Skip to content
Aventail™ investigation into UK government AI dependence featured in Computer Weekly.Read the story
Aventail™
All guides

Shadow AI: what it is, how to find it, and an AI acceptable use policy you can copy

15 min read

Shadow AI is staff using AI tools the organisation has not approved. Dated UK and global figures, how to find it, and a copyable AI acceptable use policy.

Shadow AI is the use of AI tools for work that the organisation has not approved: a personal ChatGPT account, a browser extension, a meeting recorder, a coding assistant on someone's own licence. It is a form of shadow IT, and the specific risk is that company or customer data goes into a service nobody has assessed, under terms nobody has read, with no record that it happened.

The figures below come from the current primary sources, read on 15 November 2026.

Figure

Source

Sample and date

71% of UK employees have used unapproved consumer AI tools at work. 51% do so every week

Microsoft UK and Censuswide [3]

2,003 UK employees, October 2025

31% of UK workers who use generative AI use it without their employer knowing. 17% pay for their own tools

Deloitte GenAI Workforce Survey 2026 [4]

25,000 UK workers, May to June 2026

48% of employees use unapproved AI tools at work: 65% of senior decision-makers, 31% of the staff below them

TrustedTech and Censuswide [5]

2,001 UK and US employees, March 2026

78% of people who use AI at work bring their own AI tools

Microsoft and LinkedIn Work Trend Index [6]

31,000 people in 31 countries, May 2024

43% of security incidents involved shadow AI, up from 20% a year earlier

IBM Cost of a Data Breach Report 2026 [1]

602 breached organisations, breaches from March 2025 to February 2026

USD 5.39 million average breach cost where shadow AI was involved. The global average was USD 4.99 million

IBM Cost of a Data Breach Report 2026 [1]

Same study

32% of breached organisations had a policy in place to manage AI and prevent shadow AI, down from 37%

IBM Cost of a Data Breach Report 2026 [1]

Same study

These surveys ask different questions of different groups, so the percentages cannot be compared directly. They agree on the direction: at least a third of staff, and in some samples most of them, use AI tools their employer has not approved.

What counts as shadow AI

The UK's National Cyber Security Centre defines shadow AI as "the use of AI technology which isn't captured in an organisation's approved systems and processes" and treats it as a form of shadow IT [8]. In practice it takes five forms:

  • A personal account on a public tool. Free or paid ChatGPT, Claude, Gemini or Perplexity used for work.

  • A personal account on a tool you have approved. The company pays for an enterprise plan and a member of staff signs in with a private email address instead. The account is outside your controls.

  • AI features inside software you already own. A note-taking app, PDF editor or CRM adds an assistant in an update and nobody reviews it.

  • Browser extensions and meeting recorders. They read pages, email and calls in the background.

  • Coding assistants and agents. A developer points an agent at a repository with their own API key.

Shadow AI adds two things to ordinary shadow IT. The service may keep the content or use it to improve a model, and an agent can act with the access the user gave it. The NCSC warns that an attacker who exploits an agent gains "the same data, services, and privileges that the agent has legitimate access to" [8].

How common shadow AI is

In Microsoft's UK research, 71% of employees had used unapproved consumer AI tools at work and 51% did so weekly [3]. The NCSC's shadow AI guidance also cites a 71% figure [8]. The reasons matter more than the total: 41% said the tool is what they are used to in their personal life, and 28% said their employer does not provide an approved option. Only 32% were concerned about the privacy of the company or customer data they entered.

Deloitte's 2026 survey of 25,000 UK workers found that 31% of generative AI users use it without their employer knowing, 46% use free tools at work and one in six pays for their own subscription [4]. Deloitte estimates UK workers spend about £1 billion a year on AI tools they use at work.

Use rises with seniority. In the TrustedTech survey, 65% of senior decision-makers used unapproved AI tools against 31% of the people they manage, and 73% of the C-suite did [5].

Microsoft's global Work Trend Index reported its 78% figure in May 2024 [6]. The 2025 and 2026 editions do not repeat the measure [7].

What shadow AI costs

IBM's Cost of a Data Breach Report puts a price on it. In the 2026 edition, security incidents involving shadow AI more than doubled, from 20% to 43% of those studied. Breaches where shadow AI was involved cost USD 5.39 million on average, against a global average of USD 4.99 million and a UK average of USD 4.17 million [1]. Of those incidents, 49% led to data loss or compromise, 42% to operational disruption, 35% to reputational damage, and 21% to a regulatory fine.

The 2025 edition measured it differently: organisations with high levels of shadow AI paid USD 670,000 more per breach than those with little or none [2]. Quote each figure with its year.

One cost never reaches a breach report: content pasted into a consumer account that the organisation cannot see or delete.

Why blocking alone fails

Blocking public AI domains at the firewall is a reasonable first move. On its own it does not hold, for four reasons.

  1. The block covers the network you run. A phone on mobile data and a laptop at home are outside it, and 17% of UK generative AI users already pay for their own tools [4].

  2. A network control cannot see which account is in use. The NCSC notes that a cloud access security broker, without intercepting encrypted connections, cannot identify "unapproved use of approved cloud services (i.e. personal accounts)" [9]. Allowing a domain for your enterprise workspace allows personal accounts on the same domain.

  3. AI arrives inside tools you have not blocked. A domain list does not cover the assistant a software vendor adds in its next update.

  4. People keep going. Asked whether they would keep using AI tools if their workplace banned them and they faced disciplinary action, 37% of decision-makers and 19% of other staff said yes [5].

The NCSC's position is that shadow AI "is unlikely to disappear completely" and that the goal is "to reduce risk rather than assume it can be eliminated" [8]. Its shadow IT guidance tells organisations to avoid unnecessary lockdowns, answer requests for tools quickly, and take a no-blame approach, because staff who fear punishment stop reporting what they use [9].

Keep the block for services you have assessed and rejected, and pair it with an approved route for each job people were using AI to do. The options for that route are in our guide for teams whose public AI tools have been blocked.

How to find shadow AI

No single method finds all of it. Run these in order and combine the results into one inventory.

  1. Ask. Run a short survey with an amnesty: which AI tools people use for work, on which account, for what. This is the only method that reaches personal devices, and it works only if nobody is punished for the answer.

  2. Network logs. DNS, web proxy or cloud access security broker logs show which AI domains devices reach and how often [9]. They show which service was used but usually not which account.

  3. Identity provider. List the third-party apps that staff have granted access to their work account through OAuth, and the sign-ins that used a work email on an AI service you do not manage.

  4. Endpoints. Use unified endpoint management to inventory installed AI desktop apps, command-line tools and browser extensions [9].

  5. Expenses. Search expense claims and card statements for AI subscriptions.

  6. Admin consoles. Check which AI features are switched on in the software you already license.

  7. Source control. Search repositories and CI settings for API keys and configuration that point at AI providers.

For each tool found, record the owner, the account type, the data it receives and the decision: approve, replace or block. Repeat every quarter.

Red, amber and green: the three data classes

A policy that says "do not enter sensitive data" leaves the judgement to the person in a hurry. Three classes with examples remove most of the judgement.

Class

What it covers

Where it may go

Red

Passwords, API keys, tokens and private keys. Customer personal data, health and financial records. Source code that is a trade secret. Anything under a customer NDA. Unreleased financial results, deal terms and board papers. Legally privileged material. Details of an open security incident

No external AI service, on any plan. Only a model on hardware the organisation controls, or no AI at all

Amber

Internal documents and plans. Internal code that is not a trade secret. Routine correspondence with suppliers and colleagues

Only the approved tools, on a business or enterprise plan, signed in with a work account

Green

Published material, public code, general questions with no company detail

Any tool on the approved list

Name real documents from your own organisation as examples in each class, and tell staff that unsure means amber.

Which plan of which tool is fit for amber data differs by vendor. Our AI tools at work hub links a dated guide for each tool, starting with Is ChatGPT safe for work?. How to decide and check where a request is processed is covered in the AI data residency guide.

An AI acceptable use policy outline you can copy

An AI acceptable use policy states which AI tools staff may use, with which data, on which accounts, and what happens when something goes wrong. Square brackets mark what only you can fill in. Have your data protection lead and legal adviser review it before it is issued.

[Organisation] AI acceptable use policy
Version [1.0]. Owner: [role]. Approved by: [role]. Effective: [date]. Next review: [date].

1. Purpose and scope
1.1 This policy covers every AI tool used for [Organisation] work: chat assistants,
AI features inside other software, meeting recorders, coding assistants and agents.
1.2 It applies to employees, contractors and temporary staff, on company and personal
devices, on paid and free accounts.

2. Roles
2.1 [AI owner] keeps the approved tools list in Annex A and decides requests.
2.2 [Security lead] sets the technical controls. [Data protection lead] rules on
personal data.

3. Approved tools and accounts
3.1 Staff use only the tools and plans listed in Annex A, signed in with a work account.
3.2 Personal accounts are not used for work. This includes a personal account on a
tool that Annex A approves.

4. Data classes
4.1 Red data is never entered into an AI tool that sends it outside systems
[Organisation] controls. Annex B lists red data.
4.2 Amber data is entered only into tools that Annex A approves for amber data.
4.3 Green data may be entered into any tool in Annex A.
4.4 If you are unsure of the class, treat the data as amber and ask [AI owner].

5. Requesting a tool
5.1 Anyone may request a tool at [form or channel]. [AI owner] replies within
[5] working days.
5.2 A tool on trial receives green data only.

6. Checking output
6.1 The person who uses an AI output is accountable for it.
6.2 Nothing generated by AI goes to a customer, a regulator or a production system
until a person competent to check it has done so.

7. Connectors, agents and recordings
7.1 Only [AI owner] may connect an AI tool to email, file storage, source control
or a business system.
7.2 A meeting is recorded or transcribed by an AI tool only when everyone present
has agreed.

8. Code
8.1 AI coding tools are used only in the repositories listed in Annex C.
8.2 Secrets are never placed in a prompt, a context file or a tool configuration.

9. Reporting
9.1 If you enter red or amber data into the wrong tool, report it to [channel] the
same day. A report made in good faith is not a disciplinary matter.

10. Monitoring and breaches
10.1 [Organisation] monitors the use of AI services on its network and devices as
described in [monitoring policy or privacy notice].
10.2 Deliberate or repeated breach of this policy is handled under
[disciplinary policy].

11. Review
11.1 [AI owner] reviews this policy and Annex A every [6] months and after any incident.

Annex A: approved tools, the plan, and the data class each may receive
Annex B: red, amber and green data, with examples from [Organisation]
Annex C: repositories approved for AI coding tools

Three clauses decide whether the policy works. Clause 3.2 covers personal accounts, where most unapproved use sits. Clause 5.1 gives a request an answer within a week, which keeps people inside the process. Clause 9.1 means you hear about a mistake while it can still be fixed.

Enforce the red class on the device

A policy tells people where red data may go. It cannot stop a request. For the workflows you bring under it, Aventail™ can.

Aventail™ Link is an inference engine and gateway that runs on each machine. Aventail™ Control holds the routing rules and distributes them ahead of time, and Link evaluates them on the device for each model request. Aventail™ Connect configures supported apps (Claude Desktop, Claude Code, OpenCode and Codex) to send their model requests through Link. A rule names a primary destination (the device, another approved machine running Link, or an approved provider through your own account) and up to three fallback hops. When no permitted destination exists, the request fails with an error. It is not sent to a disallowed destination. The hosted Control service receives management, device and routing events. Prompts and responses do not pass through it.

That gives the policy two lanes:

  • Green and amber: the approved business tool, with the settings in our safe-for-work guides.

  • Red: an assistant routed by Aventail™ under a policy that permits only On device, so the request is processed on that machine or it fails.

Three limits apply. Aventail™ governs model requests routed through Link. It does not discover or block AI tools used outside it, so the discovery steps and the network block still matter. Policies are assigned to devices by tag, so the red lane is a team or a set of machines. Routing to other machines and to hosted providers is in beta, and routing policies need Control for Enterprise.

The Free plan covers three connected devices, including business use. The rule set-up and the test that shows where a request ran are in how to route between local models and approved cloud APIs.

Frequently asked questions

What is an example of shadow AI?
An account manager pastes a customer contract into a personal ChatGPT account to get a summary. Others: a browser extension that rewrites email, a meeting recorder nobody agreed to, a developer running a coding agent with a personal API key.

What is the difference between shadow IT and shadow AI?
Shadow AI is one kind of shadow IT. The difference is what happens to the data: an AI service may retain it or use it to improve a model, and an agent can act with the user's access.

Is using ChatGPT at work shadow AI?
It is if the organisation has not approved that tool, that plan and that account. A personal account is shadow AI even when the company has bought ChatGPT Enterprise.

How do you detect shadow AI?
Combine a no-blame staff survey with network logs, identity provider app grants, endpoint and browser extension inventories, expense claims and the admin consoles of software you already own.

Should we ban AI tools at work?
A ban on its own moves the use to personal devices. In a March 2026 survey, 37% of decision-makers said they would keep using AI tools even if banned. Block what you have rejected and give people an approved tool for the same job.

What should an AI acceptable use policy include?
The approved tools and accounts, data classes with examples, how to request a new tool, who checks output, rules for connectors, agents and recordings, how to report a mistake, and a review date. The outline above has all eleven sections.

Sources

All read on 15 November 2026.

  1. IBM and Ponemon Institute, Cost of a Data Breach Report 2026: https://www.ibm.com/reports/data-breach (full report: https://assets.ibm.com/adobe/assets/urn:aaid:aem:21111142-1251-4369-86fb-57b82f5bb108/original/as/Cost%20of%20a%20Data%20Breach%20Report%202026.pdf)

  2. IBM Newsroom, Cost of a Data Breach Report 2025 press release (30 July 2025): https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls

  3. Microsoft UK Stories, Rise in 'Shadow AI' tools raising security concerns for UK organisations (13 October 2025): https://ukstories.microsoft.com/features/rise-in-shadow-ai-tools-raising-security-concerns-for-uk

  4. Deloitte UK, GenAI Workforce Survey 2026, Shadow AI (fieldwork May to June 2026): https://www.deloitte.com/uk/en/issues/generative-ai/genai-workforce-survey-shadow-ai.html

  5. TrustedTech, Bosses using banned AI are the biggest threat to workplace data security (Censuswide survey, 19 to 24 March 2026): https://www.trustedtechteam.com/blogs/security/shadow-ai-executives-workplace-risk

  6. Microsoft and LinkedIn, 2024 Work Trend Index Annual Report (8 May 2024): https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part

  7. Microsoft, Work Trend Index Annual Reports 2025 (23 April 2025) and 2026 (5 May 2026): https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born and https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization

  8. National Cyber Security Centre, The hidden risks of shadow AI (7 September 2026): https://www.ncsc.gov.uk/sites/default/files/2026-09/The-hidden-risks-of-shadow-AI.pdf

  9. National Cyber Security Centre, Shadow IT guidance (published 27 July 2023, reviewed 14 August 2026): https://www.ncsc.gov.uk/guidance/shadow-it

  10. Aventail™, Connect overview: https://docs.aventail.co.uk/docs/connect/overview

  11. Aventail™, Routing: https://docs.aventail.co.uk/docs/fleet/routing

  12. Aventail™, home page (scope of controls) and plans: https://aventail.co.uk/ and https://aventail.co.uk/#pricing

We use optional cookies and tracking to identify visiting companies, load our contact form, measure Google Ads download clicks and understand page use with PostHog on our homepage, Connect and How it works pages, Ollama and LM Studio comparison guides and campaign pages. Cookie policy