Banning AI does not make shadow AI disappear. It can make it harder to see.

Unsanctioned AI use is already common. Effective governance starts with visibility, a usable approved alternative and controls that reflect the risk.

Most organisations have now taken some kind of position on AI. Some have approved tools, bought licences and started training people. Others have added a line to an acceptable use policy telling staff not to put company information into public AI tools and assumed that closes the issue.

The problem is that AI adoption does not wait for policy. Shadow AI describes the use of AI tools, accounts or features that an organisation has not approved or does not have visibility over. In many cases, people are using them for an understandable reason: they are trying to get work done faster.

Research published by BlackFog in January 2026, based on a survey of 2,000 employees in the UK and US at organisations with more than 500 staff, found that 49% used AI tools that had not been sanctioned by their employer. Around half, 51%, said they had connected or integrated AI tools with other work systems or apps without IT approval or oversight.^[1]

Among respondents using unapproved AI tools, 58% relied on free versions. BlackFog notes that these can lack the enterprise-grade security, data governance and privacy protections organisations expect from approved business tools. The same research found that 33% had shared enterprise research or datasets, 27% employee information such as payroll or performance data, and 23% financial statements or sales data.^[1]

That does not look like a fringe behaviour problem. It looks like a visibility and governance problem.

The behaviour is not limited to junior staff

One of the more useful findings from the BlackFog research is that unsanctioned AI use is not confined to employees quietly working around the rules. Sixty-three per cent of respondents said using AI without IT oversight was acceptable when no approved option was available.^[1]

At senior levels, the appetite for speed was even more pronounced. BlackFog reported that 69% of president and C-suite respondents, and 66% of directors and senior vice presidents, believed the productivity benefit could outweigh privacy or security concerns.^[1]

That matters because a policy only works when the organisation around it makes the safer route realistic. If people believe the approved option cannot do the job, some will find another route. A blanket prohibition may reduce visible use while leaving the underlying demand unchanged.

The exposure is no longer theoretical

IBM’s 2026 Cost of a Data Breach Report adds a different perspective. The study covered 602 organisations that experienced data breaches between March 2025 and February 2026 across 16 countries and 17 industries.^[2]

According to IBM’s report, as summarised by Cybersecurity Dive, the share of security incidents involving shadow AI more than doubled year on year to 43%. More than two thirds of breached organisations said they did not have governance processes in place to limit shadow AI.^[2][3]

IBM put the global average cost of a data breach in the study at $4.99 million. That is a global average across breached organisations, not a forecast for what an incident would cost every UK business. The more transferable point is that uncontrolled AI use is now appearing in real security incidents while governance is still catching up.^[2]

Why a blanket ban can leave a blind spot

Traditional security controls were designed around familiar routes for data movement such as email, file sharing, managed applications and known destinations. AI changes the pattern. Sensitive information can leave through a prompt pasted into a browser, a file uploaded to a consumer tool or an AI feature embedded inside another application.

That means an organisation can have a clear policy and still lack a clear picture of what is happening in practice, particularly where personal accounts, unmanaged devices or rapidly changing AI services are involved.

The answer is not to remove every restriction. Some data should have firm boundaries. The stronger approach is to understand current behaviour first, then apply controls that reflect the sensitivity of the data, the user, the destination and the business need.

A practical visibility-first approach

A useful starting point is less dramatic than a company-wide AI transformation programme. It is a short piece of discovery work followed by a policy people can actually follow.

1. Find out what is already being used. Ask teams which AI tools and features are part of day-to-day work, what tasks they use them for and whether those tools have been approved. Policy written without that picture is based on assumptions.
2. Give people an approved option that solves the real problem. If the sanctioned route is slower, harder to access or unsuitable for the task, workarounds will remain attractive. Licensing and security matter, but so does usability.
3. Define the information that needs the strongest protection. Source code, customer records, HR information, commercially sensitive pricing and confidential contract terms are obvious examples. Clear categories are easier to understand and enforce than a vague instruction not to share company information.
4. Match the response to the risk. Some activity warrants a hard block. Some needs a warning, guidance or a better approved alternative. Treating every interaction as equally dangerous can create unnecessary friction and encourage people to work around the control.
5. Review the position regularly. AI services, features and working habits change quickly. Governance needs to be maintained as part of normal security and technology management rather than written once and forgotten.

Visibility also needs to be handled responsibly

There is an important UK data protection point here. Technology that monitors worker activity can itself process personal information. The ICO says employers must have a lawful basis for monitoring, be transparent with workers about what is being monitored and why, and consider whether a Data Protection Impact Assessment is required. The ICO recommends carrying out a DPIA even where the organisation concludes that the processing is not high risk.^[4]

That makes transparency part of the design. AI governance should protect organisational data without creating an unnecessarily intrusive monitoring environment. Security, HR, legal or data protection teams may all need to be involved depending on the approach being considered.

The VitrX view

Shadow AI is a useful example of why security and productivity cannot be planned separately. People adopt tools because they see a practical benefit. The business still needs to understand where data is going, what should be allowed, and what needs stronger control.

A sensible first step is to establish what is already in use, identify the highest-risk data flows, review the approved alternatives available to staff and agree who owns the policy. From there, the technology controls can be matched to the risks that actually exist rather than the risks the organisation assumes it has.

VitrX can help organisations review the technology around AI use, from approved workplace and Microsoft tools through to cyber security, endpoint, identity, network and data protection controls. The aim is to give people a productive route to use AI while keeping sensitive business information inside an environment the organisation can understand and govern.

If the honest answer to ‘which AI tools did our people use last week?’ is ‘we do not know’, that is a useful place to start the conversation.

References