Call · 30 min
GlossaryMattia Esposito8 November 20267-minute read

Shadow AI. The AI your team already uses, with no rules on paper.

Shadow AI is the use of artificial intelligence tools by employees and contractors without the company having chosen them, approved them or even known about them. This page also has a one-page AI use policy you can adopt.

In brief

The risk is the data that leaves: whatever gets pasted into a personal account leaves the company without a contract, and on individual plans the provider may use it to train its models.

It can be measured. In the 2024 Work Trend Index by Microsoft and LinkedIn, 78% of people who use AI at work bring their own tools; in small and medium-sized companies the figure rises to 80%.

The practical answer is a one-page policy: which tools, which data never goes in, who signs off before a text goes out, what customers are told. The template downloads in Word and PDF, in Italian and English, with a filled-in example and no email required.

Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.

This entry is part of the AI and automation glossary. Here the definition opens up: the template to adopt, where the habit comes from, the risks with their sources, what the AI Act asks and what to do short of banning everything.

The AI acceptable use policy template, to download

A single page, to fill in and have signed by everyone who uses AI for work. Six sections with suggestions in grey, to replace with your company's own choices, and an acknowledgement line. The filled-in example completes the same sections for a made-up services company with 14 people.

Filled-in example of an AI acceptable use policy for a made-up services company
FileContentsLink
TemplateWord, to fill in

The six sections with suggestions, and the signature table.

regole-uso-ai-modello-it.docx

TemplateWord, to fill in, in English

The same page in English, as an AI acceptable use policy, for a parent company or a foreign partner.

regole-uso-ai-modello-en.docx

TemplatePDF

The Italian page, ready to print.

regole-uso-ai-modello-it.pdf

TemplatePDF, in English

The English page, ready to print.

regole-uso-ai-modello-en.pdf

ExamplePDF, filled in, in Italian

Gamma Servizi srl, an invented name: building services maintenance, 14 people, two approved tools.

regole-uso-ai-esempio-it.pdf

ExamplePDF, filled in, in English

The same example, in English.

regole-uso-ai-esempio-en.pdf

Each section matches a decision that, in many companies, every person currently makes alone. Written down, they become eight rows anyone can read in two minutes. The full policy for a company of 10 to 50 people, with the notice to workers, monitoring and suppliers, is in the company AI use policy template.

SectionWhat you decide
Header

Who is responsible for the policy, and when it gets reviewed: every six months, because the tools change before the rules do.

Approved tools

Which tools, on which account, for which tasks and for whom. Company accounts only.

Data that never goes in

Personal data about customers, staff and suppliers, confidential documents, passwords, information covered by a confidentiality agreement.

Before a text goes out

Who reads and approves texts to customers, suppliers and public bodies, and who checks the figures against the source.

What customers are told

When you disclose that a reply or a text comes from a system.

Training and record

The date of the session on the tools in use, and where the record is kept.

If something goes wrong

Who to tell, straight away. If personal data got out, the controller has 72 hours to assess notifying the Italian data protection authority, the Garante.

Acknowledgement

Name, role, date and signature of everyone who has read it.

What is shadow AI

IBM defines it as “the unsanctioned use of any artificial intelligence (AI) tool or application by employees”, without the formal approval or oversight of the IT department. The most common example is a generative assistant used on a personal account to rewrite an email or summarise a document.

The term comes from shadow IT: software used at work without the IT department knowing, such as a personal cloud drive or a downloaded program. With shadow AI, what changes hands is different: texts, documents and data pasted into an outside tool. Microsoft calls the same behaviour BYOAI, bring your own AI: everyone brings their own tool to work.

Why it happens

It happens because use runs ahead of the rules. The 2024 Work Trend Index by Microsoft and LinkedIn, a survey of 31,000 knowledge workers in 31 markets carried out between February and March 2024, found that 75% use generative AI at work and that 78% of those users bring their own tools. Only 39% had received AI training from their company.

The same study measures the silence: 52% of people who use AI at work are reluctant to admit it for their most important tasks. So the first step of any policy is to ask who uses what, in a way that doesn't sound like an investigation.

In Italy, companies that have adopted AI are still a minority. According to Istat, 16.4% of Italian companies with at least 10 employees used at least one AI technology in 2025, up from 8.2% in 2024. The two figures count different things, Italian companies on one side and workers in 31 markets on the other; side by side, they suggest that in many businesses AI comes in through people before it comes in through a decision.

The real risks, one by one

Shadow AI carries four risks: data that leaves, no contract with the provider, errors that reach customers, and breaches the company finds out about too late. The most concrete is the first, and it depends on the plan: OpenAI says that on its services for individuals (Free, Plus, Pro) “we may use your content to train our models”, unless the user turns this off in settings.

RiskWhat happens
Data used for training

On individual plans, content can be used to train models. On OpenAI's business plans (ChatGPT Business, Enterprise, API), by default it isn't.

No contract covering the data

Article 28 of the GDPR requires anyone processing data on the company's behalf to be bound by a contract and to give sufficient guarantees. A personal account has neither.

Fines

For controller and processor obligations, up to €10 million or 2% of worldwide annual turnover; for the core principles and transfers to non-EU countries, up to €20 million or 4% (Article 83 of the GDPR).

Breach discovered late

The controller must notify a personal data breach within 72 hours of becoming aware of it (Article 33). If nobody knows the tool exists, awareness comes late.

Errors that reach customers

A generated text can contain false facts stated with confidence, known as hallucinations, and the person who sends it is responsible for it.

Cost of a breach

In the breaches studied for IBM in 2025, organisations with high levels of shadow AI paid on average $670,000 more than the others.

That figure comes from the Cost of a Data Breach 2025 report, carried out by the Ponemon Institute for IBM across 600 organisations that suffered a data breach between March 2024 and February 2025. One in five was breached because of shadow AI, and only 37% have policies to manage AI or detect unsanctioned use.

It's a study sponsored by a company that sells security, mostly covering large organisations: read it as an order of magnitude, not as a forecast for a small business.

What to do without banning everything

Four steps, in order: ask, choose, write, train. A ban with no alternative risks pushing use onto personal phones, where nobody sees it, and leaves the work as it was.

Ask. A round of questions, face to face or with an anonymous form, on which tools people use and for what. With more than half of users reluctant to admit it, the question has to make clear there are no consequences.

Choose. One or two tools on a business plan. OpenAI, for instance, says that by default it doesn't use data from ChatGPT Business, Enterprise or the API for training. With any provider, read that condition in the contract before you sign.

Write. The one-page policy in the template: approved tools, data that never goes in, who signs off, what customers are told. The point where a person approves before anything is sent is human oversight, and it gets decided text by text.

Train. A session on the chosen tools, covering the real risks, and a line in the record. It's also the simplest way to show the measures the AI Act asks for.

Shadow AI and the AI Act

Article 4 of the AI Act requires anyone using AI systems for work to look after their staff's skills, and it has applied since 2 February 2025. In the text replaced by Regulation (EU) 2026/1744, deployers “shall take measures to support the development of AI literacy of their staff”.

The European Commission, in its AI literacy questions and answers, covers exactly the case of staff using ChatGPT to write advertising copy or translate: they should be told about the specific risks, such as hallucination. No certificate is needed, an internal record is enough, and national authorities supervise from 2 August 2026.

A company that doesn't know which tools are in use can't tailor those measures to the systems actually running. That's why shadow AI touches the AI Act even when no tool is high-risk. The obligation is covered in detail on the page about mandatory AI Act training.

Can I use ChatGPT at work?

It depends on two things: your company's rules and the data you put in. Where there are no rules, the safe choice is to ask in writing before pasting in data about customers, colleagues or suppliers, and to use a company account rather than a personal one.

For the people writing the rules it works the other way round: if the company says nothing, every person decides alone what is confidential and what isn't. A one-page policy takes that decision off the shoulders of the people doing the work, whether in a professional practice, a workshop or a sales office.

Related terms

Human in the loop

The point where a person approves, corrects or stops what the system proposes. In an AI use policy, it's the section on who signs off before anything is sent.

AI hallucinations

A false statement produced with confidence by a model. It's the risk that reaches the customer when nobody rereads.

LLM

The language model behind generative assistants. Whatever you type travels to the provider running it.

Data processor

Anyone processing personal data on the company's behalf, bound by a contract. A personal account on an outside tool isn't one.

Questions and answers

What is shadow AI?

It's the use of artificial intelligence tools by employees and contractors without the company's approval or oversight, for example a generative assistant used on a personal account.

IBM defines it as the unsanctioned use of AI tools or applications, without the formal approval of the IT department.

What's the difference between shadow AI and shadow IT?

Shadow IT is any software or service used at work without the IT department knowing. Shadow AI is the case where the tool is an artificial intelligence.

The practical difference is what leaves: texts, documents and data pasted into an outside tool, which on individual plans the provider may use to train its models.

What are the risks of shadow AI?

Data leaving the company without a contract, none of the guarantees Article 28 of the GDPR requires from anyone processing data on the company's behalf, errors reaching customers and breaches discovered late.

The GDPR requires a breach to be notified within 72 hours of becoming aware of it, with fines of up to 10 or 20 million euros depending on the obligation breached.

Can I use ChatGPT at work?

It depends on your company's rules and on the data you put in. Where there are no rules, ask in writing before pasting in data about customers, colleagues or suppliers.

On business plans OpenAI states that by default your data isn't used to train its models; on a personal Free or Plus account it can be, unless you turn it off in settings.

Does the AI Act require internal rules on AI use?

Not as a document with that name. Article 4 requires anyone using AI systems for work to take measures on their staff's skills, from 2 February 2025.

The European Commission says an internal record of the initiatives is enough. A signed one-page policy, with the date of the session on the tools, is a simple way to show them.

Notes on sources

  1. The definition comes from IBM, What is shadow AI?, read on 8 November 2026; the distinction from shadow IT comes from the same page.
  2. The 75%, 78%, 80% in small and medium-sized companies, 52% and 39% come from the 2024 Work Trend Index by Microsoft and LinkedIn, 8 May 2024: 31,000 knowledge workers in 31 markets, surveyed from 15 February to 28 March 2024. They are global figures, not Italian ones.
  3. The share of companies using AI comes from Istat, Imprese e ICT (ICT in enterprises), 2025, covering companies with at least 10 employees.
  4. The use of content for training comes from OpenAI, How your data is used to improve model performance, and from the Enterprise privacy page updated on 8 January 2026, both read on 8 November 2026. Providers' terms change: check them again in the contract.
  5. Articles 28, 33 and 83 of Regulation (EU) 2016/679, the GDPR, read in the official text.
  6. The cost and shares on shadow AI come from IBM's press release of 30 July 2025 on the Cost of a Data Breach 2025 report, carried out by the Ponemon Institute and sponsored by IBM, across 600 organisations with a breach between March 2024 and February 2025.
  7. The Article 4 quotation comes from the new text introduced by Regulation (EU) 2026/1744, amending Regulation (EU) 2024/1689, read in the official English version; the ChatGPT case, the internal record and the 2 August 2026 date come from the European Commission's AI literacy questions and answers, updated on 20 October 2026.
  8. This page is a survey, not legal advice: which data can be handled with which tool is for the people who advise your company to decide.
·The next step

First you decide what AI does. Then you build it into the work.

The first step with Itria is a thirty-minute video call: we look at which steps in your work an AI system can prepare, with which tool, and which person approves before anything is sent. Drop us a line about what's slowing you down. We'll make the first move: we'll look at what a customer sees when they search for you, and tell you what we found. Even if we never end up working together.