What information you can put into an AI tool (and what you can't)

The most common question when an organisation starts using AI is a very specific one: can I paste this here? The answer depends on two things: the type of information and the type of tool. With a clear table, most doubts can be resolved without asking anyone.

Three types of tool

• General-purpose tool: a public assistant, or one without a contract with the organisation, such as the free version of a chatbot. There are no guarantees about where the data is processed or whether it is used to train the model. Only public information is allowed.
• Corporate tool under contract: a contracted assistant, with corporate accounts, a data processing agreement, a commitment not to train on the data, processing in the EU and, in the public sector, certification under Spain's National Security Framework (ENS). For example, Microsoft 365 Copilot with a business licence. Internal information is allowed and, under certain conditions, pseudonymised personal data.
• Specific assessed system: a system for a specific use case that has gone through a full assessment (registration record, impact assessment where applicable, security and authorisation). The data provided for in its authorisation is allowed, including personal data and, if approved, sensitive data.

An important nuance: a tool under contract that processes data outside the EU involves an international transfer. It does not allow personal data without appropriate safeguards, an impact assessment and a report from the data protection officer.

The table

Information	Examples	General-purpose	Corporate under contract	Assessed system
Public	Legislation, published information, press releases, open data	Yes	Yes	Yes
Internal	Procedures, technical inventories, drafts without personal data	No	Yes	Yes
Personal data	Customers, employees, citizens with name or ID number	No	Only pseudonymised and if the use is authorised	Yes, if provided for and assessed
Special categories	Health, minors, victims, disciplinary proceedings	No	No	Only with an impact assessment and a favourable report from the DPO
Confidential or protected	Bids in a tender, security plans, passwords, privileged legal reports	No	No	No, unless specifically authorised

Rule of thumb: if you would not email that information to an external person, do not put it into a non-corporate tool.

Quick examples

• Summarising published legislation with a general-purpose assistant: yes.
• Pasting the facilities maintenance sheet to prepare a report: only in the corporate tool.
• Pasting a customer list to look for duplicates: no, except in a system assessed for that purpose.
• Asking for a report based on an employee's health file: no, except in an assessed system with an impact assessment.
• Uploading the bids received in a tender to compare them: no. They are confidential and, in addition, evaluating them is the job of the tender evaluation committee.

Basic rules for AI users

1. Use AI as support, never to replace professional judgement or a signature.
2. Use only authorised tools, with corporate accounts and organisation devices.
3. Never enter credentials or passwords.
4. Do not send or sign anything generated with AI without having read it in full.
5. Complete the basic training before using it.
6. Report any incident or possible data leak immediately.

How to verify what the AI returns

• Check data, figures, dates, names and legal references against the original source: AI can invent articles or court rulings.
• Check that no personal data has slipped in that should not be there, taken from another document.
• Be wary of very confident answers about things the tool cannot know: internal data it has not been given or very recent legislation.
• State, where appropriate, that a document has been prepared with the support of AI.
• Report any incorrect, biased or offensive answer.

What to do if something goes wrong

If a tool shows another person's data, someone enters confidential information where they should not, or the AI follows instructions hidden in a document:

1. Stop: stop using it for that task and do not forward the output.
2. Preserve the evidence: screenshot, prompt, response, tool, date and time.
3. Report it immediately through the incident channel.
4. Alert the person in charge if the output has already reached a customer, a citizen or a case file.

Those in charge assess the severity, contain the problem and notify when required: a personal data breach that poses a risk must be reported to the Spanish Data Protection Agency (AEPD) within 72 hours at most.

Questions staff ask

Can I connect the corporate assistant to a folder with confidential information? By default, no. Moreover, with tools such as Microsoft 365 Copilot there is no need to “connect” anything: they can already read everything the person can see. If you have access to ongoing bids, privileged legal reports or security plans, ask for them to be protected with sensitivity labels or excluded from search, and do not ask for summaries based on those documents.

Can I ask it to summarise a case file containing a person's data? With a general-purpose tool, no. With the corporate one, only if that use is authorised and with the minimum data necessary. If it involves health, minors or social services, only in a system assessed for that purpose.

Can I use a free chatbot on my phone to draft a report? No. The rules require authorised tools, corporate accounts and organisation devices. If the report only uses public information, use the authorised general-purpose tool from your work computer.

Can AI reply directly to a complaint? Not directly. It can prepare a draft, but a person reviews it, corrects it and signs it, because the reply has consequences for the complainant.

Can I record a meeting where people are discussed and summarise it with AI? Only under certain conditions: notify participants at the start, use the corporate tool, restrict access to the summary and set how long it will be kept. If health, sanctions or performance reviews are discussed, it is better not to record.

Training by role

No one should use AI without basic training; moreover, the EU AI Act requires AI literacy measures. A simple plan:

Role	Training	Indicative duration	Content
All staff	Responsible use of AI	1–2 h	What AI is and its limits, what information can be entered, verification, authorised tools, incidents
Users of corporate assistants	Practical day-to-day use	4 h	Email, documents, presentations, meetings, how to ask well and how to verify
Supervisors and managers	Overseeing AI systems	3–4 h	Automation bias, test benches, incidents, lifecycle
Technical, security and data protection staff	Governance and compliance	6–8 h	Impact assessments, security, agents and permissions, contract clauses

How to write your own policy

1. Inventory the tools already in use, authorised or not. There are usually surprises.
2. Classify them into the three types (general-purpose, corporate under contract, assessed system) and check where they process data.
3. Adapt the table in this article to your information categories and your real examples.
4. Set the basic rules and the incident channel.
5. Publish the list of authorised tools and the procedure for requesting a new one.
6. Train before you enforce.
7. Review it every year or whenever a relevant new tool or regulation appears.

Types of incident worth recognising

• An incorrect answer with consequences: a deadline, a requirement or an amount wrongly stated to a customer or a citizen.
• A biased, discriminatory or offensive answer.
• Data exposure: the tool shows another person's data, or someone has entered personal or confidential information where they should not have.
• Anomalous behaviour: the AI follows instructions hidden in a document or performs actions that were not requested.
• Unauthorised use: a tool that is not on the list or a use outside the approved purpose.

Every incident is logged, its cause analysed and corrected so that it does not happen again: adding the case to the test bench, adjusting a permission or reinforcing training.

The table only works with technical controls

The rules are reinforced with data loss prevention, blocking of unauthorised tools and sensitivity labels. If you are going to roll out Copilot, also read how to prepare the organisation before enabling it.

At Bidatia we help organisations draft and implement AI use policies tailored to each one, with their technical controls and training.