LOKALIT

Staff are already pasting documents into AI chats. What it costs and what to do about it

· 3 min read · LOKALIT

It takes just one situation: someone at the front desk wants to reply to a letter faster, so they paste it into a free chat in the browser. The letter contains a first name, surname, address and PESEL national ID number. Nobody means any harm, and the data has just left the institution.

Figures from the IBM study

The IBM Cost of a Data Breach 2025 report looked for the first time at how breaches are affected by AI tools that staff use without the employer's approval1.

20%of organisations had a breach involving AI tools used without the employer’s knowledge
+USD 670kaverage extra cost of such a breach
63%of organisations have no AI governance policies

Source: IBM, Cost of a Data Breach 2025

The average cost of a data breach worldwide in this study was USD 4.44 million2. In breaches involving unauthorised tools, personal data was leaked more often (65% of cases compared with 53% on average)1. Among organisations that had an incident with AI tools, as many as 97% lacked proper access controls3.

Poles trust it less, but use it anyway

According to a global study by KPMG and the University of Melbourne, only 41% of Poles say they trust AI, less than the global average (46%). At the same time, 77% of respondents accept its presence in everyday tasks4. The same study shows that people use it regularly at work, mostly without any training and often without the employer's knowledge5.

An obligation few people know about

Since 2 February 2025, Article 4 of the EU AI Act has applied. Organisations that use AI systems must ensure their staff have a sufficient level of knowledge about these tools6. This also applies to a public office or cultural centre where someone writes letters with the help of a chat every day.

The tools themselves can have holes too

In June 2025 researchers disclosed the EchoLeak vulnerability in Microsoft 365 Copilot. A single crafted email was enough for the assistant to extract data from the user's mailbox and documents without their knowledge. The flaw was rated 9.3 out of 10; Microsoft fixed it, and there are no reports of it being exploited in practice7. It is a reminder that a tool with access to the whole mailbox and all documents is a new kind of risk.

Four steps that really help

  1. An AI usage policy. Short and specific: which tools are allowed and what must not be pasted (personal data, children's data, HR documents).
  2. Team training on real tasks, with certificates that document the AI Act obligation.
  3. The right tools for the right data. Business versions with a data processing agreement for ordinary tasks, and for sensitive data local AI running on your own hardware, without sending anything to the cloud.
  4. A register of tools. A list of what the team uses, so you know where the data goes.

What the training looks like and what it covers is described in LOKALIT services.

Sources

  1. 2025 Cost of a Data Breach Report: Navigating the AI rush without sidelining security, IBM
  2. Average global data breach cost now $4.44 million, Help Net Security, 4 August 2025
  3. IBM: Shadow AI breaches cost $670K more, 97% of firms lack controls, VentureBeat
  4. Zaufanie do AI w Polsce niższe niż na świecie, KPMG (in Polish)
  5. Korzysta z AI regularnie, przeważnie bez żadnego przeszkolenia i często bez wiedzy pracodawcy, KPMG (in Polish)
  6. Regulation (EU) 2024/1689 (AI Act), EUR-Lex
  7. First-ever zero-click attack targets Microsoft 365 Copilot, CSO Online

Read also