Shadow IT and generative AI at work: how to govern employee use of ChatGPT, Copilot and Gemini

The short answer: your employees already use generative AI, whether you allowed it or not. The risk is not the tool itself, it is invisible use: a contract pasted into a free chatbot, a customer file summarised with a personal account, a browser extension that reads every page. To govern it, use four levers in this order: provide an approved tool under a business contract, write simple rules on forbidden data, discover and then block unapproved tools, and train your teams. Banning without offering an alternative only moves usage onto personal phones.

At ALLSAFE SOLUTIONS, we support SMBs in Casablanca and across Morocco on these questions, as part of our business cybersecurity in Morocco offering. This guide relies on official sources: Morocco’s data protection authority (CNDP), the French cybersecurity agency ANSSI’s recommendations on generative AI, and documentation from Microsoft, Google and OpenAI. We quote no usage statistics: available studies diverge too much to be cited seriously.

Shadow IT, shadow AI: what are we talking about?

Shadow IT covers everything employees use for work without IT approval: a personal Dropbox to send a file that is too large, WeTransfer, private messaging to talk to a supplier, a note-taking app, a “handy” browser extension. It is almost never malicious: it is an employee trying to work faster because the official tool does not meet their need.

Shadow AI is its newest and most delicate form. A generative AI tool does not just store a file: you hand it the very content of the work, so it can summarise, translate, correct or analyse it. A spreadsheet of customer reminders, a draft contract, minutes of a management meeting, a piece of code: all of it can leave in seconds for a service whose terms, retention period and hosting country the company does not know.

The concrete risks for an SMB

Risk	Typical example	Consequence
Confidential data leak	A salesperson pastes the price list and customer list to draft a proposal	Data out of control, unknown terms of use
Personal data outside any framework	An HR file or patient records summarised with a free account	Processing not compliant with law 09-08
Unmanaged accounts	AI subscription paid by the employee, tied to a personal address	Conversation history lost, or kept after they leave
Third-party extensions and apps	“AI assistant” extension reading page content, including webmail	Broad data access, unknown publisher
Wrong answers used as is	Legal clause or calculation reused without checking	Professional error, company liability
Unreviewed generated code	Script or macro copied into a production tool	Vulnerability or unexpected behaviour
Targeted phishing	Fake “free ChatGPT” site or fake extension	Credential theft, see phishing scams targeting SMBs

In its security recommendations for a generative AI system, published in April 2024, ANSSI groups the threats specific to these systems into three families: manipulation of the AI, infection of the data, and exfiltration of information. For an SMB that uses these tools, the third one matters most: what is typed into the tool can leave the company’s perimeter. The same guide recommends systematically reviewing AI-generated code, a common-sense rule that also applies to legal text and figures.

What the Moroccan framework says

There is, to date, no Moroccan law specific to generative AI at work. But the existing framework already applies:

• Law 09-08 on personal data protection. In its statement of 18 March 2025, the CNDP reminded that artificial intelligence processing using personal data falls under this law, and that the principles of transparency, fairness, specified purpose, proportionality and security apply. The Commission has also started work on a dedicated deliberation on AI. We detail your obligations in Morocco’s data protection law 09-08.
• Transfers of data abroad, regulated by law 09-08: most consumer AI services process data outside Morocco. Before approving a tool for personal data, check the legal basis for that transfer with your legal adviser.
• Law 05-20 on cybersecurity, for the entities it covers, which requires an information systems security policy: AI use must have its place in it. See cybersecurity law 05-20 and the DGSSI.
• Professional secrecy and contractual commitments: an accounting firm, a clinic or a law firm remains bound by secrecy, whatever the tool.

Personal accounts or business plans: the real difference

The decisive question is not “which chatbot?” but “under which contract?”. Major vendors clearly separate their consumer and business offerings:

Tool	Business offering	Commitment published by the vendor
Microsoft 365 Copilot and Copilot Chat	Sign-in with the Microsoft 365 work account	Enterprise data protection: prompts, responses and data accessed through Microsoft Graph are not used to train foundation models
Gemini in Google Workspace	Eligible Workspace editions	Customer data, including prompts, is not used to train models without the customer’s prior permission or instruction
ChatGPT	Business, Enterprise, API	Business data is not used for training by default

Sources: Microsoft Learn “Enterprise data protection”, Google Workspace generative AI Privacy Hub, OpenAI enterprise privacy policy. These commitments change: re-read them when you make your choice.

Two points to watch. First, Microsoft states that web search queries made by Copilot go through Bing, which is governed by terms separate from Microsoft 365. Second, third-party agents and connectors added to these tools have their own terms. A well-chosen business plan settles the model-training question; it does not replace usage rules or proper access rights.

If you already run Microsoft 365 or Google Workspace, start with the assistant built into your environment: it inherits your accounts, your multi-factor authentication and your sharing permissions. To compare both environments, see Microsoft 365 vs Google Workspace.

Our 6-step governance method

1. Take stock without a witch hunt. Firewall and DNS filtering logs, the list of browser extensions installed, and a short anonymous survey: what do teams use AI for, and why?
2. Choose one or two approved tools, under a business contract, tied to company accounts with multi-factor authentication. A useful official tool is the best protection against shadow AI.
3. Write one-page usage rules, included in the IT acceptable use policy (see the template below).
4. Classify data into three simple levels: public, internal, confidential. Only the confidential level is forbidden in any unapproved AI tool.
5. Put technical controls in place: blocking unapproved tools, controlling extensions, data loss prevention where licences allow.
6. Train and review. A short session with concrete examples from your business, then a quarterly review of discovered tools and team requests.

The usage rules to give employees

A rule nobody reads protects nobody. Here is the outline we suggest, to adapt to your business:

• Allowed: the approved tools, listed by name, with the work account.
• Forbidden in any unapproved tool: customer and supplier data, personal data of employees or patients, prices and margins, contracts, management documents, credentials and passwords, source code.
• Always review: an AI answer is a draft. Figures, legal references and commitments are checked by a person before anything is sent.
• No AI extension installed without IT approval.
• No personal subscription used for work: ask for a business licence.
• Report: a doubt, data sent by mistake, a useful tool to approve. Reporting must never be punished, or nobody will report anything.

These rules fit naturally with your phishing awareness training programme: same sessions, same reflexes.

Technical controls that work

Control	Possible tool	What it brings
Cloud app discovery	FortiGate or WatchGuard logs, Defender for Cloud Apps	List of services actually used; Microsoft offers a “Generative AI” category with a risk score per app
Category blocking	Firewall web filtering, DNS filtering	Blocks unapproved tools at network level, in the office and for remote work depending on the architecture
Sanctioned / unsanctioned status	Defender for Cloud Apps	Tags apps and blocks unsanctioned ones on onboarded devices
Data loss prevention	Microsoft Purview DLP on devices and in the browser	Can block pasting or uploading sensitive data to generative AI sites, depending on licences and supported browsers
Extension control	Intune or group policies	Allow list of browser extensions
Strong authentication	MFA, conditional access	Approved tools only open with a protected work account
Smartphones	MDM	Work apps separated from personal apps

These features depend on your licences and your devices: we size them case by case, without making you buy what you will not use. For the Microsoft baseline, see securing Microsoft 365 in 10 settings and Microsoft Intune for SMBs; for smartphones, managing business smartphones with an MDM. These measures are part of a broader approach, described in Zero Trust for an SMB: where to start.

A word of caution: no block is total. An employee can always photograph a screen with a personal phone. That is why the approved tool and the rules matter as much as the technology.

The mistakes we see

1. An outright ban with no replacement tool: usage moves to phones and becomes invisible.
2. AI licences bought without reviewing sharing permissions: an assistant connected to email and files finds everything the user can already open, including folders shared too widely.
3. No inventory of browser extensions.
4. A policy written by legal alone, incomprehensible to teams, and never presented.
5. AI accounts tied to personal addresses, impossible to recover when the employee leaves.
6. No review of AI answers before they are sent to a customer.

Getting started checklist

• Inventory of AI tools and extensions actually in use.
• One or two approved tools, under a business contract, with MFA.
• Review of file sharing permissions before enabling an assistant connected to your data.
• One-page usage rules, included in the IT policy and signed.
• Simple data classification: public, internal, confidential.
• Unapproved tools blocked at the firewall or DNS filtering level.
• Data loss prevention if your licences allow it.
• Training session and a contact point for new tool requests.
• Quarterly review.

How we support you

We start with an assessment: tools discovered in your logs, extensions, sharing permissions, existing licences. We then propose one or two tools suited to your environment, as a Microsoft partner and Google Workspace partner, write the usage rules with you, and configure the controls on your firewalls, computers and smartphones. The equipment and services we manage are monitored 24/7, and a critical incident, such as a suspected data leak, is handled in under 15 minutes.

To go further, explore our business cybersecurity in Morocco services and our Microsoft 365 and Azure cloud offering, then request your free initial audit: we reply within 24 business hours.