The software company that created FT Mercati Go to ftmercati.com

Guide · Technology for decision makers

ChatGPT, Copilot and Claude: using AI at work

Can I let staff use ChatGPT, Copilot or Claude? The business version, settings, data not to paste, notice to workers and AI Act obligations.

By FT Support Updated on 18-minute read

The short answer

Yes, you can let your staff use ChatGPT, Microsoft Copilot, Claude and Gemini, on three conditions: company accounts, not personal ones, because in the plans for organisations content does not train the models by default and there is a data processing agreement; written usage rules, with the list of what is never pasted; and a contact person who answers for the choices and explains the risks.

  • Company accounts, not personal ones: in the plans for organisations, content does not train the models by default.
  • In the personal versions of ChatGPT, Claude and Gemini you switch training off by hand, each with its own setting; no personal version comes with a contract with the company.
  • Price lists, quotes, drawings, contracts and employee data are not pasted into a tool without contractual guarantees.
  • A company that makes these tools available is a deployer: measures for AI literacy (Article 4) and transparency (Article 50).
  • A ban does not stop the use: it moves it to personal accounts, where the company sees nothing.
  • Microsoft Copilot works inside the tenant and shows only what the user can already read: the risk is internal sharing that is too broad.

Note. Informational content updated as of 12 October 2026; the providers' terms were reread on 17 November 2026 and change often: before deciding, reread the official pages cited below and, for privacy and employment matters, consult your advisers. The national rules described are those of Italy; other EU member states have their own.

Can I let my staff use ChatGPT?

Yes, and in many companies it is already happening, with or without permission. In Italy, in 2025, 84% of large companies bought licences for at least one generative artificial intelligence (AI) tool, mainly Microsoft Copilot, ChatGPT Plus and Gemini Advanced, that is, largely personal plans; among small and medium-sized enterprises, 9% use paid tools and another 9% free tools (Artificial Intelligence Observatory of the Politecnico di Milano, February 2026). In Italy, 34% of workers use AI tools not promoted by their own organisation (HR Innovation Practice Observatory, May 2026). According to Eurostat, in 2025, 16.4% of Italian enterprises with at least ten employees used at least one AI technology, double the year before; 6.2% used generative AI for language (it was 3.7% in 2024), one in twenty among those with 10-49 employees; in basic metals and fabricated metal products the share is stuck just above 2%, while the European average for the sector has more than doubled.

The clearest public answer comes from one country, Italy: the guidelines of its Ministry of Labour on AI in the world of work (December 2025, non-binding) include a handbook for SMEs on the informed use of generative AI, which says to prefer the professional or business versions, to switch off in personal accounts the option that uses data to improve the model, not to upload confidential content without guarantees, to train staff and to set internal rules. The European Commission, asked whether a company whose employees use ChatGPT to write or translate texts must comply with Article 4 of the AI Act, answers yes: they must be informed about the specific risks, for example hallucinations.

Banning does not work. A ban with no alternative moves the use to personal accounts and to phones, where the company sees nothing: this is shadow AI. There are three conditions for saying yes: company accounts, written usage rules, a contact person who answers for the choices. The rest of this guide explains how.

Personal or business version: what really changes

Four things: whether content trains the models, who administers the accounts, where the data is kept and what the contract says. The comparison below applies the same criteria to each tool, without prices: the figure depends on the number of people and on the plan, and is a matter for a quote.

CriterionChatGPT (OpenAI)Microsoft CopilotClaude (Anthropic)Gemini (Google)
Personal versionsFree, Go, Plus, ProCopilot with a personal Microsoft accountFree, Pro, Max, UltraGemini app
Versions for companiesBusiness (called “Team” until August 2025), Enterprise, Edu; API separatelyCopilot Chat with a work account (included in eligible Microsoft 365 licences); Microsoft 365 Copilot, in two tiers: Basic, with no add-on licence, and Premium, paidTeam and Enterprise (“Claude for Work”); API; also available through Microsoft 365, Amazon Bedrock, Google Cloud and Microsoft FoundryGemini in Google Workspace, included in the subscriptions; separately, Gemini Enterprise, with its own licences
Training with the personal versionYes, until “Improve the model for everyone” is switched offNew app (from 18 August 2026): no, Microsoft states that prompts, responses and files do not train foundation models. Previous version: yes, unless you opt out in Settings, PrivacySince 28 August 2025 the user chooses: 5 years of retention if yes, 30 days if noYes, until training is switched off in the Gemini History settings; some of the chats are read by human reviewers in any case
Training with the business versionNo by defaultNo: requests, responses and tenant data do not train the modelsNo: the commercial terms exclude itNot outside the domain without permission
Where the data is kept for a European customerResidency in Europe for Enterprise, Edu and API, where eligible; Business is not among the eligible plansEU Data Boundary for European tenants, except web searches, Anthropic’s models and, unless turned off, processing during peak load periodsStorage in the United States; in Europe through Bedrock or Google CloudData regions in some editions
RetentionDeleted chats removed within 30 days; in Enterprise the administrator decides itWith the other Microsoft 365 content, under Purview retention policiesTeam and Enterprise: indefinitely, unless a custom period is set in Enterprise (minimum 30 days); deleted chats removed within 30 days; API 30 days3, 18 or 36 months at the administrator’s choice, 18 by default
Company sign-in, administration, logsSSO in Business; SSO, SCIM and compliance logs in EnterpriseMicrosoft Entra account, AI Administrator role, audit logs in PurviewSSO in Team; SCIM, audit log and Compliance API in EnterpriseWorkspace Admin console
Data processing agreement and declared certificationsYes, on request, for Business, Enterprise and API; SOC 2Yes: the same DPA as for email and files in Microsoft 365; ISO 27001, ISO 42001Yes, incorporated into the commercial terms; ISO 27001, ISO 42001, SOC 2Yes, with the Workspace terms; ISO 42001

A word on the providers’ language. The sales pages write “your data is never used to train the models”; the legal pages write “by default”. The second is the one that counts: in personal ChatGPT, for example, a thumbs-up or thumbs-down rating causes the whole conversation to be used even when training is switched off. The data processing agreement (DPA) is the document that turns promises into commitments: it exists in the company plans and not in the personal ones, and it is the first thing to ask for.

ChatGPT, Copilot, Claude and Gemini: two lines each

  • ChatGPT is OpenAI’s assistant. In the versions for organisations (Business, Enterprise, Edu) administrators manage users and access and enable the connections to company data: projects, plugins (from 11 December 2026 they take the place of custom GPTs, which OpenAI is retiring), connected apps (previously called “connectors”) and the “Company Knowledge” plugin, available since October 2025, which answers from the connected sources, respecting each person’s permissions and with citations. You activate the Business plan yourself, with at least two seats, and it includes sign-in with the company identity; Enterprise is bought with a contract and adds centralised control of users, models and tools and a log platform for compliance checks.
  • Microsoft Copilot (called “Microsoft 365 Copilot” until autumn 2026) is three things with the same name: Copilot with a personal account, a consumer product; Copilot Chat with the work account, included in eligible Microsoft 365 licences, which uses mainly web data and attached files; Microsoft 365 Copilot, in two tiers: Basic, with no add-on licence, with Copilot inside the apps but without the tenant’s data; Premium, paid, which works inside Word, Excel, PowerPoint, Outlook and Teams with the data of the tenant read through Microsoft Graph. It shows only the content on which the person already has at least read permission: the risk is internal sharing that is too broad in SharePoint, not Copilot. Anthropic’s models can also work inside Copilot, as a Microsoft subprocessor: for European customers they are disabled by default.
  • Claude is Anthropic’s assistant. The Team and Enterprise plans give an administered workspace, with sign-in with the company identity; Enterprise adds automatic user management, an audit log, custom retention and roles. It can also be used inside Microsoft 365 and through the cloud platforms: Amazon Bedrock, Google Cloud and Microsoft Foundry. For European customers the point to know is that the data is stored in the United States: to keep it in Europe, Claude is used through Bedrock or Google Cloud.
  • Gemini is Google’s assistant. With a personal account it is the Gemini app; with Google Workspace it works inside Gmail, Docs and Sheets with the domain’s content, which is not used to train the models outside the domain without permission, and the administrator decides whether and for how long conversations are kept.

The personal versions differ in one more way: the terms of use. Claude requires users to be eighteen; ChatGPT at least thirteen and, under eighteen, the permission of a parent or guardian. ChatGPT’s European terms make clear that the user owns what they produce, but that the results “may not always be accurate”, and they forbid presenting a generated text as written by a person. These are clauses that a company usage rule must repeat.

The settings for day one

In personal accounts, if they remain in use for a time:

  • ChatGPT: Settings, Data controls, “Improve the model for everyone” switched off, or “Do not train on my content” in OpenAI’s privacy portal; temporary chats do not train the models but are kept for up to 30 days.
  • Claude: the choice on training, asked of all users since 28 August 2025, can be reviewed in Settings, under Privacy (Privacy Settings); with the choice set to no, retention stays at 30 days.
  • Gemini: in the Gemini History settings you switch off the use of chats to train the models. Switching History off does not make them private: chats are still kept for around two months, some are read by human reviewers and the reviewed ones are retained for up to three years. Temporary chats are not used to train the models.
  • Copilot with a personal account: since 18 August 2026 there has been a new version of the app (web, desktop and mobile), for which Microsoft states that prompts, responses and file contents are not used to train foundation models. In the previous version they could be, unless you opted out with the “Training on conversation activity” switch in Settings, Privacy: anyone still using it should switch that off, and sign in to the new app by 31 December 2026 to keep their chats. It remains in any case a different product from the business one and outside the contract: for work, use the work account.

In company plans, on the day of activation:

  1. Sign-in with the company identity (single sign-on): in Microsoft 365 it is the existing work account; ChatGPT Business and Claude Team include it; Claude Enterprise adds automatic user creation. That way access is removed from a single place when a person leaves the company.
  2. Who administers. One name, with a least-privilege role: in Microsoft 365 there is the AI Administrator role, separate from the Global Administrator.
  3. Retention. How long conversations and files are kept: in Microsoft 365 with Purview policies, in Workspace with the administrator’s choice, in Claude Enterprise with custom retention, in ChatGPT Enterprise with the workspace settings.
  4. The logs. Every interaction with Copilot generates an audit record in Purview; ChatGPT Enterprise has a log platform for checks; Claude Enterprise an audit log and a Compliance API. They are needed when something goes wrong, and to show that the rules work.
  5. Connectors and agents. Which internal sources the assistant can read and which agents can act: the administrator authorises them, one by one, not the individual users.
  6. Data residency, if it matters: Enterprise or API for OpenAI, a European tenant with “flex routing” turned off for Microsoft, Bedrock or Google Cloud for Claude, an edition with data regions for Workspace.

What you never paste

In Italy, the Ministry of Labour’s handbook lists the uses that are risky “in the absence of adequate technical and contractual guarantees”: contracts and invoices; CVs and employee data, salaries, evaluations, disciplinary notes; price lists, quotes, commercial strategies, margin analyses and sales plans, tenders and negotiations; technical documents, design diagrams, code, prototypes, draft patents; material covered by confidentiality agreements; automated decisions without supervision; content with legal value. In a manufacturing company these are everyday things: the quote with its margins, the drawing received from the customer, the inspection certificate, the price list with the metal premiums, the bill of materials.

There are two reasons, beyond common sense.

  • Trade secrets. In Italy, the Industrial Property Code (Codice della proprietà industriale) protects business information that is secret, has economic value and is subject to “measures to be considered reasonably adequate to keep it secret” (Article 98). Pasting a price list or a drawing into a personal account that can use them for training risks leaving exactly that requirement unmet. Written rules, the business version and training are also the measures that preserve the protection.
  • Personal data. With company plans the provider processes the data on the company’s behalf, under the data processing agreement; with personal accounts it also processes it for its own purposes, such as training or advertising. The personal data of customers and employees, therefore, does not go into a personal account, and goes into a company one only if it is needed, with the privacy notice updated.

A good checklist is the set of questions that the Garante (Italy’s data protection authority) put to DeepSeek in January 2025 before blocking it: what data it collects, from which sources, for what purposes, on what legal basis, where the servers are, what it uses for training. They are the same questions to ask of any tool, before activating it. Anyone who has to work on a sensitive document removes names, amounts and references before pasting it: usually the question can still be asked.

The obligations that apply even with a ready-made tool

  • Whoever uses them is a deployer. In the language of the AI Act, the company that makes ChatGPT, Copilot or Claude available to its staff is the deployer, in plain terms the user; the provider is OpenAI, Microsoft, Anthropic or Google. An individual who uses an assistant at home for themselves is not.
  • AI literacy (Article 4). Providers and deployers “take measures to support” the AI literacy of their staff, taking into account knowledge, experience and the context of use; no level, course or certificate is required. The Commission specifies that this also applies to those who use ChatGPT for texts and translations. What to do in practice is in the guide AI literacy: what the AI Act requires of companies.
  • Transparency (Article 50), since 2 August 2026. A customer-facing chatbot must make it clear that you are talking to a machine; anyone who generates or manipulates realistic images, audio or video (deep fakes) says so; anyone who publishes generated text to inform the public on matters of public interest declares it, unless there is human review with editorial responsibility. As the article is written, a newsletter or a quote written with an assistant and reviewed by a person does not fall within the obligation. The automatic marking of generated content is up to the providers, who for systems already on the market must comply by 2 December 2026; Anthropic, for example, states that it marks texts with an invisible watermark and files with C2PA content credentials.
  • High risk. The obligations of the deployer of high-risk AI systems (Article 26) do not concern an assistant used to write and summarise; they apply, from 2 December 2027, if the tool is used, for example, to select or evaluate workers; and anyone who adapts a general-purpose assistant for that purpose also takes on the provider’s obligations (Article 25). The risk levels are in the glossary.
  • Notice to workers. In Italy, to take one country’s example, Law No. 132/2025, the country’s AI law (Article 11), refers to Article 1-bis of Legislative Decree No. 152/1997, which concerns “fully automated” decision-making or monitoring systems: not a writing assistant. Since 23 October 2026 Legislative Decree No. 179/2026 has added that final decisions on hiring, on changing or ending the employment relationship, disciplinary measures and evaluations cannot be taken by an automated system alone, otherwise they are void, and that AI systems that affect the organisation, pace or safety of work enter the workplace risk assessment, with information and training on the specific risks.
  • Who supervises in Italy. In Italy the market surveillance authority for AI systems is the National Cybersecurity Agency, designated by Law No. 132/2025; since 23 October 2026 Legislative Decree No. 179/2026 has given it powers and fines: the fines follow those of the AI Act, with the lower value for SMEs; no fine is provided for AI literacy, but one is for transparency; for minor infringements the authority may limit itself to ordering a remedy within a set time.
  • The Garante and the providers. In Italy, the Garante temporarily blocked ChatGPT in March 2023; in December 2024 it adopted a corrective and sanctioning measure against OpenAI; in March 2026 the Court of Rome upheld OpenAI’s appeal and the authority temporarily removed the measure from its website; in January 2025 the Garante blocked DeepSeek; in April 2025 it recalled that the right to object to the use of one’s own data for training can also be exercised against other providers, such as OpenAI, DeepSeek and Google. For a company there is one lesson: the authority’s questions about data are the same ones to ask before choosing a tool.

The minimum rules on one page

The eight questions that the handbook of Italy’s Ministry of Labour suggests asking yourself before uploading any content, turned into rules. How to write the whole document, with an example, is in the guide Company AI policy: usage rules and an example.

Rules for the use of ChatGPT, Copilot, Claude and Gemini

1. Company accounts only. For work, use the tools on the list, with the work account: [name of the tool, plan]. Personal accounts are not used for work activities.

2. Training switched off. Where there is a setting on the use of data to improve the model, it is switched off; in company plans this is so by contract.

3. Personal, confidential or strategic data. These do not go into any tool: customer and employee data, purchase prices and price lists, quotes and margins, contracts, drawings and technical documents, code, material under confidentiality, login credentials. If you need to work on them, first remove names, amounts and references.

4. Where the data goes. For each tool on the list it is written down where the data is processed and for how long it is kept, taken from the contract, not from the advertising.

5. Results checked. What AI produces is a draft: numbers, names, dates, quotations and references to rules are checked by the person who uses it, who answers for it as for any other work of theirs.

6. Errors and made-up answers. The models can make up plausible answers: for technical, legal, tax or safety matters the result is reread by a competent person.

7. Transparency. Anyone who receives generated content, inside or outside the company, knows it when it matters; a customer-facing chatbot declares itself as such.

8. Sector obligations. Use is consistent with the rules of the sector, with the confidentiality agreements signed with customers and with the rules on personal data; in case of doubt, ask [name].

Anyone who enters confidential data by mistake reports it at once to [name]. Approved by senior management on [date]. Version 1.

Three things tell you whether the rules work: an internal register of the activities carried out with AI, which the handbook recommends; the audit logs of the company tools, to be used for security and to reconstruct an error, not to monitor people, because the rules on remote monitoring of workers apply (in Italy, Article 4 of the Italian Workers’ Statute: consult your employment adviser first); a human review of the outputs, on a sample, from time to time.

When a ready-made tool is not enough

A general-purpose assistant answers with what it knows and with what is pasted into it. When you need answers on the company’s documents, with the source cited, the next step is an assistant on company know-how built on your own archives; when you need AI to do something in the ERP (the company’s management software), the answer is AI agents, with written rules and confirmations; when the data cannot leave the company, the choice between cloud and on-premise must be made first. In the first two cases the models of OpenAI, Anthropic, Google or Microsoft remain in use, through the APIs and with a company contract; in the third you move to open-weight models installed in the company. In all of them, what changes is who keeps control of the data and of the actions.

How we work

In the companies we work with, we start from the tools people already use: we put them on a list with a company account, we write the usage rules with the people who run the company and we put in the plan who needs to know what, so that rules and plan also count as AI literacy measures under the AI Act. We apply the same rules to the systems we build, for ourselves and for our customers. When the ready-made tool is not enough, the next step is on the page AI consulting for companies.

Sources

  1. OpenAI Help Center, How your data is used to improve model performance
  2. OpenAI Help Center, ChatGPT Business overview
  3. OpenAI Help Center, What is ChatGPT Enterprise?
  4. OpenAI Help Center, Data residency and inference residency for ChatGPT
  5. OpenAI, Enterprise privacy at OpenAI
  6. OpenAI, Europe Terms of Use
  7. Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot
  8. Microsoft Learn, Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat
  9. Microsoft Learn, What is Microsoft Copilot? (overview and licences)
  10. Microsoft Learn, What is the EU Data Boundary?
  11. Microsoft Learn, Connect to AI subprocessor models (Anthropic) in Microsoft Copilot
  12. Microsoft Learn, Audit logs for Copilot and AI applications (Microsoft Purview)
  13. Microsoft, Privacy Statement, Microsoft Copilot section
  14. Anthropic, Updates to Consumer Terms and Privacy Policy (28 August 2025)
  15. Anthropic, Consumer Terms of Service (in force since 8 October 2025)
  16. Anthropic, Commercial Terms of Service (in force since 17 June 2025)
  17. Anthropic Privacy Center, Is my data used for model training?
  18. Anthropic Privacy Center, Where are your servers located?
  19. Anthropic, Claude in Microsoft Foundry
  20. Google, Gemini Apps Privacy Hub
  21. Google Workspace, Generative AI in Google Workspace Privacy Hub (updated as of 6 October 2026)
  22. Regulation (EU) 2024/1689 (AI Act), Article 4: AI literacy, text consolidated as of 27 July 2026 (AI Act Service Desk)
  23. AI Act, Article 50: transparency obligations for providers and deployers of certain AI systems
  24. AI Act, Article 3: definitions of provider, deployer and AI literacy
  25. AI Act, Article 111: deadline of 2 December 2026 for marking the content of systems already on the market
  26. European Commission, AI literacy: questions and answers (updated as of 27 July 2026)
  27. Italy's Law No. 132 of 23 September 2025, Article 11, on Normattiva (in Italian)
  28. Italy's Legislative Decree No. 152 of 26 May 1997, Article 1-bis, on Normattiva (in Italian)
  29. Italy's Ministry of Labour and Social Policies, Ministerial Decree No. 180 of 17 December 2025: guidelines for the implementation of AI in the world of work, Appendix 2, handbook for SMEs (in Italian)
  30. Italy's Legislative Decree No. 179 of 7 October 2026, powers of the national AI authorities and penalties, Gazzetta Ufficiale No. 234 of 8 October 2026 (in Italian)
  31. Italy's Industrial Property Code (Legislative Decree No. 30/2005), Article 98: trade secrets, on Normattiva (in Italian)
  32. Garante per la protezione dei dati personali (Italy's data protection authority), press release of 31 March 2023 on the temporary limitation of ChatGPT (in Italian)
  33. Garante per la protezione dei dati personali, press release of 20 December 2024 on OpenAI, with the note on judgment No. 4153/2026 of the Court of Rome (in Italian)
  34. Garante per la protezione dei dati personali, press release of 30 January 2025 on DeepSeek (in Italian)
  35. Garante per la protezione dei dati personali, press release of 29 April 2025 on the right to object to training (in Italian)
  36. ISTAT (Italy's national statistics institute), Imprese e ICT (enterprises and ICT), year 2025, report of 15 December 2025 (PDF, in Italian)
  37. Eurostat, Artificial intelligence by size class of enterprise (isoc_eb_ai), update of 15 June 2026
  38. Artificial Intelligence Observatory, Politecnico di Milano, press release of February 2026 (in Italian)
  39. HR Innovation Practice Observatory, Politecnico di Milano, press release of 12 May 2026 (in Italian)
  40. Microsoft, Privacy FAQ for Microsoft Copilot (applies only to the version of the app before 18 August 2026)
  41. Microsoft, Microsoft Copilot privacy controls (training settings in the previous version of the app)
  42. Anthropic Help Center, What certifications has Anthropic obtained? (16 March 2026)
  43. Anthropic Help Center, How Claude marks AI-generated content (8 October 2026)
  44. Anthropic Help Center, Configure custom data retention controls for Enterprise plans (7 October 2026)
  45. Anthropic Privacy Center, How long do you store my organization's data?
  46. OpenAI, Services Agreement (in force since 1 January 2026): ownership of input and output, confidentiality
  47. Italy's Law No. 300 of 20 May 1970 (the Italian Workers' Statute), Article 4, on Normattiva (in Italian)
  48. Garante per la protezione dei dati personali, press release of 28 April 2023 on the reopening of ChatGPT (in Italian)
  49. OpenAI, Make the most of company knowledge with ChatGPT (23 October 2025)
  50. OpenAI Help Center, Setting up single sign-on (SSO) for ChatGPT Business
  51. Eurostat, Artificial intelligence by economic activity (isoc_eb_ain2): manufacturing and metals
  52. OpenAI Help Center, Company knowledge in ChatGPT (plugin for Business, Enterprise and Edu, updated as of October 2026)
  53. Microsoft, Microsoft Copilot for individuals: your activity history (new app: prompts, responses and files do not train foundation models)
  54. Microsoft Learn, Flex routing (EU and EFTA): processing outside the EU Data Boundary during peak load periods
  55. OpenAI Help Center, Custom GPT retirement and migration FAQ
  56. Claude Help Center, Set up single sign-on (SSO): Team and Enterprise plans
  57. Google Workspace, Turn Gemini Enterprise on or off for users (Gemini Enterprise is not included in Workspace)

Checked on 17 November 2026. Rules and deadlines change: always check the official source before you decide.

Frequently asked questions

The short answers

Does ChatGPT use the texts I paste to train itself?

In the personal version yes, until you switch off "Improve the model for everyone" in Settings, Data controls, or choose "Do not train on my content" in OpenAI's privacy portal; a thumbs-up or thumbs-down rating still causes the whole conversation to be used. In the Business, Enterprise and Edu versions and in the API, by default, no.

Can I paste a quote or a customer's drawing into ChatGPT?

In a personal account, no: price lists, quotes, margin analyses, drawings and technical documents, contracts and material covered by confidentiality are exactly the content that the handbook of Italy's Ministry of Labour indicates as risky without technical and contractual guarantees, and pasting them into a tool that can use them for training can cost a trade secret its legal protection. In a company account with a data processing agreement, you can only if the usage rules provide for it for that type of document and if the confidentiality agreement with the customer does not forbid it, and in any case after removing the names, amounts and references that the question does not need.

Does Copilot see all the company's documents?

No. With the work account, Microsoft Copilot shows only the content on which the user already has at least read permission in Microsoft 365: the real risk is internal sharing that is too broad, not Copilot. Requests, responses and data read through Microsoft Graph do not train the models; for European tenants processing stays within the EU Data Boundary, with three exceptions: web searches, Anthropic's models (disabled by default for European customers) and processing during peak load periods ("flex routing"), which can take place outside Europe unless the administrator turns it off.

What is the difference between free Copilot, Copilot Chat and Microsoft 365 Copilot?

They are three things with the same name. Copilot with a personal Microsoft account is a consumer product, governed by the Microsoft Privacy Statement and with no contract with the company: for the new version of the app, available since 18 August 2026, Microsoft states that prompts, responses and files do not train foundation models; in the previous version they could, unless you opted out in the settings. Copilot Chat with the work account is included in eligible Microsoft 365 licences, uses mainly web data and the files attached to it, and is covered by enterprise data protection. Microsoft 365 Copilot, which since autumn 2026 has been called Microsoft Copilot in the app, in its paid version (Premium) works inside Word, Excel, Outlook and Teams with the tenant's data. Only the last two are covered by the data processing agreement.

Do I have to train employees if they only use ChatGPT?

Yes. Article 4 of the AI Act, rewritten by Regulation (EU) 2026/1744, requires providers and deployers to take measures to support the AI literacy of their staff, without prescribing a level, a course or a certificate. The European Commission answers precisely on the case of a company whose employees use ChatGPT to write or translate texts: they must be informed about the specific risks, for example hallucinations. In Italy, to take one country's example, Legislative Decree No. 179/2026 provides for no fine for this obligation.

Do I have to give a notice to workers who use AI?

In Italy, to take one country's example, the notice required by Law No. 132/2025 (Article 11) concerns fully automated decision-making or monitoring systems that affect the employment relationship, not a writing assistant. Since 23 October 2026 Legislative Decree No. 179/2026 has added that final decisions on hiring, on changing or ending the employment relationship, disciplinary measures and evaluations cannot be taken by an automated system alone. An internal communication with the usage rules remains advisable, and the guidelines of Italy's Ministry of Labour recommend it.

Does the data end up in Europe or in the United States?

It depends on the provider and the plan. Microsoft keeps European tenants within the EU Data Boundary, with the exception of web searches, Anthropic's models and processing during peak load periods ("flex routing"), which the administrator can turn off. OpenAI offers data residency in Europe for ChatGPT Enterprise, Edu and for the API; Business is not among the eligible plans. Anthropic stores data in the United States: to keep it in Europe, Claude is used through Amazon Bedrock or Google Cloud. Google has data regions only in some editions of Workspace. The question should be asked before signing, together with the one about retention.

Has the Garante, Italy's data protection authority, banned ChatGPT?

No: today there is no ban and the rules of the GDPR apply. In March 2023 the Garante temporarily blocked ChatGPT in Italy; the service resumed at the end of April 2023. In December 2024 it adopted a corrective and sanctioning measure against OpenAI; in March 2026 the Court of Rome upheld OpenAI's appeal and the authority temporarily removed the measure from its website. In January 2025 it blocked DeepSeek, and in April 2025 it recalled that the right to object to the use of one's own data for training can also be exercised against other providers, such as OpenAI, DeepSeek and Google.

What is shadow AI and how do you manage it?

It is the use of artificial intelligence tools not approved by the company, usually with personal accounts: in Italy, a third of workers say they do it. You manage it with three things: one or two company tools that really work, written usage rules and a contact person who answers for the choices and explains what the tools can get wrong. A ban with no alternative moves the use to where it cannot be seen.

Do you have a process you would like to see working better?

Tell us about it in a first conversation, with no obligation: together we work out whether and where AI brings a measurable benefit, and which step to start with.