Data protection in AI use

1. Introduction: the easiest AI risk is also the most mundane

AI use rarely starts with a grand strategy. It starts with a small, everyday moment.

An employee is in a hurry. In front of them is a long customer contract, a messy memo, a complaint thread or an Excel sheet full of customer data. They open an AI tool and type: “Summarize this.” Then they paste in the entire document.

The work gets done faster. The moment feels innocent. That is exactly why it is dangerous.

Data protection in AI is not about AI being inherently bad or about organizations needing to fear everything new. It is about AI making data processing so easy that the old routines of caution cannot always keep up.

In the past, a confidential document had to be sent by email, saved in a system or shared through access rights. Now it can be copied in seconds into a chat box backed by an external provider, cloud infrastructure and possibly several technical subcontractors.

That is why data protection in AI use is one of the most important leadership questions of the AI era. Not just a matter for the legal department. Not just for IT. Not just for the data protection officer. It is an everyday way of working.

2. What does data protection mean in AI use?

Data protection in AI use means that an organization understands and controls what personal data and confidential information is processed in AI systems.

Personal data means any information from which a person can be identified, directly or indirectly. A name and an email address are obvious examples. But a customer number, an IP address, location data, a voice recording, a photo, health data, an employee evaluation or a combination of different data points can also be personal data.

In practice, data protection in AI use comes down to at least these questions:

What information may be entered into AI?
Where is the data processed?
Is the entered data stored by the provider?
Is the data used to train models?
Who can access the data?
Can the data be transferred outside the EU?
Is there a legal basis for the processing?
How are data subjects’ rights fulfilled?
How are errors, leaks and misuse detected?

So data protection does not just mean “we don’t publish customer data.” It means managing the entire data lifecycle. Where does the data come from? Where does it go? Who uses it? How long is it kept? Can it be deleted? Can it be traced?

3. Why does this matter right now?

AI has moved from experiments to everyday work. The conversation is no longer just about whether AI can write a poem or create an image. Now it is used in customer service, HR, legal, marketing, software development, healthcare, education and to support executive decision-making. At the same time, regulation is getting stricter and more specific.

The EU General Data Protection Regulation (GDPR) has applied since 2018, and it covers the processing of personal data even when that processing takes place in AI systems. The EU AI Act, in turn, entered into force on August 1, 2024, and its obligations apply in phases. According to the European Commission, the AI Act’s transparency rules take effect in August 2026, and the regulation is built on a risk-based approach.^[1]

In Finland, the Office of the Data Protection Ombudsman has emphasized that when developing and using AI systems, an organization must choose a legal basis for processing personal data that fits the purpose. Its guidance also stresses data accuracy, risk assessment and respect for data subjects’ rights.^[2]

This makes the topic timely for three reasons. First, AI is already used far more than organizations’ official policies often acknowledge. Second, data moves faster and less visibly than before. Third, the consequences of mistakes can be serious: a data leak, a breach of contract, reputational damage, a fine, lost customer trust or a wrong decision about a person.

4. AI, data and trust: what is this really about?

AI needs data. But not all data is the same.

There is public information, such as website copy, product information and open reports. There is internal information, such as strategy documents, sales pipelines and financial forecasts. There is personal data, such as customer lists, employee records and recruitment materials. And there is especially sensitive data, such as health information, children’s data, biometric data, or data revealing trade union membership, religion or political opinions.

To an AI tool, all of these may look like plain text. To an organization, the difference between them is enormous.

If a marketer enters a public product description into an AI tool and asks for a social media version, the risk is usually small. If an HR specialist enters an employee’s sick leave records, performance review notes and a manager’s assessment, it is a completely different situation.

The key question in data protection is trust. A customer, employee or partner gives an organization their data for a specific purpose. They have not necessarily agreed to that data ending up in an external AI service, in a model’s training data or in a system the organization does not properly understand.

AI does not remove this relationship of trust. It makes it more fragile.

5. What information should you not enter into AI?

A good rule of thumb is simple:

Never enter anything into an AI tool that you would not send to an unknown external provider without a contract, approval and a risk assessment.

Particular caution is needed with at least:

personal data
customer data
employee data
health data
children’s data
contracts
trade secrets
internal strategy documents
board and executive team materials
requests for proposals and bids
product development materials
source code
security-related information
unfinished M&A or financing materials
legal assessments
documents covered by non-disclosure agreements

This does not mean AI can never be used with these topics. It means that such use requires an approved solution, contracts, access controls, security, documentation and clear rules.

The difference is significant. Asking an AI tool to “Draft a general template for reviewing the risks in a customer contract” is one thing. Pasting the customer’s entire contract, names, prices, negotiation history and confidential terms into a public AI service is quite another.

6. A prompt is not just a question — it can constitute processing of personal data

Many people think of a prompt, the instruction given to an AI tool, as just text. From a data protection perspective, however, a prompt can constitute processing of personal data.

If you tell an AI tool, “Here is our customer John Doe’s complaint, purchase history and payment details — write a reply,” you are processing personal data in an AI system. If you attach an employee’s performance review, it may involve the employee’s personal data. If you attach a patient record, it may involve health data, which falls under the special categories of personal data.

For many organizations, this is a wake-up call. The data protection risk does not arise only from what AI answers. It already arises from what you give it.

That is why an organization’s AI guidelines should be very concrete about inputs: what may be attached, what may not, when data must be anonymized, when an approved enterprise solution must be used, and when to ask the data protection officer or IT for permission.

Anonymization means that a person can no longer be identified from the data. Pseudonymization, on the other hand, means that identifiers have been replaced with, for example, codes, but the person can still be linked to the data with the help of additional information. This distinction matters, because pseudonymized data is still personal data.

7. The business perspective: efficiency must not override accountability

For companies, the promise of AI is huge. It can speed up customer service, summarize documents, help with sales, run analyses, support software development and free up experts’ time. But it is precisely this efficiency that makes it risky.

When a tool is easy, people use it. If there are no rules, everyone makes their own interpretation. One employee uses AI only for public texts. Another pastes in customer contracts. A third uses it to evaluate job applications. A fourth copies in source code. A fifth uses AI to summarize the executive team’s memo. All of them may think they are acting sensibly.

From the organization’s perspective, this is an uncontrolled data ecosystem. A company’s job is not just to prohibit. A ban alone easily leads to shadow use: people keep using the tools, just quietly.

A better solution is to provide a safe way to use AI. That means approved tools, clear guidelines, training and easy examples.

8. The startup and growth company perspective: speed needs boundaries

For a startup, AI is a tempting tool. A small team can get more done. But a startup’s strength, speed, can also be its weakness. In the early stages there is not always a legal team, a data protection officer or a ready-made governance model.

Even so, customers, investors and partners increasingly ask:

How do you handle data?
Do you use AI?
Is data transferred to third parties?
Is customer data used to train models?
Where does the data live?
Can a customer opt out of the use of their data?
Do you have security and data protection policies?

These are not questions you want to answer for the first time in the final week of an enterprise customer’s procurement process. For a growth company, data protection can be a sales enabler. When things are in order, a large customer feels confident enough to buy.

This is an important insight: data protection is not just a brake. It can be the infrastructure of trust.

9. The individual perspective: the employee is often the first point of risk

The weakest point in data protection is often not a malicious attacker. It is a busy person.

An employee may not know that information entered into a public AI service can be stored by the provider. They may not realize that a customer ID combined with purchase history is personal data. They may not see that a draft contract can contain trade secrets.

That is why employees should not be left to guess. A good organization gives clear rules: these tools are allowed, this data may be used, this data may not be used, this is how you anonymize text, this is what you do when in doubt, this is who you can ask.

Tone matters too. If the guidelines are nothing but scare tactics, people will not learn to use AI productively. If they are too loose, the risks grow. What you need is a practical middle ground: encourage use, but teach the boundaries.

10. The technological perspective: where does data travel and what happens to it?

To the user, an AI service often looks like a single text box. In reality, a complex technical chain may sit behind it.

Data may travel from the user’s device to the provider’s cloud, and from there on to the model, logging systems, monitoring tools and possibly subprocessors’ infrastructure. In some services, inputs may be used to improve the service or train models unless this has been excluded by contract.

That is why an organization needs to find out at least:

Is this a consumer version or an enterprise version?
Are inputs used to train models?
Is chat history retained?
In which country or region is the data processed?
Who are the subprocessors?
Can data be transferred outside the EU?
Can data be deleted?
Is logging enabled?
How are access rights managed?
Is there a data processing agreement for the service?

A data processing agreement, or DPA, is a contract that defines how a provider processes personal data on behalf of an organization. Without such an agreement, putting personal data into the service can be problematic.

11. The legal and ethical perspective: GDPR, the AI Act and accountability

In data protection, the legal foundation is clear: personal data may only be processed lawfully, transparently and for a specified purpose.

The GDPR principles are especially important in AI use: lawfulness, fairness, transparency, purpose limitation, data minimization, accuracy, storage limitation, integrity and confidentiality, and accountability.

Data minimization means processing only the data that is genuinely needed. In AI use, it is an excellent principle. If you want AI to draft a customer message, it does not need to see the entire customer database. If you want to summarize contract risks at a general level, AI does not necessarily need to see the parties’ names, prices or signatories.

The European Data Protection Board (EDPB) issued an opinion in December 2024 on the processing of personal data in the context of AI models. Among other things, the opinion addresses when an AI model can be considered anonymous, how legitimate interest can be assessed as a legal basis, and what follows if a model was developed using unlawfully processed personal data.^[3]

This says something essential: data protection in AI is no longer a gray side issue. It is at the core of European digital regulation.

The ethical question goes even further than the law. Even if something is technically possible and legally justifiable, it is not always sensible or trust-building. A customer will not necessarily start by asking: “Was there a legal basis for this under Article 6?” They will ask: “Why did my data end up in an AI system?” If the answer sounds careless, trust crumbles.

12. Concrete examples: the bad, the better and the excellent way to work

Example 1: summarizing a customer contract

Bad: An employee copies the customer’s entire contract into a public AI service and asks for a summary. The contract contains the customer’s name, prices, confidentiality terms, personal data and negotiation history.

Better: The employee removes names, prices and identifying details, summarizes only the general contract clauses and uses AI to outline the structure. The final assessment is done internally.

Excellent: The organization has an approved enterprise-grade AI solution in which data is not used for model training, usage is logged, contracts are in place and employees have guidelines on which contracts may be put into the system.

Example 2: a customer service reply

Bad: A customer service agent pastes the customer’s full message history, address, payment details and complaint into an AI tool.

Better: The agent writes an anonymized description: “A customer received the wrong product and is asking for a refund. Draft a friendly reply template.”

Excellent: An approved AI assistant is integrated into the customer service system; it processes only the necessary data, respects access rights and produces a suggestion that a human reviews.

Example 3: HR and recruitment

Bad: A recruiter enters applicants’ CVs, names, contact details and interview notes into an AI tool and asks it to pick the best candidates.

Better: The recruiter asks AI to draft a general interview framework based on the role’s requirements, without any applicant personal data.

Excellent: The organization assesses the risks of using AI in recruitment, ensures non-discrimination, documents decision-making, keeps a human accountable for final selections and uses only approved systems.

13. Risks and pitfalls

1. “But this is just a small test”

Many data protection risks start with an experiment. An experiment does not remove accountability. If a test involves personal data, it is still processing of personal data.

2. Using consumer tools for company data

A free or personal AI account can be convenient, but it may not meet the company’s security and data protection requirements.

3. Overconfidence in anonymization

Removing a name is not always enough. If a text contains enough detail, a person or company can still be identifiable. For example, “a 43-year-old CEO based in Northern Finland whose company was sold to a Swedish buyer last year” may be identifiable even though no name is mentioned.

4. Training a model on customer data without a clear legal basis

If an organization wants to use customer data to train an AI model, the questions immediately become more serious: what is the legal basis, is the purpose compatible, how are customers informed, and can the data be deleted?

5. Hallucinations and false conclusions

AI can produce a convincing but incorrect answer. If that answer is based on personal data, the error can affect a real person.

6. Blurred accountability

AI is not accountable. The organization is. If AI makes a flawed recommendation, classifies people in a discriminatory way or discloses data to the wrong person, accountability does not disappear behind the technology.

7. Shadow AI

If an organization does not provide safe tools, employees may use their own. This is one of the biggest risks in practice.

14. The future: toward controlled, safe and useful AI

Data protection in AI use is not going away. On the contrary, it will become part of normal management. In the future, we are likely to see five trends.

1. Enterprise-grade AI solutions become the norm

Organizations move away from ad hoc use of consumer tools toward managed solutions in which access rights, logging, data retention and training use are defined by contract.

2. Data governance moves to the core of AI strategy

Data governance means the model for managing data: who owns the data, who may use it, for what purpose and on what terms. Without data governance, an AI strategy easily remains just a nice slide deck.

3. AI use starts to be audited

Companies are increasingly asked where AI is used and with what data. The questions may come from customers, investors, authorities or the board.

4. AI literacy becomes a basic skill for every employee

AI literacy does not just mean knowing how to write good prompts. It also means understanding the risks, the limits and the accountability.

5. Trust becomes a competitive advantage

Companies that know how to use AI safely and transparently stand out. Customers do not just want efficient service. They want to know their data is safe.

15. Conclusions: what should you take away from this?

Data protection in AI use is, above all, a practical matter.

It is decided in the moment an employee wonders whether they can paste a document into an AI tool. It is decided in the moment leadership chooses whether to buy an approved enterprise solution or let people use whatever they happen to find. It is decided in the moment a startup chooses whether to get its data practices in order right away or only when its first big customer demands it.

The most important lesson is this: AI is worth using, but not blindly.

A good organization does not just say “don’t use it.” It says: “Use it like this.” It provides safe tools, clear boundaries and sufficient training. It recognizes that the benefits of AI only materialize when people feel confident using it the right way.

Data protection is not the opposite of AI. It is a prerequisite for using AI sustainably, in a way that builds trust and strengthens the business.

16. Practical frameworks, checklists and templates

A. The one-minute decision model for employees

Before you paste anything into an AI tool, ask:

Does this contain personal data?
Does this contain customer data?
Does this contain trade secrets?
Does this contain contracts, prices or strategy?
Is this public information?
Is this an approved tool?
Do I know whether the data is stored by the service?
Could I explain to a customer or my manager why I put this into an AI tool?

If in doubt, don’t paste it. Ask first.

B. The green–yellow–red model

Green (generally allowed): public web copy, general ideas, anonymized examples, your own drafts without confidential information, generic message templates, summaries of public reports.

Yellow (requires judgment or an approved tool): internal documents, draft contracts, customer cases without identifying details, sales materials, technical documentation, project plans, anonymized customer feedback.

Red (do not enter without separate permission and an approved solution): personal data, customer lists, health data, employee data, confidential contracts, trade secrets, board materials, M&A documents, sensitive recruitment data, security-related details.

C. A minimum template for an organization’s AI policy

Which AI tools may be used?
For which purposes may they be used?
What data must not be entered?
What data may be entered only into an approved enterprise solution?
How is data anonymized?
Who is responsible for approving AI tools?
How are employees trained?
How is use monitored?
How are errors reported?
How is the policy updated?

D. A data protection checklist for leadership

Has the organization mapped where AI is already being used?
Have approved tools been defined?
Have prohibited uses been defined?
Have employees been trained?
Have provider contracts been reviewed?
Are data processing agreements in place?
Do you know whether inputs are used to train models?
Has a risk assessment been carried out?
Have high-risk use cases been identified?
Are logging and access management in order?
Is there a procedure for security incidents?
Does the board have visibility into AI use?

E. Questions to ask a provider

When evaluating an AI service, ask:

Where is the data processed?
Where is the data stored?
Is customer data used to train models?
Can training use be excluded by contract?
How long are inputs and outputs retained?
Who can access the data?
Which subprocessors are used?
Is data transferred outside the EU or EEA?
Is a DPA available for the service?
How is data deleted?
How are security incidents reported?
Has the service been audited?
Can usage be logged at the organization level?
Can user groups and access rights be managed?
How does the service meet GDPR requirements?

F. A beginner’s model: rolling out AI safely

Choose 3–5 low-risk use cases.
Deploy an approved tool.
Write a one-page guide for employees.
Train people with concrete examples.
Prohibit entering personal data and confidential documents into public tools.
Set up a channel for questions about unclear situations.
Collect feedback and update the guide monthly.

G. A model for an advanced organization

Map how AI is being used.
Classify use cases by risk level.
Create a register of approved tools.
Define data classes: public, internal, confidential, restricted, personal data, special categories of personal data.
Carry out a data protection impact assessment (DPIA) for higher-risk use cases.
Build logging and monitoring.
Define the human role in decision-making.
Document providers and data transfers.
Train employees according to their roles.
Report on AI use to leadership and the board.

H. A practical everyday tip

If you want to use AI safely, first turn the original material into a general description.

Don’t write: “Summarize this contract with customer X, including these prices, people and terms.”

Write instead: “Draft a checklist for assessing the risks of a B2B service agreement. Focus on liability, termination, data protection, confidentiality and price changes.”

This way you get value from AI without needlessly handing it confidential information.

I. The one-sentence summary

Data protection in AI means that an organization is not just excited about what AI can do, but also understands what data may be given to it, on what terms and under whose responsibility.