assistant | Answering customers | limited | tell the user |
| CV screening | Hiring | high | register and oversight |
| Image generator | Marketing | limited | label the content |
Invented for this page. Your Your AI systems collection starts empty and fills as you work.
It wears your name, not ours.
The same lab, under your name, your colours and your logo. Your customers never see ours.
AskMOJOOn the Pro plan, and not needed to take a lab. A fully bespoke theme exists as an upgrade, hand built rather than self-serve.
Web search
FirecrawlAnd whatever you already use: any tool available through an API or an MCP server, connected with your own account.
A tool built for exactly your job, and only that, will go deeper at that one job. This is for the jobs nothing covers, or the ones spread across three tools today.
The regulation applies in stages, and the dates are in the text itself. The lab holds no calendar of its own: it reads the version of the text you put in, which is the only way an answer stays true after the next amendment. A hard coded date would be wrong within a year.
Regulation (EU) 2024/1689 is published in the Official Journal of the European Union, and that published text is what you put in as a resource. The lab works from it rather than from a summary of a summary, which is exactly where most of the confusion on this subject comes from.
What decides is the use, not the headcount: what the system does, whether you use it or put it on the market, and what it decides about people. A six person company can hold an obligation a large one does not. The first magik asks exactly those questions.
It depends on the use, and that is exactly what the first magik asks. Using an AI system does not carry the same duties as placing one on the market, and the same tool changes category depending on what it decides. The lab works from the text itself, which you put in, not from a second hand summary.
No, and a lab that claimed otherwise would be selling you a false sense of safety. It produces an inventory and a reading of the text against your uses. Compliance is a set of decisions and documents that your management signs, not the output of a run.
For turning up at your counsel with the inventory already done: which systems you use, for what, who decides. That is the part nobody but you can produce, and the part that costs the most to reconstruct in a meeting. It does not replace the advice, it stops you paying for a list.
Keeping track of your systems and how they are used is worth it with or without an obligation: it is what lets you answer within the hour when a client, an insurer or an auditor asks. The lab keeps that register current run by run, rather than in one evening a year.
For the judgement calls, probably. What neither a course nor a consultant will do for free is the inventory of your own systems, and that is the work this idea takes off your hands. It makes the paid hours go to the questions that need a person.
A tool built for exactly small companies using ai, and only that, will go deeper at that one job. This lab is for the jobs no tool covers, or the ones spread across three tools today, and it runs under your name.
Every audience is still open.
Nobody has taken a square on this idea yet. You declare your audience when you take the lab (“roofers in Brittany”, “plumbers around Paris”), and two people can start from the same idea on two different audiences without getting in each other’s way.
And it wears your name, not ours.
The bundle describes a lab that does not exist yet : the jobs it would do, and an example of each output.
Open Knowledge Format : plain markdown, re-importable into any agent. No account, no lock-in.
You hear about a competitor's price change from your lab, not from a customer asking why you cost more.
Web searchFirecrawl
ExaGive it five things you wrote. Everything it writes after that sounds like you, instead of sounding like a chatbot.
A prospect describes what they want at midnight. You wake up to a brief, a price range, and whether it is worth a call.
The scoping, the brief and the price range are done before you open your laptop.
Gmail
Notion
WhatsApphome screen appStill comparing tools? Best AI for research.