PREVIEW · nothing can be bought yet · not an offer, no contract can be concluded

You use AI.You just can't prove it.

Not through negligence. No one procured ChatGPT — somebody opened it. The EU AI Act has required measures to support AI literacy for 608 days; it prescribes no record of them today. Whoever gets asked still has only what they wrote down.

KI Akte turns that into a file: you list your tools, confirm each classification yourself, and receive your dated documents — generated, not copied out. No consultancy, no project, no compliance department you don't have.

ki-akte · Example GmbH
[RECORDED] 7 AI systems in use
[PROPOSED] risk class per system
[CONFIRMED] by management, 7 of 7
[GENERATED] file, 7 documents
{
  "as_of": "2026-10-03",
  "systems": 7,
  "high_risk": 0,
  "regenerable": true
}
608days Art. 4 AI Act in force
62days Art. 50 AI Act in force
7documents per file
anytimeregenerable
20 seconds

Does this even apply to me?

Three questions. The first two decide whether the Regulation reaches you at all — a no there and you are finished here. The third decides the opposite: if you answer it with no, that is precisely the gap.

Start the pre-check — free, no account →

Question 01

Do you use AI in the business?
ChatGPT, Copilot, Gemini, an AI feature inside software you already run, a chatbot on your site. Occasional use by one person counts.

Question 02

Are you a deployer, not a provider?
You use other people's AI; you don't build and sell your own. Anyone placing high-risk AI on the market needs far more than this file.

Question 03

Could you show today which systems are in the house and who was trained?
If yes, you already keep a file — you don't need us. If no, that gap is exactly the point.

Three times yes? Then Art. 4 AI Act applies to you (AI literacy — you must take measures so that staff using AI understand it sufficiently) since 2 February 2025. Unconditionally, for every deployer.

Art. 50 AI Act reaches you as a deployer in two cases only — and most companies fall into neither: para. 3, if you operate emotion recognition or biometric categorisation — then you inform the people exposed to it. And para. 4, if you publish deep fakes or AI-generated text on matters of public interest — then you disclose it. Both since 2 August 2026.
That a chatbot must identify itself (para. 1) and that AI content must be machine-readably marked (para. 2) is the provider's duty — not yours. This distinction decides whether you owe anything at all — and your file records it per system, with the rule branch that produced the classification.

What each month without a file costs. Not a fine — provability. A record you write today is dated today. Art. 4 AI Act has applied for 608 days; every further month is operation you can describe later but cannot evidence contemporaneously. You cannot back-date a record.

Who checks: the national market surveillance authorities. The Regulation provides for penalties (Art. 99 AI Act); which tier applies depends on the duty breached and the size of the company. We deliberately quote no figure — that is a legal assessment of your case, and not ours to make.

You already do this

You keep one of these for GDPR.

Every company maintains a record of processing activities because the GDPR demands it. For AI, the AI Act does not demand one today. It is still the same job: knowing what runs in the house — and being able to produce it when someone asks. The KI-Akte is your record for AI, kept voluntarily.

Like an external data protection officer — but for the file.
The responsibility stays with you. We keep the records.

That is a comparison, not a role: the AI Act knows no “AI officer”. We take on no duty the law assigns to you and make no decision that is yours — we keep the records with which you evidence them.

The difference

Today — and with a file

Today
  • A spreadsheet last touched in the spring
  • Nobody knows which AI tools are actually in use
  • Training happened — but cannot be evidenced
  • "Show me your AI governance" costs you a week of searching
With KI Akte
  • A maintained register of every system, structured
  • A confirmed classification per system, with its reasoning shown
  • The literacy record exists as a document, not a memory
  • The same question is answered with a dated PDF
The file 01

Documents, not advice

In German administrative practice an Akte is precisely this: the collection you maintain and produce on demand. You enter your AI tools into a guided form — name, vendor, purpose, who comes into contact with it. The full set follows.

Everything datedEach document carries the date it was generated. An undated file is not evidence when it matters.
What you contributeFour steps: pre-check · list your AI tools · confirm a proposed classification per system · training details. No preparation and no paperwork needed — only knowing which tools are in the house. How long it takes depends entirely on that number.
Table of contents
01 AI system register             basis
02 Risk class per system          Art. 6
03 Transparency notices           Art. 50
04 AI usage policy DE/EN          governance
05 AI literacy record             Art. 4
06 Logging record                 internal
07 Human oversight evidence       internal
08 Operational monitoring         high-risk
09 ISO 42001 gap analysis         optional
-- As of: 03.10.2026
-- 01-04 generated · 08 and 09 in progress
-- 05-07: you attach your own evidence
The file 02

We assemble. The verdict stays yours.

This is not a disclaimer, it is how the product is built. KI Akte proposes a risk class per system and shows its reasoning — you confirm it. Your confirmation is held together with the rule branch that produced the proposal, and the date. That is the construction intended to keep this software rather than a reserved legal service under § 2 RDG. Whether the line holds is for the independent legal review to confirm — it is still outstanding.

No conformity verdictNo declaration of conformity, no CE marking, no certificate, no advice on your individual case. That is your lawyer's work, and it stays theirs.
Classification · system 03
System   "Website chatbot"
Contact  "customers"

[PROPOSED] limited risk
Reasoning: no Annex III purpose ·
           public contact → Art. 50

[?] Confirm this classification.
[✓] confirmed · MD · 03.10.2026

What happens to what you enter. Your input is processed to generate your file and is not retained afterwards. It does not feed a training or knowledge system. The data processing agreement and the full processing details will be in place at release — until then we describe how it is built and make no contractual commitment here.

The file 03

Not an audit that expires in three months

A one-off report describes a single day. Your AI landscape moves faster: a new tool, a changed policy, a training session. So the file is not written — it is generated, and it can be regenerated whenever you need it.

The actual valueNot the first document. The twelfth, which is as current as the first — without anyone having had to remember.
And when the law itself movesYour tools are one half. The other is the Regulation, and it is not finished: the next deadline falls on 2 December 2026, in 60 days. A spreadsheet you start today is silently out of date that day, and nobody in the company notices. KI Akte keeps the normative index the classification itself reads — what applies today and what is still to come. That you need not remember the date — that you hear from us and the file is regenerated — is what comes with launch.
Regenerate
Change recorded:
  + "Notion AI" now in use
  + training, 4 people, 19.08.2026

[REGENERATED] 7 documents
[AS OF]      03.10.2026
[PREVIOUS]   02.07.2026 (archived)
The file 04

On high risk, the product stops

If a system looks like Annex III — recruitment, creditworthiness, biometrics, critical infrastructure — KI Akte does not quietly classify it as harmless to keep the flow moving. It stops and refers you to your lawyer.

Who this is built forMittelstand, 10 to 99 staff, using other people's AI. Anyone building or placing high-risk AI on the market needs more than this file — and gets told so, rather than having it overlooked.
Check · system 06
System  "CV pre-screening"
Purpose "recruitment"

[STOP] possible Annex III purpose
This system is not classified
automatically. We do not guess here.

[NEXT] with your legal counsel
[FILE] other 6 systems unaffected
The alternatives

What else exists — and why it doesn't fit

There are tools for this problem. They are simply not built for you.

The platforms

Continuous-scan compliance platforms. In English, designed for AI providers, price on request, access via a demo call. For a twenty-person company that means, in practice: not built for you. They monitor continuously — you need something you can produce.

The checklists

Free self-assessments. They tell you what you would have to do and generate nothing. You end up with a list of open points and still no evidence. That part we give you for free — the pre-check. What you pay for is what comes after it.

The consultancy

An opinion as at a cut-off date. It describes a single day, costs many times more, and ages the moment somebody installs a new tool. For an individual legal question it remains right — for a register that has to stay current it is the wrong instrument.

The constraint we impose on ourselves is not a feature: KI Akte does not render the legal verdict. Anyone who renders the individual legal conclusion leaves § 2 RDG behind. That limits what we may promise you — and it is why every classification in your file is yours and not ours.

Evidence

The first file is our own.

This is normally where customer quotes go. We have none — KI Akte has not been sold yet, and invented evidence is not something we will publish. What exists is a file we generated for ourselves, from a dated snapshot of our own agent registry.

Evidence 01

13 entries, register snapshot of 22 June 2026: 12 agents from that snapshot of our agent registry, plus one project. The file proposes a class for each entry and confirms none: 11 × transparency under Art. 50 AI Act, 1 × minimal, 1 × Annex III. AI we use beyond these entries is not in the file.

Evidence 02

Same input, same answer. The classification is a fixed decision tree under the AI Act — Art. 5 → Annex III → Art. 50 → minimal — not a language model. There is exactly one implementation of it, shared by German and English — the language cannot reach the result at all. Across all 36 purpose-and-contact combinations: 15 halts. Each of the five Annex III purposes reaches the Annex III branch — asserted by a test that reads the list of purposes out of the questionnaire itself, so a sixth cannot fall through. You can show your lawyer which branch applied.

Still missing

Three real customer files. Only then does the brand go public. Until then we say what we have — and what we don't. Nobody can sell you compliance. Evidence, yes.

How it will work

Four steps — and the fourth is the reason we exist

So you can judge what you would be getting: what already runs today is marked as such. So is what comes with launch. We describe nothing as present that does not yet exist.

01
You enter your tools Name, vendor, purpose, who comes into contact with it. No account, nothing transmitted. For each system the file proposes a risk class — you confirm it, or you do not. runs today
02
The file is generated Register, confirmed classification per system with its rule branch, transparency notices and usage policy in German and English. Dated. Regenerable at any time. with launch
03
We keep the normative index Which articles apply, which were deferred, which deadlines are still ahead — the same source the classification reads. Public at Legal status. runs today
04
When something changes, you hear from us If a deadline moves or a duty is added, we regenerate your file and tell you what changed — without you having to remember. That needs an account and your consent to store your entries. with launch

Why step 04 is the actual point. A spreadsheet you start today is silently out of date on 2 December 2026 — in 60 days — and nobody in the company notices. That is not a knowledge gap you close by reading harder: it needs someone watching on your behalf, continuously. That is the difference between a document you buy once and a file that stays true.

Steps 02 and 04 are not yet released. This description is not an offer within the meaning of § 145 BGB and deliberately names no price.

Status

Not for sale yet — deliberately

We have publicly committed to an independent legal review of the boundary to § 2 RDG, and will publish its outcome whatever it is. It is still outstanding. Until it lands, we do not sell publicly. The one exception is a founder cohort of at most three pilot customers, each agreed individually and by name. These customers confirm in writing, before buying, that the review is open — and know what happens if it comes back qualified: full refund, and the file stays with them.

Price

There is deliberately no number here yet.

A price is a promise. We can make one once the legal review has concluded and it is settled what we are allowed to deliver — not merely because a number looks good on a page. If you would like to know sooner, write to us and we will tell you where we stand.

The offer will be addressed exclusively to businesses within the meaning of § 14 BGB. All prices are net; statutory VAT is added where it applies. For businesses established in another EU member state, the tax liability shifts to the recipient under § 3a(2) UStG (reverse charge). This page is not an offer within the meaning of § 145 BGB.

Release 01

Legal review. A lawyer confirms the boundary to reserved legal services. Without it, no public sale. What speaks for the boundary today, said openly: the German Federal Court of Justice, judgment of 9 September 2021 – I ZR 113/20 (smartlaw) on a rule-based document generator · a fixed decision tree with no language model · and the halt at Annex III, Art. 5 and provider status, where we do not classify and produce no record. We publish the opinion when it lands — including if we do not like it.

Release 02

Terms, liability, data processing. Contract, insurance and the DPA are in place before we sell publicly.

Release 03

Three real files. The brand goes public only once it has shown on real cases that it holds.

Notify me at release

One message, when KI Akte goes live. No marketing, nothing passed on.

Write to us

Deliberately a mail link and not a form: until the data-processing agreement is in place we do not start a list with your address on it. Your message sits in our mailbox. If we reply, our support register records who was written to and with what subject — not the content. No marketing list and no waiting list is created from it.

Reference

One question, answered on its own page

Each page answers exactly one question — in the first paragraph, with the article the answer follows from. Not legal advice, and no assessment of an individual case.

KI Akte is software, not legal advice. The product assembles evidence from your own, self-confirmed statements. It does not assess your individual case, does not declare conformity, and does not replace a legal review. You remain responsible for the substantive correctness of the classifications you confirm.

Preview. Not an offer within the meaning of § 145 BGB, no binding prices, no contract can be concluded. Statements about deadlines under Regulation (EU) 2024/1689 are given in good faith, without warranty.

What the pre-check asks · Legal status · Imprint · Privacy© 2026 KI Akte