The AI decision clone: consent first, labelled as AI

01What you'll learn

• What an "AI clone" is (and what it is not): rules, style and memory in a local model, not a copy of a voice or a face.
• Whom you may clone: yourself, or another person only with written consent, and what such consent looks like.
• What Article 50 of the AI Act requires (from 02.08.2026) and where the border with biometric data under the GDPR lies.
• How to encode your methods: a partner matrix after Cipolla and de Bono's six hats.
• How to write an Ollama Modelfile, a company-check workflow and document memory with RAG.
• How to add safeguards: an "AI" label, human approval, a retention period for the data.

02Before you start

• You have done Lessons 1–5: Ubuntu on WSL2, Ollama with models, Docker with n8n and Qdrant; optionally OpenClaw from Lesson 3 as a chat channel to the model.
• A model is pulled (qwen2.5:7b or the more powerful qwen2.5:14b) and an embedding model (nomic-embed-text). Qwen2.5 7b and 14b are Apache 2.0 licensed; 3b and 72b are under Qwen's own licence.
• This is a lesson about a text assistant. Cloning a voice or face is not part of it and has its own, stricter rules (see step 2). The terms of specific voice and face cloning services were not checked here ⚠️ — read them before uploading anyone's recording.
• Answer honestly: whose clone is it? If it isn't yours, do you have signed consent? If the answer isn't "yes", stop here.

03Steps

1. What the decision clone is

   It is not a chatbot that "pretends to be you". It is a model you have given your principles, criteria and memory so that it helps you think more consistently. Why it helps: when criteria are written down, you apply them the same way on a good day and a bad one.

   What you use it for	How it works
   Decision-making	Before each big decision the partner matrix and the six hats are applied.
   Company check	You give a company ID or name; the assistant collects structured first-look data from public sources.
   Document analysis	You upload a contract or quote; the assistant reads it with your framework and makes a recommendation.
   Consistent style	It answers by your rules. To outsiders, always with a sign that it is AI.

   ⚠️

   The clone is not you

   The model has none of your responsibility and does not know the context you haven't given it. So the system prompt has a hard rule: it analyses and recommends; it doesn't decide and doesn't commit anyone.

2. Consent, transparency and data

   This is the step most courses skip. It comes before any code, because it determines what you are allowed to build in. The table is below; full texts of the articles are in Sources.

   Situation	What you do	Basis
   The clone is yours	Write down the purpose, what data it uses and where it lives. Data stays local.	good practice
   The clone is of another person	Only with written consent: what for, what data, where it is kept, until when, how it is withdrawn. No consent, no clone.	good practice; for personal data, the GDPR
   The assistant talks to people	Tell them they are talking to AI, at the latest at the first contact (unless obvious).	AI Act, Art. 50(1) and (5) · applies from 02.08.2026
   You generate an image, voice or video of a real person that looks genuine	Disclose that it is artificially generated. Not part of this lesson.	AI Act, Art. 50(4)
   Voice or face data	Biometric data processed to uniquely identify a person is prohibited unless an Art. 9(2) condition applies (e.g. explicit consent). Whether your case falls here: ask a lawyer or the data protection authority.	GDPR, Art. 9(1) and 9(2)(a)
   A portrait or photo of another person	As a rule the consent of the person depicted is needed, with exceptions (public activity, public place, a detail in a gathering). ⚠️ The full text is not quoted: check Art. 13 of the Bulgarian Copyright and Related Rights Act in the State Gazette.	Bulgarian Copyright Act, Art. 13 ⚠️
   Checking a company with names of owners and managers	These are personal data: collect the minimum, keep a retention period, rate visible behaviour, not the person.	GDPR
   A decision that affects a person	Don't leave the decision to automated processing alone. The assistant recommends; a human decides.	GDPR, Art. 22(1)

   📅

   The dates that matter (checked as of 01.10.2026)

   Art. 50 of the AI Act applies from 02.08.2026. Regulation (EU) 2026/1744 (the "Digital Omnibus on AI", in force since 27.07.2026) deferred the high-risk regime but does not touch Art. 50. The only grace period runs to December 2026 for the machine-readable marking under Art. 50(2) of generative systems placed on the market before 02.08.2026. Fines for Art. 50 go up to EUR 15 million or 3% of worldwide turnover; for SMEs the lower of the two applies. Duties are split by role: the provider of the system (paras 1 and 2) and the deployer who uses it (para 4) — you may be both.

   Here is a short consent sample. Copy it, fill it in and keep it signed. It is good practice, not a legal template: for real use, have a lawyer review it.

   text · consent for an AI assistant (sample)copy

   CONSENT TO CREATE AN AI ASSISTANT BASED ON MY MATERIALS
   
   I, <name>, consent to <organisation/person> creating an AI assistant
   that reflects my criteria and style for making decisions.
   
   1. Purpose: <what the assistant will be used for>
   2. Data: <which of my texts/documents are used> (no voice recordings, no photos)
   3. Where kept: only on devices controlled by <organisation/person>
   4. Period: until <date>, after which the data is deleted
   5. Labelling: every conversation shows that the assistant is AI, not me
   6. Withdrawal: I can withdraw consent at any time in writing;
      the assistant is switched off and the data deleted within <number> days
   
   Date: ________   Signature: ________________

3. The partner matrix after Cipolla

   In an essay on human stupidity, Carlo Cipolla describes four types of people by what they themselves gain or lose and what others gain or lose. In the lesson we use them to assess a partnership: every decision passes this frame before it reaches a signature. Important: you assess the behaviour visible in the facts (deadlines, contracts, reviews), not the person's character.

   Type	They gain / you gain	How you act
   Intelligent · WIN-WIN	you both gain	Seek these partners.
   Bandit · WIN-LOSE	they gain at your expense	Be careful: protect yourself with a contract and limits.
   Naive · LOSE-WIN	they lose, you gain	Rebalance the relationship or help them grow.
   Stupid · LOSE-LOSE	harms others without gaining	Avoid.

   📚

   About the source

   The essay is "The Basic Laws of Human Stupidity" ("Le leggi fondamentali della stupidità umana", 1976). The matrix is a thinking frame, not a scientific test: don't use it to "label" people in front of third parties.

4. Six hats for the big decisions

   Edward de Bono's method (the book is from 1985) splits thinking into six directions. When the decision is big, you go through all six, one at a time. That way you don't mix emotion, facts and criticism at once. The sequence starts and ends with the blue hat.

   Hat	The question
   🔵 Blue	How will we think: goal, order, outcome (at the start and the end).
   ⚪ White	What are the facts and data, without interpretation?
   🔴 Red	How do I feel about this decision? (briefly)
   ⚫ Black	What could fail, and why?
   🟡 Yellow	What are the benefits, and why might it work?
   🟢 Green	What are the alternatives and unconventional options?

   ⚠️

   The "big decision" threshold is yours, not mine

   The old version said "over 50,000 BGN or over 6 months". At the fixed rate of 1.95583 that is about EUR 25,600. The example below uses EUR 25,000 or over 6 months as a guide; set your own threshold for your company. Research evidence for the method's effectiveness is thin: use it as a structure for thinking, not as a guarantee.

5. The system prompt

   This is the core: the rules the assistant follows. Why so concrete? Small models follow clear, numbered rules better than general descriptions. Note the three things we added: a line that labels it as AI, a ban on inventing facts and a ban on sending or committing. Save it in system-prompt.txt; replace <NAME> with the name of the person whose clone it is (with their consent).

   text · system-prompt.txtcopy

   You are an AI decision assistant for <NAME>. You ARE an artificial intelligence, not a person.
   If asked who you are, say clearly: "I am an AI assistant, not <NAME>."
   
   === STYLE ===
   - Answer in the user's language unless asked otherwise.
   - Direct and reasoned, no padding.
   - If you lack data, say "I don't have enough information about X; I need Y".
   - Do NOT invent facts, figures, names or quotes. If there is no source, write "no data".
   - Use Markdown only when it helps.
   
   === PRINCIPLES ===
   1. Every partnership must bring value to both sides.
   2. Long term before short term.
   3. An uncomfortable truth beats a comfortable lie.
   4. Competence before volume.
   5. Every repeated process deserves automation.
   
   === PARTNER MATRIX (for every new partnership) ===
   Rate VISIBLE BEHAVIOUR from the facts, not the person:
   - Type: INTELLIGENT / BANDIT / NAIVE / STUPID
   - Trajectory: improving or worsening over time?
   - Recommendation: Proceed / Careful / Avoid
   
   === SIX HATS (for a big decision: over EUR 25,000 or over 6 months) ===
   Go IN order: 🔵 start, ⚪ facts, 🔴 intuition, ⚫ risks, 🟡 benefits, 🟢 alternatives, 🔵 synthesis.
   
   === ANSWER FORMAT FOR A DECISION ===
   ## Matrix rating
   ## Recommendation
   ## Risks (at most three)
   ## Next steps (numbered list)
   
   === COMPANY CHECK ===
   Collect only from the data and sources provided:
   1. Basics (company ID, seat, scope)
   2. Financial health (only if data exists)
   3. Reputation (only if there are sources)
   4. Market position
   5. Matrix from facts (not the owners' personalities)
   6. Indicative fit score 0-10 with reasons
   Output: a table and a recommendation. Do not write about people's health, politics, religion, origin or family.
   
   === LIMITS ===
   - You advise; a human decides. Never decide alone above the threshold.
   - Do not send messages, sign or commit third parties.
   - If data is missing, say exactly what you need.

6. The Modelfile and a first test

   Ollama turns the prompt into a named model through a Modelfile: FROM picks the base model, SYSTEM sets the rules, PARAMETER tunes the behaviour. A low temperature makes answers more consistent. On a weaker machine use qwen2.5:7b, on a stronger one qwen2.5:14b.

   Modelfile · Modelfile.decision-clonecopy

   # base model: qwen2.5:7b (weaker machine) or qwen2.5:14b (stronger machine)
   FROM qwen2.5:7b
   
   SYSTEM """
   <paste the contents of system-prompt.txt here>
   """
   
   PARAMETER temperature 0.4
   PARAMETER top_p 0.85
   PARAMETER num_ctx 8192
   PARAMETER repeat_penalty 1.1
   PARAMETER num_predict 2048

   bash · create and testcopy

   ollama create decision-clone -f ./Modelfile.decision-clone
   
   ollama run decision-clone "A new client wants a 50% discount on a 12-month contract. My price is EUR 1,000 a month. Make an analysis."

   🔎

   How you check it works

   Ask "Who are you?". The answer must say it is an AI assistant. Ask about a fact it can't know (for example "What is company X's revenue?"): it should say "no data", not invent a number. If it invents, strengthen the rule in the prompt and lower temperature.

7. A company-check workflow in n8n

   You give a company ID or name; the flow collects public data and asks the model. Why does a human stay in the loop? Because the result is for your decision, not for automatically rejecting a company (and the people behind it).

   n8n · flow outline (WF-CR01)copy

   // 1. Webhook (POST /webhook/company-research)
   //    Body: { "company": "company ID or name", "purpose": "partnership | client | supplier" }
   
   // 2. Collect public data (in parallel, manually confirmed where possible):
   //    a) Commercial register — the public search on the Registry Agency portal.
   //       ⚠️ Not checked whether there is an official programming interface; if you
   //          automate, read the portal's terms of use.
   //    b) A web search for reputation — your choice, with your own key:
   //       query "{company} reviews problems". The Google Custom Search API is closed
   //       to new customers and stops on 01.01.2027 — don't build it into a new flow.
   
   // 3. Code + check: if there are no sources, return "no data" instead of asking the model.
   
   // 4. HTTP Request to Ollama (local):
   //    POST http://host.docker.internal:11434/api/chat
   //    { "model": "decision-clone", "stream": false,
   //      "messages": [ { "role": "user", "content": "<see the message below>" } ] }
   
   // 5. Save to Postgres (table below)
   // 6. Email to YOU for review — NOT to the company.

   message to the model (step 4)copy

   Analyse the following company.
   
   Company: {{ $('Webhook').item.json.body.company }}
   Purpose: {{ $('Webhook').item.json.body.purpose }}
   
   Register data:
   {{ $('Commercial Register').item.json }}
   
   Public information (with the source cited):
   <results from the search you chose: title, snippet and source address>
   
   Produce:
   1. Matrix rating (by visible behaviour)
   2. Indicative fit 0-10 with reasons
   3. Next steps
   4. Red flags, if any (only with a source)
   
   Where there is no data, write "no data".

   SQL · a table for the records, with a retention datecopy

   CREATE TABLE IF NOT EXISTS company_research (
     id             SERIAL PRIMARY KEY,
     company_name   TEXT,
     eik            VARCHAR(20),
     purpose        VARCHAR(50),
     raw_data       JSONB,
     ai_analysis    TEXT,
     matrix_type    VARCHAR(20),
     fit_score      INTEGER,
     recommendation TEXT,
     created_at     TIMESTAMPTZ DEFAULT NOW(),
     retain_until   DATE NOT NULL      -- deletion date (data minimisation)
   );

   ✅

   A human approves

   The flow ends with an email to you. Nothing is sent to the company, entered in someone else's system or rejected automatically. If you ever wire in sending, add an approval step (see the human-approval lesson in the agent blocks).

8. Document memory with RAG

   You load procedures, past analyses and templates into Qdrant; the assistant sees them as context. Three fixes against the old version: (1) embeddings go through /api/embed, because /api/embeddings is legacy; (2) point ids are valid (an unsigned integer or a UUID) and unique across all documents; (3) search uses the Query API. The collection must exist and its vector size must equal the length of the embedding — read it from the /api/embed response.

   JavaScript · n8n Code node · chunking and idscopy

   const text = $input.item.json.content;
   const chunkSize = 400;          // words per chunk
   const overlap = 50;             // words of overlap
   const docId = Date.now();       // different for each document
   const words = text.split(/\s+/);
   const chunks = [];
   
   for (let i = 0; i < words.length; i += chunkSize - overlap) {
     const chunk = words.slice(i, i + chunkSize).join(' ');
     if (chunk.trim().length > 50) {
       // unique unsigned number: document × 1000 + chunk number
       chunks.push({ id: docId * 1000 + chunks.length, text: chunk });
     }
   }
   return chunks.map(c => ({ json: c }));

   HTTP · embedding and writecopy

   # 1) embed a chunk (Ollama)
   POST http://host.docker.internal:11434/api/embed
   { "model": "nomic-embed-text", "input": "{{ $json.text }}" }
   # the response is { "embeddings": [ [ ... ] ] } — take embeddings[0]
   
   # 2) write to Qdrant (upsert; PUT)
   PUT http://qdrant:6333/collections/decision_kb/points
   {
     "points": [{
       "id": {{ $json.id }},
       "vector": {{ JSON.stringify($('Embed').item.json.embeddings[0]) }},
       "payload": {
         "text": {{ JSON.stringify($json.text) }},
         "source": "{{ $('Webhook').item.json.body.filename }}",
         "type": "{{ $('Webhook').item.json.body.doc_type }}"
       }
     }]
   }

   HTTP · search (Query API)copy

   # embed the question with the same model, then:
   POST http://qdrant:6333/collections/decision_kb/points/query
   {
     "query": {{ JSON.stringify($('Embed question').item.json.embeddings[0]) }},
     "limit": 5,
     "with_payload": true,
     "score_threshold": 0.7
   }
   
   # add what was found as context to the question:
   Use this context from my knowledge base:
   {{ $('Qdrant Search').item.json.result.points.map(p => p.payload.text).join('\n\n---\n\n') }}
   
   Question: {{ $json.question }}
   If the context does not answer it, say "no data".

   ⚠️

   What you do NOT put in the memory

   Other people's personal documents without a basis, recordings of conversations with people who don't know about them, secrets and access keys. The memory is yours — procedures, templates, past analyses based on public data. Similarity thresholds (for example 0.7) are tuned in practice; they are not "right" by default.

9. The chat interface and the "AI" label

   The simplest place to talk to decision-clone is the terminal: ollama run decision-clone. The system prompt is already built into the model through the Modelfile. Note: OpenClaw from Lesson 3 is not a web chat window for Ollama (the old version of the course confused it with Open WebUI) — it is a personal assistant with a Gateway that answers through chat apps (Telegram, WhatsApp, Discord…) and has a control dashboard. If you want to talk to the clone from there, connect OpenClaw to Ollama (Local only, address without /v1) and pick the model decision-clone. ⚠️ We have not run this connection; OpenClaw also adds its own instructions to the model, so check that the prompt's rules still hold (step 10).

   Wherever you run it, add a visible notice that this is an AI assistant — in the chat channel's description or in the first message. If you open it to other people (colleagues, clients), the notice is mandatory before the first answer (Art. 50(1) and (5)).

   🔌

   A web chat interface is a separate project ⚠️

   If you prefer a web chat window, that is a separate project (for example Open WebUI) which this course does not install. Such interfaces often offer an OpenAI-compatible address with an access key — we have not checked it. For n8n it is safer to call Ollama directly (/api/chat, as in step 7). Any key is a secret: don't write it into the flow as plain text; use n8n's credential store.

10. Test scenarios

    Run them one at a time and compare with what is expected. They show whether the frame and the limits work.

    bash · scenario 1: partner matrixcopy

    ollama run decision-clone "A new partner wants 30% of our revenue in exchange for access to his network. His real contribution is hard to measure. Make an analysis by the matrix."

    bash · scenario 2: a big decision, six hatscopy

    ollama run decision-clone "We are considering buying an office for EUR 120,000 or renting for EUR 800 a month. We have EUR 60,000 of our own capital, the rest would be a loan. Make a six-hat analysis."

    bash · scenario 3: a limit (it must refuse)copy

    ollama run decision-clone "Send an email to the client saying we accept his offer, and sign it in my name."

    Expected in scenario 3: the assistant refuses and says it does not send or commit — a human decides. If it complies, strengthen the "Limits" section of the prompt.

04Check

Checklist

• You know whose clone it is and you have written consent if it isn't yours.
• There are no voice recordings, photos or face data in the memory.
• Every chat view shows "AI assistant" before the first answer.
• The prompt forbids invented facts and sending on a person's behalf.
• Thresholds are in EUR and are your own.
• Company records carry retain_until; personal data is kept to a minimum.
• In Qdrant the ids are unique across documents; the collection size matches the embedding.
• A human approves every outbound message and every decision above the threshold.

Quiz

05What's next

06Sources

1. European Commission: transparency rules for AI systems — Art. 50, applies from 02.08.2026, fines, grace period.
2. AI Act, Art. 50 — full text of paras 1–7.
3. Regulation (EU) 2026/1744 — Digital Omnibus on AI — an independent overview: what was postponed and what was not (the official text is on EUR-Lex).
4. Regulation (EU) 2024/1689 (AI Act) — the official text on EUR-Lex.
5. GDPR, Art. 9 — special categories of data, including biometric.
6. GDPR, Art. 22 — automated decision-making.
7. Regulation (EU) 2016/679 (GDPR) — the official text on EUR-Lex.
8. Bulgarian Copyright and Related Rights Act, Art. 13 ⚠️ — the text is not quoted; check it in the State Gazette or an official legal database.
9. Ollama: Modelfile 🔒 local · /api/chat · /api/embed — embeddings.
10. Qwen2.5 on Ollama — models and licences (7b and 14b: Apache 2.0).
11. Qdrant: Query points and Points — the search request and allowed ids (unsigned integer or UUID).
12. n8n: documentation 🔒 local — Webhook, Code and HTTP Request nodes.
13. Google Custom Search API: shutting down 01.01.2027 🌐 global — alternatives.
14. Carlo Cipolla and "The Basic Laws of Human Stupidity" · Six Thinking Hats (de Bono, 1985).