Shadow AI: The Invisible Risk of Ungoverned AI Use in Remote Work

Shadow AI: el riesgo invisible del uso no gobernado de IA en el trabajo remoto

While leadership is still debating whether to officially adopt generative AI, a good chunk of the team is already using ChatGPT, Claude or other AI tools every day — just off the books, without notice and without any kind of oversight. That phenomenon has a name: Shadow AI. It can be especially difficult to manage in remote and distributed teams, exactly where visibility into day-to-day work is already harder to come by. This article covers what Shadow AI is, why it’s a real risk — not just a reputational one, but a data-exposure one — and how Hubstaff AI offers a governed alternative to that informal use.

The path has six steps: what Shadow AI is and what the research says, why banning it rarely works, what data gets exposed, which Hubstaff AI features relate to the topic, how to build workable governance, and how far — and where — a tool like this stops.

What is Shadow AI

Shadow AI is the use of artificial intelligence tools by employees without approval, tracking or any visibility from the company — the same problem “Shadow IT” already posed with unauthorized software, now applied to AI assistants. Research identifies AI use that conflicts with internal policies, information sharing with public tools, and limited transparency about that use.

Several studies point in the same direction, although they measure different things and deserve a careful reading:

• KPMG and the University of Melbourne. In their 2025 global study, among workers who use AI, 44% reported use contrary to organizational policies or guidelines; separately, 48% reported uploading company information to public AI tools. These are separate measures, not percentages of the entire sample.

• PagerDuty. Its 2026 survey, run by Wakefield Research among 1,250 office professionals at large companies, found that, among those who had previously used AI in their work responsibilities, 66% had used AI tools at work despite believing they were not permitted, and that 88% of the overall sample had shared work data with a public AI tool. The study covered companies with annual revenue of at least US$500 million in Australia, Japan, the UK and the US, excluding IT and technology roles.

• Gartner. A survey of 302 cybersecurity leaders (2025) indicates that 69% of organizations suspect, or have evidence, that employees use prohibited public generative AI.

None of these figures describes one particular company. What they show is a pattern: informal use is not the exception, and it almost never starts from bad intent.

The everyday examples are simple: pasting the text of a contract to ask for a summary, uploading a customer spreadsheet to tidy it up, copying a snippet of code for someone to explain it, or transcribing a meeting with an assistant the company never approved. Each action is small. Added up across dozens of people and personal accounts, they form a leak of information that nobody sees.

Shadow AI: the invisible risk of using AI governance in remote work

Why banning it rarely works

The most common reaction of a company that discovers Shadow AI is to ban it. The PagerDuty figure already hints at why that fails: two-thirds of respondents who had previously used AI for work used AI even though they believed it was not allowed. When a tool saves time and there is no equally practical official alternative, a ban alone may not stop use and may encourage concealment.

This gets worse on remote teams for three reasons:

• There is no looking over a shoulder. Visibility into how work gets done is already limited, and AI use happens in the browser, on the phone or in personal accounts.

• Pressure to deliver. When a deadline closes in, people pick the tool that solves the problem, not the one that is approved.

• No alternative on offer. If the company does not provide an official channel, each employee decides on their own.

That is why the problem is not employee curiosity but the absence of an official channel that is as convenient as the informal one. That idea organizes the rest of this article.

What data gets exposed, and why it matters

Not everything pasted into a chat carries the same risk, but some categories are more worrying because, once sent, the company loses control over them:

• Customer data. Names, contacts, histories and any personal information the company handles under its responsibility.

• Code and technical documentation. Snippets of proprietary or client software, which can reveal business logic or vulnerabilities.

• Contracts and legal documents. Commercial terms, clauses and third-party data.

• Financial and strategic information. Spreadsheets, forecasts and plans that should not leave the organization.

• Data about people on the team. Reviews, salaries and other HR information.

The underlying problem is the lack of control over what happens to that information afterwards: where it is stored, who can access it and whether it is reused. Personal data protection is also regulated by each country’s laws, and a company that does not know which data flows through external tools will struggle to show that it handles it properly.

The table compares informal use with a governed channel:

AspectInformal use (Shadow AI)Official, governed channel
Who knows about itEach employee, privately.The company, which defines and communicates the channel.
How data is accessedCopying and pasting into an external chat.Through an official connection with permissions.
Who sees whatNo rules: it depends on what each person decides to paste.Access follows the roles each person already has.
Record and visibilityNone.Official connection with permissions; usage visibility depends on features and configuration, not MCP alone.
Standard across teamsEach person, a different method.The same criterion for teams in different countries.

Hubstaff AI features relevant to this topic

Open API for centralized visibility. Lets the IT team pull activity and productivity data from Hubstaff and cross-reference it with other internal systems, creating one source of visibility instead of relying on scattered, unmapped tools. The enhanced API publishes a machine-readable schema, and access respects existing permissions: each person sees what their role allows.

Official MCP server. Connects compatible assistants such as Claude, ChatGPT and Gemini to Hubstaff data through an official integration and each user’s permissions. It is read-only: this connector does not create, modify or delete Hubstaff data, or execute approvals or payments in Hubstaff. This does not restrict the assistant’s other tools or guarantee that data stays within company infrastructure. The organization must approve the connected assistant, its data-processing terms and the account used.

Official CLI. Gives technical teams an auditable path to automate queries and integrations from the command line, reducing the need to manually export data into unapproved external tools. The CLI turns API endpoints into commands that run directly and refreshes access tokens automatically.

Unusual activity detection. Flags patterns that fall outside the expected range in team activity for a human manager to review — a first filter that helps spot routine changes without relying on constant surveillance. Hubstaff makes clear that unusual does not necessarily mean malicious or wrong.

Benefits for businesses

MCP, API and CLI provide official ways to query or integrate Hubstaff data. With corporate accounts, appropriate permissions and internal policies, they can reduce manual copying into unevaluated tools and support consistent practices across teams. App and URL tracking complements this approach with tool-usage visibility; that is not an MCP feature and does not eliminate all Shadow AI.

• Potentially lower exposure. An approved integration can reduce manual copying; protection also depends on the assistant, account and policies.

• A standard across countries. The same channel and the same rules for teams in different places and time zones.

• Complementary visibility. App and URL tracking helps understand tool usage; MCP enables permission-based data queries.

• The decision stays human. AI prepares and flags; whoever manages the team decides.

Use cases

IT companies and agencies. They have informally discovered that the team already uses AI to summarize meetings, documents or code with zero record of it. An official channel lets them keep the convenience and establish usage rules for Hubstaff data.

HR and compliance teams. They need to map where customer data might be circulating through unapproved AI tools. The first step is to offer an alternative and a clear criterion.

Managers of remote teams. They want to offer an official AI channel before informal use becomes an entrenched habit, because it is easier to guide a new habit than to change an established one.

Companies migrating from another platform. They can use the switch as a chance to also organize the team’s AI usage, instead of leaving it for later.

How to build governance in five steps

A tool helps, but AI governance starts with decisions the company makes. A reasonable path has five steps:

• 1. Diagnose. Ask the team, without an investigative tone, which tools it uses and what for. The goal is to understand, not to punish.

• 2. Write a policy. What kind of information can be used with AI, what cannot, and which tools are approved.

• 3. Provide company tools and accounts. Company-managed accounts give more control than personal ones.

• 4. Open an official channel for work data. This is where MCP, API and CLI fit: queries over activity and productivity data, within known permissions.

• 5. Train and review. Explain why the rules exist and review periodically what works and what is being worked around.

Steps 1, 2, 3 and 5 depend on the organization. Step 4 is the one Hubstaff AI makes easier.

What Hubstaff AI does not solve

It is worth being clear about the limits, because that is what prevents false security. Hubstaff AI offers a more governed channel for using AI over Hubstaff data. It does not control everything the team does with AI outside that channel, it does not stop someone from pasting information into another tool, and it does not replace an internal policy.

• MCP is not application monitoring. Hubstaff’s app and URL tracking can identify AI tools and time spent, depending on configuration, for comparison with internal policy. It does not capture prompts or conversations. Unusual activity detection is a separate feature: it flags patterns for human review and does not itself determine whether a tool is authorized.

• It does not guarantee regulatory compliance. It is a more governed channel, not a certification of compliance with any country’s rules.

• It does not diagnose mental health. It works with activity and productivity data. It can flag patterns, but that kind of diagnosis requires specific technical evaluation.

• It depends on the AI account that is connected. The MCP connects Hubstaff to the assistants the team already uses. For the channel to be truly governed, those assistants should be used through company accounts rather than personal ones.

• Availability depends on the plan. According to Hubstaff, API, CLI and MCP access follows the Team and Enterprise plan rule; confirm what each account includes.

Why choose this solution

Hubstaff AI follows the same principle as always: the person decides, AI prepares and recommends. Applied to Shadow AI, that means giving the team an AI channel within the company’s own rules — which the organization must assess against its policies — instead of leaving it to each employee to decide, on their own, where to paste work information.

For teams evaluating a platform switch, Aufiero Informática — Hubstaff’s official distributor in Latin America — offers the first 2 hours of migration consulting at no cost, to help reorganize teams and departments during the transition.

Frequently asked questions

Does using Hubstaff’s MCP server with ChatGPT or Claude guarantee compliance with occupational or psychosocial-risk regulations?

No. Hubstaff AI does not guarantee or ensure compliance with any country’s occupational or psychosocial-risk regulations. It simply offers a more governed channel for using AI over productivity data.

Does Hubstaff AI fully eliminate the risk of Shadow AI at a company?

No. It can help reduce risks when combined with internal policies, corporate accounts and appropriate permissions, but it does not control all AI use or guarantee the elimination of Shadow AI.

Does unusual activity detection identify who is using unauthorized AI tools?

Not by itself. Unusual activity detection flags patterns for human review. Separately, Hubstaff’s app and URL tracking can identify AI tools and time spent, depending on configuration, without capturing prompts or conversations. The company must compare that use against its policies.

Does Hubstaff AI diagnose stress or burnout caused by AI tools?

No. It works with activity and productivity data, not mental-health assessments — that kind of diagnosis requires specific technical evaluation.

Do I need a technical team to use the API, MCP or CLI?

The API and CLI are built for teams with their own IT or development resources. The MCP server, on the other hand, was built for teams that don’t want to touch code at all: it connects Hubstaff to the AI assistants the team already uses.

Does Aufiero help with migrating from another platform to Hubstaff?

Yes. Aufiero Informática offers the first 2 hours of migration consulting free of charge to help reorganize teams and departments during the switch.

Conclusion

Shadow AI does not automatically disappear when an official channel is offered. Hubstaff can support a responsible-use strategy: MCP, API and CLI facilitate access to its data, while app and URL tracking provides complementary visibility. Results depend on company policies, accounts, configuration and human review.

Aufiero Informática is Hubstaff’s distributor in Latin America.

AI

Aufiero Informática

Embajadores de marca virtuales en Latam. Distribuidores oficiales de software de gestión, productividad y seguridad.