assurance · Level 2
What is sovereign AI, and why it matters to you
Who actually controls your data and the AI tools you use, in plain English, and the real questions worth asking before you trust a free AI tool with anything that matters.
Your next step
Start this course online
Register your start, then work through the lessons and exercises at your own pace. You can download the PDF separately. Where a scored exam is available, you will find it after the lessons.
Register or resume your course
Start with the essentials
The short answer
Sovereign AI means having real control over your data, the model you use and the infrastructure it runs on, not just access to a tool. A free AI tool can still store your input, use it to improve its models, or move it across borders, unless its terms say otherwise. Read the terms, and ask a vendor directly where control actually sits.
What you will learn
- You will be able to explain what sovereign AI means, for a country, an organisation and a person, in your own words.
- You will know why a free AI tool does not automatically mean your data stays under your control, and what to check instead.
- You will understand the difference between an open-weight model and a closed model, and why that affects who controls it.
- You will be able to recognise the signs of vendor lock-in before you are stuck with them.
- You will have a written list of specific questions you can put to any AI vendor about your data and your control over it.
- You will know where to check what UK data protection law actually says about moving your data across borders.
Who it is for
Anyone who uses AI tools and wants to understand who really controls their data and their work, plus anyone in a small organisation choosing which AI vendor to trust. No technical background is needed.
Before you start
- None beyond curiosity. What is AI? is helpful background if you are new to AI tools altogether.
Learn in your browser · Chapter 01 of 06
What sovereign AI actually means
The term usually appears in headlines about governments and billion-pound infrastructure. Strip that away and it describes something you can check for yourself, today, about the AI tools already on your phone or your laptop.
Sovereign AI is usually defined as a country's or an organisation's ability to develop, run and govern its own AI using its own infrastructure, its own data, its own models and its own people, rather than depending entirely on someone else's. That is the definition you will find in explainers written for governments and large companies. Underneath the scale, the same idea applies to you: do you actually control what happens to what you type into an AI tool, or has someone else quietly taken that control while you were focused on getting an answer.
Control, in this sense, has three separate parts, and it helps to keep them apart rather than treating "AI" as one single thing you either trust or do not.
Scroll sideways to see every column.
| Part of the stack | The question it raises | A plain example |
|---|---|---|
| Data | Where does my input go, who can see it, and can I get it back out? | A chat message you typed: is it stored, and for how long? |
| Model | Who built it, who can change it, and could I run an equivalent myself if I had to? | A model only reachable through one company's app, versus one you could download |
| Infrastructure | Whose computers is it actually running on, and in which country? | A server you have never seen, owned by a company you have never dealt with directly |
Two friends give a useful worked example. Priya uses a free chatbot from a large consumer AI company for everything, including drafting messages that mention her clients by name. Tom uses a similar free tool, but first checked its settings and turned off the option that lets his conversations be used to improve the model, and he keeps anything genuinely sensitive out of it entirely. Both are using AI. Only one of them has actually thought about where control sits, and it took him fifteen minutes, not a computer science degree.
Try it yourself · Activity 01
15 minMap your own AI tools
List every AI tool you have used in the last month, from a chatbot to a writing assistant built into another app.
- Write down each tool's name and roughly what you use it for.
- For each one, write what you currently believe about where your input goes, from memory, not by checking yet.
- Mark any tool you have used for something genuinely sensitive: client work, health, finances, anything about another person.
- Circle the one tool on your list you know the least about, despite using it the most.
Were you able to answer confidently for most tools, or mostly guessing? Most people, honestly, are mostly guessing. The rest of this workbook gives you a way to stop guessing.
Worked answer
Example: ChatGPT-style assistant (drafting emails, daily), a grammar checker built into a word processor (editing, daily), a company chatbot on a retailer's website (customer queries, occasionally). Belief: the assistant probably stores conversations somewhere; the grammar checker seems local but is not certain; the retailer's chatbot is a complete unknown. Sensitive use: none deliberately, though a draft email once mentioned a colleague's diagnosis in passing. Least known: the retailer's chatbot, used most days for work questions, never checked.
Learn in your browser · Chapter 02 of 06
Where your data actually goes when a tool is free
Free does not mean nobody is being paid. It usually means your input, in some form, is part of the arrangement. This chapter teaches you how to actually check that, instead of guessing.
A free AI tool still costs real money to run: the computers, the electricity, the people who built it. That cost has to be covered somehow, and "your data helps improve the product" is one of the most common ways it is covered. That is not automatically wrong. What matters is whether you know it applies to you, right now, for the tool in front of you, because that changes what happens to what you type.
- Whether your input can be used to train or improve future versions of the model, by default or only if you opt in.
- Whether a human being might review some of your conversations, and under what circumstances.
- Which country or countries your data is stored and processed in, and whether it can move between them.
- How long your data is kept, and what "deleted" actually means: removed from your account view, or genuinely gone from every backup.
When personal data leaves the UK, UK data protection law has something to say about it regardless of which AI tool is involved. A transfer to a country without a UK adequacy decision needs an appropriate safeguard, such as a standard contract clause, before it can happen lawfully. You do not need to become a data protection lawyer to use this. You need to know the question exists: does my data leave the country it was collected in, and if so, under what safeguard.
Try it yourself · Activity 02
25 minActually read one tool's terms
Pick the AI tool from Chapter 1 that you circled as least known, or any tool you use regularly for anything beyond idle curiosity.
- Find its privacy policy and its terms of service, usually linked at the very bottom of its website or inside its account settings.
- Search the page for the words "train", "improve" and "third part" and read the sentences around each match.
- Find any setting that lets you opt out of your input being used to improve the model, and note whether it is on or off by default.
- Note which country or countries the policy says your data is processed in, if it says at all.
Was the answer easy to find, or spread across several pages of dense legal wording? A policy that hides the answer is itself useful information about how much that company wants you to know.
Worked answer
Example, for an invented consumer chatbot: the privacy policy states that conversations may be used to improve the service unless the account holder turns on a specific privacy setting, which is off by default. Data is processed on servers in the United States and the European Union, with international transfer safeguards named but not explained in plain language. No fixed deletion time is given for backups. None of this makes the tool unsafe to use casually, but it does mean sensitive material should not go into it without turning that setting on first.
Learn in your browser · Chapter 03 of 06
Who owns the model, and why that matters
Two AI tools can look identical on screen and sit on completely different foundations underneath. Ownership decides who can change the rules, and who you would have to trust if things went wrong.
Every AI tool is built on a model, and models come in two broad shapes. A closed model is only reachable through its owner's own app or service. You send a request, their computer answers, and the model's inner workings are never published. An open-weight model has its trained parameters published, under a stated licence, so that other people or organisations can download it and run it on their own computers, or a cloud provider of their choosing.
Scroll sideways to see every column.
| Closed model | Open-weight model | |
|---|---|---|
| Who can run it | Only through the owner's own service | Anyone, on hardware they control or choose, within the licence terms |
| What happens if the owner changes the terms | You accept the new terms, or you stop using it entirely | You can usually keep running the version you already have |
| How much you can inspect | Little to nothing about how it actually works | Independent researchers can generally test and publish findings on it |
| What it still leaves unanswered | Nothing extra beyond what is stated | Nothing about who trained it, on what data, or whether you should trust that |
Ownership matters most at the moment things go wrong for you: the vendor changes its terms in a way you dislike, discontinues the exact model you had built a workflow around, or is bought by a different company with different priorities. With a closed model, your only choice at that point is to accept the change or leave with whatever you can export. With an open-weight model you already have a copy of, at least that specific version keeps working, on hardware you or someone you trust controls, regardless of what its original publisher decides next.
Try it yourself · Activity 03
20 minCheck what you could actually keep
Pick one AI tool you rely on for something that would genuinely disrupt you if it disappeared tomorrow.
- Find out, from its own documentation or support pages, whether it is built on a closed model, an open-weight model, or does not say clearly.
- Look for an account setting that lets you export your data: conversation history, saved documents, custom instructions.
- Try the export, if one exists, and check whether the file you get is actually usable elsewhere, or locked to that one tool.
- Write one sentence on what you would genuinely lose if this specific tool stopped existing tomorrow.
Did the answer change how comfortable you are relying on this tool for something important, or did it confirm you were already comfortable for good reasons?
Worked answer
Example: a note-taking app with a built-in AI assistant. The assistant runs on a closed model reached only through the app's own service. An export option exists and produces a plain text file of notes, but any AI-generated summaries and the assistant's memory of past conversations are not included in the export. What would be lost: the assistant's accumulated context about ongoing projects, which would need rebuilding from scratch elsewhere. That is a real, specific cost, now written down instead of vaguely assumed.
Learn in your browser · Chapter 04 of 06
The lock-in trap: how switching gets harder over time
Nobody usually decides on day one to become dependent on a single AI vendor. It happens gradually, through small choices that each felt reasonable at the time. This chapter shows you the pattern early enough to still act on it.
Vendor lock-in is the state of being stuck with one supplier because switching away would cost more time, money or lost work than most people are willing to pay. It rarely announces itself. It builds up through convenience: a custom instruction you spent an hour tuning, a habit your whole team has learned, a format only that tool understands, an integration wired into three other systems you also depend on.
- Your work, prompts or custom instructions only exist inside that one tool, with no working export.
- Switching would mean retraining yourself or a whole team on a different way of working.
- Other tools you use are wired directly into this one, so leaving breaks more than just this one habit.
- You genuinely do not know what a comparable alternative would cost you in time to set up.
- The tool has changed its terms, its behaviour or what it offers before, and you stayed anyway because leaving felt harder than adapting.
None of this means you should avoid convenient tools, or refuse every integration. It means going in with your eyes open, and building in an exit before you need one, the same way you would keep a spare key rather than only discovering you are locked out once the door is already shut.
Try it yourself · Activity 04
20 minScore your own lock-in
Take the same tool from Chapter 3's exercise, the one that would disrupt you most if it disappeared.
- Go through the five signs above and mark each one yes or no for that tool.
- Count how many are yes. Three or more is a genuine lock-in position, whatever the tool's other merits.
- For each yes, write one small, realistic step that would reduce it: a regular export habit, keeping a copy of your custom instructions in a plain document, learning one alternative tool well enough to switch quickly if you had to.
- Pick one of those steps and actually do it this week, not just plan to.
Did scoring it change how you feel about the tool itself, or just about how prepared you are to leave it if you ever needed to?
Learn in your browser · Chapter 05 of 06
Questions you can actually put to a vendor
This is the practical core of the workbook: a real set of questions, worded so you can send them or check for the answers yourself, whether you are one person or a small organisation choosing a vendor.
You do not need special standing to ask an AI vendor these questions. Most are answered somewhere in a privacy policy, a terms page or a support article; some genuinely need a direct email to a vendor's support or sales team, which a small organisation is entitled to send before signing anything.
- Is my input used to train or improve your models by default, and can I turn that off?
- Where is my data stored and processed, and does it ever move to a different country?
- Can I export everything tied to my account, in a format I could actually use somewhere else?
- If I delete something, is it actually gone, including from backups, or only hidden from my own view?
- Who inside your organisation can see my data, and under what circumstances would they?
- What happens to my data and my work if I stop using your service, or if you discontinue it?
- Is the model you use something I could run elsewhere if I genuinely needed to, or is it only ever available through you?
- What would happen to my data if your company were bought by another one?
- Where can I read your actual current terms in full, rather than a marketing summary of them?
Scroll sideways to see every column.
| If the honest answer is | What it tells you |
|---|---|
| "We cannot say" or no clear answer at all | A real gap worth weighing before you rely on this vendor for anything sensitive |
| A clear answer, even an unwelcome one | At least you can make an informed choice, which is the actual goal here |
| A confident answer that does not match what the written terms actually say | Trust the written terms, not the reassurance, and ask again in writing |
Try it yourself · Activity 05
25 minAsk three questions for real
Pick one AI tool you rely on for something that actually matters to you or your work.
- Choose three questions from the list above that you genuinely do not know the answer to for this tool.
- For each one, find the answer in the tool's own published terms or settings, or send a short, direct email asking.
- Write down what you found, including if the answer was unclear or missing entirely.
- Decide, honestly, whether anything you learned should change how you use this tool.
Which question was hardest to get a straight answer to? That is usually the question worth asking again, and worth asking of every vendor you consider afterwards.
Learn in your browser · Chapter 06 of 06
What you can do now
You now have a working definition of sovereign AI that applies at your own scale, not just a government's, and a real set of questions to use whenever you pick up a new AI tool.
You can now look at any AI tool, consumer or business, and ask the three questions that actually matter: what happens to my data, who owns and controls the model behind it, and could I leave within a week if I needed to. Those three questions cover data, model and infrastructure, the same three parts of the stack that sovereign AI is actually about, whether the sovereignty in question belongs to a country, a company or a single person typing into a chat box.
https://trust-agent.ai/exam/what-is-sovereign-ai-explainedThis is free educational content published by Mickai LTD, not professional, legal, financial or compliance advice, and Mickai accepts no liability for how it's used.
This is the same question Mickai asks about its own stack: who actually controls the data, the model and the infrastructure behind an AI system, and can that answer be checked rather than just taken on trust. Sovereignty, at any scale, means being able to see and verify the answer for yourself, not just being told it.
If your company needs this built properly rather than explained, Mickai builds bespoke AI workflows and code enhancements for businesses, running on Mickai's own sovereign operating system with zero data egress: your data and code never leave your own environment. See what that looks like at mickai.co.uk/bespoke.
Keep learning
The complete workbook
This workbook takes sovereign AI out of government policy documents and puts it in your hands. You will learn what the term actually covers, where your data goes when a tool is free, why it matters who owns the model behind it, how vendor lock-in happens without anyone deciding it on purpose, and a real set of questions you can put to any vendor, whether you are one person or a small organisation.
- 01What sovereign AI actually meansRead here · 1 exercise
The term usually appears in headlines about governments and billion-pound infrastructure. Strip that away and it describes something you can check for yourself, today, about the AI tools already on your phone or your laptop.
- 02Where your data actually goes when a tool is freeRead here · 1 exercise
Free does not mean nobody is being paid. It usually means your input, in some form, is part of the arrangement. This chapter teaches you how to actually check that, instead of guessing.
- 03Who owns the model, and why that mattersRead here · 1 exercise
Two AI tools can look identical on screen and sit on completely different foundations underneath. Ownership decides who can change the rules, and who you would have to trust if things went wrong.
- 04The lock-in trap: how switching gets harder over timeRead here · 1 exercise
Nobody usually decides on day one to become dependent on a single AI vendor. It happens gradually, through small choices that each felt reasonable at the time. This chapter shows you the pattern early enough to still act on it.
- 05Questions you can actually put to a vendorRead here · 1 exercise
This is the practical core of the workbook: a real set of questions, worded so you can send them or check for the answers yourself, whether you are one person or a small organisation choosing a vendor.
- 06What you can do nowRead here · Reading
You now have a working definition of sovereign AI that applies at your own scale, not just a government's, and a real set of questions to use whenever you pick up a new AI tool.
Also inside: a 8-point checklist, a glossary of 12 terms and 10 questions and answers to test yourself. 5 hands-on exercises, each with a worked answer at the back where the workbook gives one.
The PDF is free. We ask for your name and email before downloading. Read the privacy notice at the download form. Free to download and use for personal learning, study groups and inside your own team. Please do not resell the workbooks or republish them as your own. Link people to trust-agent.ai instead.
Get your free PDF workbook
Enter your name and email, then download the complete workbook. This form also works without JavaScript.
Test yourself
Questions and answers
What does sovereign AI actually mean?
Having real control over the data, the model and the infrastructure behind an AI tool, not just access to it. It is usually described at government scale, but the same three questions apply to a single person choosing an app.
Do I need to own my own servers to have sovereignty over my AI use?
No. What matters is whether you can see, check and leave if you need to. A rented tool can still leave you in control, and a tool you technically own can still leave you locked in if you never checked its terms.
Does a free AI tool mean my data is being sold or misused?
Not necessarily, but free tools do have to cover their costs somehow, and your input being used to improve future models is one common way. Read the specific tool's terms rather than assuming either the best or the worst.
How do I find out where my data actually goes?
Read the tool's privacy policy and terms of service, usually linked at the bottom of its website or inside account settings, and look specifically for whether your input trains the model, whether it crosses borders, and how long it is kept.
What is the real difference between an open-weight model and a closed model?
A closed model is only reachable through its owner's own service and cannot be inspected. An open-weight model has its trained parameters published so others can download, run and independently test it, under a stated licence.
Is an open-weight model automatically the safer or more trustworthy choice?
No. Being downloadable is one axis of control, not proof of the data it was trained on or how well it was tested. It tells you who can run it, not whether you should trust it.
What is vendor lock-in, in plain terms?
Being stuck with one AI supplier because switching away would cost more time, money or lost work than you are willing to pay, usually built up gradually through convenience rather than a single decision.
How do I know if I am already locked into a tool?
Ask whether you could leave within a week, if the vendor changed its terms tomorrow, without losing anything that matters. If the honest answer is no, you are already locked in, whatever the tool's other strengths.
What questions should I actually ask an AI vendor?
Whether your input trains their models by default, where your data is stored, whether you can export it in a usable format, what happens if you stop using the service, and whether the underlying model could be run elsewhere if you needed to.
Do I need to be a large organisation for a vendor to answer these questions?
No. Most answers are already published in a privacy policy or terms page. For anything not published, a short, direct email is a reasonable request from a single person or a small organisation, and worth keeping in writing.
Write as you go
My notes
Use this scratchpad for your own notes, working or answers as you read. Notes are saved in this browser on this device. Downloading makes a local copy; it does not send your notes to Mickai.
Consolidate your learning
Review your work
Work through the exercises and questions, then compare your reasoning with the worked answers. A scored exam and verifiable certificate are not yet available for this course.
Review the questionsLearn the language
Key terms
- Sovereign AI
- Having genuine control over the data, the model and the infrastructure behind an AI tool, not just the ability to use it.
- Data residency
- Where your data is physically stored and processed, and which country's laws apply to it as a result.
- Data sovereignty
- The principle that data is subject to the laws of the country in which it is collected, stored or processed.
- Vendor lock-in
- Being stuck with one supplier because switching away would cost more time, money or lost work than most people are willing to pay.
- Model ownership
- Who legally controls an AI model, including the organisation that trained it and whether anyone else can inspect, copy or run it.
- Open-weight model
- An AI model whose trained parameters are published so that others can download and run it themselves, under a stated licence.
6 of the workbook's 12 terms. The complete glossary is in the workbook.
Follow the evidence
Sources and checks
Facts last checked: .
These workbooks use AI assistance. See how the workbooks are made.
- What is sovereign AI?McKinsey & Company
- International transfersInformation Commissioner's Office (ICO)
- UK GDPR, Article 20: Right to data portabilitylegislation.gov.uk (The National Archives)
- AI Act (includes the application timeline)European Commission, Shaping Europe's digital future
Created by Mickarle Wagstaff-Irons - Micky Irons with the Mickai team. Published by Mickai LTD. Last updated 29 September 2026.
NextKeep going
Where to go next
Recommended for you
Privacy and data minimisation in AI applications
Keep personal data under control in an AI feature: UK GDPR principles, DPIAs, roles and transfers, and a redaction pipeline you run and test to see what it misses.
Recommended for you
Open model licences, datasets and deployment obligations
Decide whether you may use, host and redistribute an 'open' model or dataset: licence families, dataset rights, model cards, AI-BOMs, hash checks and a checklist tool.
Recommended for you
Running open-weight AI models on your own computer, for free
A hands-on guide to running open-weight AI models on your own computer at no subscription cost, using Ollama, LM Studio or llama.cpp, with honest hardware limits.