Secure AI Adoption in the Enterprise: Governance, Policies and Guardrails

Secure AI adoption means deciding which tools your company uses, with which data and under which rules, before each employee decides on their own. In practice it takes three layers: an owner with a governance framework, an AI use policy everyone understands, and technical guardrails that enforce that policy in Microsoft 365 and on the network. Without them, AI still enters your company, just outside your control.

AI is already in your company, approved or not

According to Microsoft and LinkedIn’s 2024 Work Trend Index, 75% of knowledge workers already use AI at work, and 78% of those users bring their own tools instead of waiting for the company to choose one. The same study found that 60% of leaders worry their organization lacks a plan and vision to implement AI.

That is shadow AI: AI used without IT approval or oversight. A salesperson pastes the customer list into a free chatbot to draft emails, an accountant uploads a financial statement to get a summary, a manager installs a browser extension that reads their inbox. Nobody acts in bad faith. They want to work faster, and the tool is one click away.

Latin America is moving the same way. The 2025 Latin American Artificial Intelligence Index (ILIA), published by Chile’s National Center for Artificial Intelligence (CENIA), places Panama in its “adopters” group, with 38.95 points against a regional average of 42.98, and highlights strong adoption of web-based generative AI. For companies operating in Panama, the question is no longer whether to use AI, but how to organize it so it works for you.

What can go wrong when AI has no rules

IBM’s Cost of a Data Breach Report 2025 measured the effect of shadow AI on breaches. One in five organizations reported a breach caused by shadow AI, and organizations with high levels of unsanctioned use paid an average of USD 670,000 more per breach. Only 37% had policies to manage AI or detect shadow AI, and 63% of breached organizations had no AI governance policy or were still developing one.

Then there is the law. In Panama, Law 81 of 2019 on personal data protection has been in force since March 29, 2021. It requires data to be processed for a specific, explicit and lawful purpose, generally with the data subject’s consent, and it sets conditions for transferring data to other countries. Pasting customers’ personal data into a tool that processes it abroad, without checking those conditions, is a legal risk as well as a technical one. If you operate in several countries, the same review applies under each local privacy law.

For a mid-sized company, the technical risks come down to four. They are described in NIST’s generative AI profile (NIST AI 600-1) and in the OWASP Top 10 for large language model (LLM) applications, two references worth keeping close:

• Data leakage. Customer data, contracts or payroll end up in a public tool whose terms nobody read. OWASP lists it as “sensitive information disclosure”, the second risk in its 2025 list.
• Internal oversharing. An assistant connected to your files finds, in seconds, documents that were shared too widely years ago: salaries, performance reviews, board minutes.
• Made-up answers. NIST calls confidently stated false content “confabulation”. If nobody reviews it, the error reaches a proposal or a customer.
• Manipulated AI. Prompt injection, first on the OWASP Top 10, happens when text hidden in an email, a document or a web page changes what the assistant does.

AI governance: who decides, and with which framework

AI governance sounds like multinational bureaucracy, but in a small or mid-sized business it is concrete: one person or a small committee decides which tools are approved, for which uses and with which data, and reviews those decisions on a regular schedule. Without a clear owner, every department ends up with its own tool and nobody knows what information has left the company.

You do not need to invent the method. The NIST AI Risk Management Framework (AI RMF 1.0, January 2023) is voluntary, works for any sector and organizes the work into four functions: Govern (culture, roles and policies), Map (understand the context and where AI is used), Measure (assess the risks) and Manage (prioritize and act). If you need certification, ISO/IEC 42001:2023, published in December 2023, sets the requirements for an AI management system in organizations that provide or use AI-based products and services.

If you sell into Europe or serve European customers, also track the EU AI Act, which entered into force on August 1, 2024. Since February 2, 2025 its prohibitions apply, along with the obligation to ensure that staff using AI systems have sufficient AI literacy.

For a company of 20 to 200 people, getting started takes three steps. First, name an owner. Second, take an inventory of the AI tools already in use: ask every department and check what the firewall logs show. Third, classify each use by risk, because brainstorming a social media post is not the same as analyzing customer files.

What your AI use policy should say

The policy turns those decisions into rules for everyone. It works best when it fits on two pages, uses examples from your own operation and is explained in a short meeting instead of arriving as an attachment nobody opens. At a minimum, it should answer these questions:

• Approved tools: which ones can be used, always with a company account and never with personal accounts.
• Allowed and forbidden data: what can be typed or uploaded (public information, internal drafts) and what never can (customers’ personal data, unpublished financial information, passwords, contracts).
• Human review: anything that goes to a customer, a regulator or an important decision is reviewed by a person.
• Transparency: when a customer must be told that content was prepared with AI assistance.
• New tools: how to request a new tool or extension, and who approves it.
• Incidents: whom to tell, and how fast, if someone shared something they should not have.

Technical guardrails: making the policy stick

A policy alone is not enough. If the approved tool is slower or harder than the free one, people go back to the free one. That is why the policy comes with a good tool, training and guardrails: the technical controls that limit what AI can see and do, and what users can share with it.

In a company that runs on Microsoft 365, there are five main guardrails. None of them is exotic; most are configured with tools that many companies already pay for in their licenses and have never turned on:

• A business tool that does not train on your data. Microsoft states that in Copilot, prompts, responses and data accessed through Microsoft Graph are not used to train foundation models. Before approving any other tool, check what its terms say about your data.
• Protected identity. MFA for everyone and Conditional Access, so corporate AI is used with company accounts and from devices that meet your rules.
• Sensitivity labels. With Microsoft Purview, a document encrypted by a label keeps its permissions, and Copilot honors the usage rights that label grants each user.
• Control over unapproved AI apps. From the firewall or your web security service you can see which AI services are used on your network and block or limit the ones that are not approved.
• Logging and retention. Microsoft Purview lets you search and retain Copilot interactions, which helps when investigating an incident or answering an audit.

Before you turn on Copilot, fix your permissions

Microsoft 365 Copilot, now called Microsoft Copilot in Microsoft’s documentation, only surfaces organizational data that the user already has at least view permission for. That is good news as long as permissions are right. The problem is that permissions drift over the years: folders shared with the whole organization, open links nobody remembers, SharePoint sites for projects that ended long ago.

Copilot does not create that problem, but it exposes it in seconds: one question is enough to surface a document nobody knew was open. That is why Microsoft’s guidance for preparing data before Copilot is built on three pillars: remediate oversharing, set up guardrails and meet regulations. It relies on Microsoft Purview and SharePoint Advanced Management; Microsoft says the latter is included with the Copilot license.

Microsoft also warns that generative AI responses are not guaranteed to be 100% factual and asks users to review them before sending. The practical advice: before assigning licenses, review who can access what in SharePoint, OneDrive and Teams, fix what is most exposed, label sensitive information and start with a pilot group.

A four-phase plan for secure AI adoption

This is how we run AI adoption with our clients, in the same order described on our AI for business page (in Spanish). Each phase leaves something finished before the next one starts:

• Assessment: inventory of the tools already in use, a review of permissions and licenses, and use cases with real value for your business.
• Preparation: AI use policy, MFA and Conditional Access, permissions fixed and sensitivity labels on sensitive information.
• Pilot: a small group with concrete use cases, such as meeting summaries, email or sales analysis, with training and measured results.
• Scale-up: more users and more use cases, including automation with Power Automate or agents in Copilot Studio, under the same guardrails.

Why the order matters

A company that jumps straight to a pilot without fixing permissions risks having its first Copilot surprise be the payroll spreadsheet. One that stops at the policy, without an approved tool, pushes its people back to shadow AI. And one that buys licenses for everyone without a pilot ends up paying for tools nobody knows how to use.

The order also makes results measurable. If the pilot has defined use cases, after a few weeks you can compare how long it takes to prepare a proposal or close the monthly report, and decide with data whether to scale up.

How Techno Sistemas helps

We do this work alongside the client’s IT team or as their IT department. We review the Microsoft 365 tenant, clean up permissions and licenses, set up identity with phishing-resistant MFA and Conditional Access, and support the pilot with training and measurement.

If your company already has a monthly IT support plan (in Spanish) with us, Microsoft 365 administration is part of the service and AI adoption builds on that work. If you would rather start with a diagnosis, we run a security assessment of your environment and hand you the findings with clear priorities. Companies setting up operations in Panama can find an overview of our services on our English page.

Frequently asked questions

What is shadow AI?

It is the use of AI tools the company has not approved or does not oversee, usually with personal accounts. In its 2025 report, IBM found that one in five organizations reported a breach caused by shadow AI.

Does Copilot use my company's data to train its models?

No. Microsoft states that prompts, responses and data Copilot accesses through Microsoft Graph are not used to train foundation models. Copilot only shows what the user already has permission to see, which is why permissions should be cleaned up before turning it on.

Does banning ChatGPT and similar tools solve the problem?

Rarely. If people have no approved alternative, they use the one they know on their phone or personal account. It works better to offer a business tool, explain which data must not be shared and block only what poses a real risk.

Do I need ISO/IEC 42001 certification?

It is not mandatory. It makes sense if your customers or your industry require it, or if you build AI products. For most small and mid-sized businesses, applying the NIST framework with a clear policy and technical guardrails is a good starting point.

What does Panama's law say about using customer data in AI tools?

Law 81 of 2019 applies to any processing of personal data, including processing done with AI: a specific purpose, the data subject’s consent where required, and conditions for transferring data to another country. Review uses involving personal data with your legal counsel.

If you want your team to use AI with clear rules and your data under control, we start with an assessment of your Microsoft 365 environment. See how we support AI adoption (in Spanish)

Sources

1. Microsoft WorkLab – Work Trend Index 2024: AI at Work Is Here. Now Comes the Hard Part (2024): 75% of knowledge workers use AI at work, 78% of AI users bring their own tools, 60% of leaders worry they lack a plan.
2. IBM Newsroom – IBM Report: 13% of Organizations Reported Breaches of AI Models or Applications, 97% of Which Reported Lacking Proper AI Access Controls (2025): One in five organizations breached through shadow AI, USD 670,000 higher cost, 37% with AI policies, 63% without an AI governance policy.
3. CENIA – Índice Latinoamericano de Inteligencia Artificial (ILIA) 2025: fichas de país y aspectos metodológicos (2025): Panama in the adopters group, 38.95 points against a regional average of 42.98, with strong generative AI adoption.
4. IAPP – Panamá y el futuro reglamento de su ley de protección de datos (2020): Law 81 in force since March 29, 2021; purpose limitation, consent and bases for international transfers.
5. NIST – Artificial Intelligence Risk Management Framework (AI RMF 1.0) (2023): Voluntary, sector-agnostic framework with four functions: Govern, Map, Measure, Manage.
6. NIST – AI 600-1 Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (2024): Risks specific to generative AI, including confabulation, data privacy and information security.
7. OWASP Gen AI Security Project – OWASP Top 10 for LLM Applications 2025: LLM01 Prompt Injection and LLM02 Sensitive Information Disclosure.
8. ISO – ISO/IEC 42001:2023 AI management systems (2023): Requirements for an AI management system in organizations that provide or use AI-based products; published December 2023.
9. European Commission – AI Act: Regulatory framework on AI: Entered into force August 1, 2024; prohibitions and AI literacy obligations apply from February 2, 2025.
10. Microsoft Learn – Data, Privacy, and Security for Microsoft Copilot (2026): Copilot only surfaces data the user can view, does not train foundation models on prompts or responses, honors sensitivity labels; product renamed Microsoft Copilot.
11. Microsoft Learn – Secure & Governed Data Foundation for Microsoft Copilot: Foundational Deployment Guidance (2026): Three pillars before Copilot: remediate oversharing, set up guardrails, meet regulations, using Purview and SharePoint Advanced Management.