Your AI Agent Reads About People Who Never Chose It. Ireland Just Drew a Line
Opinion & Thought Leadership

Your AI Agent Reads About People Who Never Chose It. Ireland Just Drew a Line

6 min read
Share:

On Monday, Ireland's Data Protection Commission asked X to stop training its AI models on data that X's AI agent had been collecting about other people, and X confirmed it had done so. Commissioner Dale Sunderland told POLITICO that the regulator had raised "significant concerns" in recent weeks: "We asked them to suspend the processing, and they've confirmed to us today that they've done that." The agent, Grok Bot, had started gathering this data in late October.

No fine has been reported and the DPC has not published a formal decision. That makes it easy to file under "Big Tech problem". I think that would be a mistake for any company that is now connecting AI agents to mailboxes, shared drives and CRMs, because the case is about a kind of data almost every such agent touches.

The data in question belongs to someone else

POLITICO describes the shift plainly: regulators are moving their attention from AI training on data people share about themselves to "third-party data that personal AI agents are accessing and gathering on other people". The example POLITICO gives is ordinary: holiday photos shared with a friend who uses an AI agent. The category also covers personal data an agent scrapes from the web.

Sunderland was careful to say the DPC has no issue with AI agents "per se". The concern is what happens next, when a provider moves into "further taking of that data and using it to further train the models". He also said the DPC is talking to another, unnamed tech company about similar plans, before any processing has started.

The DPC matters here because, as POLITICO notes, it enforces the GDPR on companies with their European headquarters in Ireland, including Apple, Google, Meta, X and OpenAI. In practice, what it objects to tends to shape what those providers offer across the EU, Bulgaria included.

Why the usual vendor question misses this

When a company in Bulgaria evaluates an AI tool, the privacy conversation usually ends with one question: do you train on our data? For business plans the answer is generally reassuring. OpenAI's enterprise privacy page says it does not train on data from ChatGPT Business, Enterprise or the API by default, and that it does not train on data accessed through connected apps by default either.

Look at what an agent actually reads, though. An agent that sorts a sales manager's inbox reads mostly about customers. An agent that screens CVs reads about candidates. An agent that summarises supplier calls reads about people at the supplier. In each case the personal data belongs to people who never chose the tool and usually do not know it exists. For them, "our data" is the wrong frame. The data is theirs, and your company is the one that decided to put it in front of an AI system.

That is also how the GDPR sees it. Where personal data is not collected from the person it describes, the controller still owes them information about the processing under Article 14 of the GDPR. Whether a model provider may reuse such data for training is exactly the kind of question the European Data Protection Board addressed in its Opinion 28/2024, which deals with legitimate interest as a legal basis for developing AI models and with the consequences of a model being built on unlawfully processed data.

Where the real exposure usually sits

In my reading, business plans with a signed data processing agreement are rarely the weak spot. The weak spot is the agent an employee connects on their own, with a personal account, because it saves them an hour a day.

Consumer settings are designed around the account holder, and they change often. This week is a good illustration. Google announced on 16 November that Gemini app users will be able to keep their chat history while opting out of model training, which until now required turning history off. Ars Technica points out that training remains opt-out, that the change applies only to future chats, and that Google says it may keep data used to improve its models for up to three years. It is a sensible improvement for the individual user. It also shows that on a consumer account, whether a client's email ends up in a training set depends on a toggle one employee may never have opened.

A hypothetical example

Take an imagined accounting firm in Varna with 40 people. A senior accountant connects her work mailbox to a personal AI assistant so it can sort incoming documents and draft replies. Her inbox contains payroll files from client companies: names, salaries and sick-leave notes for people employed by those clients.

Nobody at the firm has decided anything. The firm's contract with its clients says nothing about AI. The assistant runs on a consumer plan whose training setting is on by default. If that provider ever does what X did, data about employees of the firm's clients, people two steps removed from the firm, becomes training material. The firm will have no record of it, no agreement covering it, and no way to show a client or the Commission for Personal Data Protection what happened.

Nothing in this example requires bad intent. It only requires an agent with access and no one asking where its inputs go.

What I would check this month

I expect the Irish case will not stay limited to platforms that build their own agents. Once regulators frame agent-collected data as a separate category, the natural next question is aimed at the companies that let agents into their systems. A short internal check now is cheaper than answering that question later.

Start with a simple inventory: which AI assistants and agents have access to company mailboxes, calendars, drives or the CRM, and through which accounts. Personal accounts connected to work data deserve their own line in your list of AI systems, with a named owner.

For each one, look at the provider's terms for data the agent accesses, not only for data you type in. Some vendors make separate commitments for connected apps and agent actions; some say nothing. If the answer is unclear, treat it as yes.

Then decide where the line is. For many companies the practical rule will be that agents touching client, candidate or employee data run only on business plans covered by a data processing agreement, with training on customer data switched off.

None of this is new law. It is the GDPR applied to a new kind of tool, and it fits naturally into the AI governance work many companies are starting anyway because of the AI Act. The difference after this week is that a European regulator has said out loud that the data an agent gathers about other people is its own question. It is better to have an answer ready before someone asks.

Frequently asked questions

What did the Irish Data Protection Commission do about X's AI agent?

The DPC raised significant concerns about X using data collected by its Grok Bot agent about other people to train AI models, and asked X to suspend that processing. Commissioner Dale Sunderland told POLITICO on 17 November 2026 that X had confirmed it stopped. No fine or formal decision has been reported.

What is third-party data in the context of AI agents?

It is personal data about someone other than the agent's user, for example holiday photos a friend sends to a person who uses an AI agent, or personal data an agent scrapes from the web. In a company it typically means data about clients, candidates, suppliers or their employees that an agent reads in mailboxes, documents or the CRM.

Do business plans of ChatGPT or similar tools train on company data?

OpenAI states that by default it does not train on data from ChatGPT Business, ChatGPT Enterprise or the API, including data accessed through connected apps. Terms differ between vendors and between business and consumer plans, so each tool connected to company data should be checked separately.

Why are personal AI accounts a risk for companies?

When an employee connects a work mailbox or drive to an AI assistant on a personal account, the company usually has no data processing agreement, no record of the processing and no control over training settings. Consumer settings such as training opt-outs are designed for the account holder and can change, as Google's 16 November update to Gemini's history and training controls shows.

What should a company check first?

Start with a list of AI assistants and agents that can access company mailboxes, calendars, drives or the CRM, and through which accounts. For each, check the provider's terms for data the agent accesses, assign an owner, and decide whether agents handling client, candidate or employee data may run only on business plans with a data processing agreement and training switched off.

Related Articles

Forward-Deployed Engineers: Europe's AI Gap Is Skills, Not Models

Hg and Anthropic are backing a new European company built on forward-deployed engineers. Eurostat data shows why: lack of expertise, not models, is what stops most companies from using AI.

Nov 4, 2026

AI Innovation: Reid Hoffman’s Call for Silicon Valley Action

Discover Reid Hoffman's vision for AI innovation as a tool for good, urging Silicon Valley to act. Learn actionable insights and explore our AI services.

Jan 13, 2026

Custom AI Agents: Don’t Copy Chatbot Answers

Explore the role of custom AI agents in communication and learn proper etiquette for sharing AI outputs. Understand how to add value with human context and discover Encorp.ai's solutions.

Nov 3, 2025