Shadow AI is any use of AI tools, models or agents at work that sits outside the tools your organisation has assessed and approved: a personal ChatGPT account used for client emails, a browser extension that summarises pages, an AI feature switched on by default inside a SaaS product, or an autonomous agent such as OpenClaw installed on a laptop. It is the AI equivalent of shadow IT, and in 2026 it is the norm rather than the exception.

How common is shadow AI in the UK?

Very. Microsoft UK's October 2025 survey of 2,003 employees found 71% had used unapproved consumer AI tools at work and 51% did so every week. Okta's March 2026 research put UK staff ignoring AI rules at 55%, while 96% of UK executives said they were confident they could see what tools were in use, the widest gap of any country surveyed. SAP found 68% of organisations report unapproved AI use and 60% of UK firms have not trained staff on AI.

Is it illegal for staff to use ChatGPT at work?

No. Using an AI tool is not illegal in itself. The exposure comes from what goes into it: pasting personal data into a consumer tool can breach UK GDPR and your contracts with customers, pasting confidential material can breach NDAs, and pasting credentials is a security incident. The ICO expects you to know where personal data goes, and since 2 August 2026 the EU AI Act's Article 50 rules apply to any AI that interacts with people in the EU.

Does ChatGPT, Claude or Gemini train on what my staff paste in?

On consumer plans, by default, yes for ChatGPT and Gemini unless the user opts out, and Claude's consumer plans ask users to choose. On business tiers (ChatGPT Business and Enterprise, Claude Team and Enterprise, Gemini inside Google Workspace, Microsoft 365 Copilot) prompts are not used for training by default and you get admin controls, retention settings and audit logs. The difference between a personal account and a business seat is the single cheapest risk reduction available, at roughly $20 to $25 per person per month.

What data should never go into a public AI tool?

Anything that identifies a person (names with contact details, dates of birth, National Insurance or NHS numbers, health or HR matters), payment and bank details, credentials and API keys, unpublished financial results, contract and pricing terms, and anything marked confidential or covered by an NDA. The demo in this article scans text for exactly these patterns on your device and produces a redacted version.

Should we ban AI tools at work?

Bans fail. Microsoft found 28% of UK employees use consumer AI because their employer offers no approved alternative, and staff report saving 7.75 hours a week with it, so the demand does not go away, it goes underground. The pattern that works is to provide sanctioned tools within 30 days, block the specific high-risk categories (browser extensions that intercept prompts, agents with system access), and train people on what not to paste.

What is a shadow AI audit and how long does it take?

A four-week exercise. Week one discovers what is in use from DNS and proxy logs, SSO and OAuth grants, expense claims for AI subscriptions, browser extension inventories and an amnesty survey. Week two classifies the findings by data class and by tool risk. Week three provisions sanctioned alternatives and puts a check in the path staff already use. Week four issues a one-page policy, an hour of training, and the metrics you will track. Cloud First runs these for UK mid-market firms at a fixed price.

What does the EU AI Act require of UK businesses right now?

Since 2 August 2026, Article 50 transparency applies: chatbots must tell people they are dealing with AI, deepfakes must be labelled and AI-generated content must carry machine-readable marks. The Digital Omnibus, in force since 27 July 2026, pushed the high-risk system obligations (recruitment, credit, education and similar uses) back to 2 December 2027, and to 2 August 2028 for AI embedded in regulated products. Any UK business whose AI is used by people in the EU is in scope. The UK itself has no AI statute; a Regulating for Growth Bill with regulatory sandboxes is expected around November 2026.

What about AI agents such as OpenClaw on staff laptops?

Treat them as a separate, higher category. OpenClaw became the most-starred project on GitHub in early 2026 and, by default, gives an agent broad access to files, shell and browser. A one-click remote code execution flaw (CVE-2026-25253) and more than 21,000 exposed instances followed. The NCSC's August 2026 guidance is explicit: always run agents in a sandbox that controls what they can reach, size the controls to the autonomy granted, log everything and keep an emergency stop. An agent on an unmanaged laptop with a corporate login is none of those things.

What does a business AI plan cost per person?

In 2026: ChatGPT Business is $20 per user per month billed annually ($25 monthly, two-seat minimum after OpenAI's April price cut). Claude Team is $20 for a standard seat annually or $25 monthly, with premium seats at $100. Microsoft 365 Copilot is £16.10 per user per month annually in the UK and is now bundled into Business Standard and Premium. Google Workspace includes Gemini in Business plans from roughly $7 to $22 per user per month. For a 50-person firm that is £10,000 to £15,000 a year, against an average shadow AI breach cost IBM puts at $5.39m.

Shadow AI: The Tools Your Staff Already Use, and What to Do in the Next 30 Days

Shadow AI is the normal way UK businesses use AI in 2026. Microsoft UK found 71% of employees have used unapproved consumer AI at work (51% weekly), Okta found 96% of UK executives confident they can see AI use while 55% of staff ignore the rules, the widest gap of any country, and senior managers use unapproved tools at twice the rate of their teams (65% vs 31%, 73% in the C-suite). IBM's 2026 Cost of a Data Breach report puts shadow AI in 43% of incidents at $5.39m each; Cyberhaven finds 39.7% of AI interactions involve sensitive data. The article covers the four layers of shadow AI (consumer accounts, prompt-harvesting browser extensions, AI switched on inside SaaS, and agents such as OpenClaw on laptops), what the ICO's agentic AI report, the NCSC's August 2026 guidance and the amended EU AI Act (Article 50 live since 2 August 2026, high-risk deferred to December 2027) expect, why bans fail, a 30-day audit (discover, classify, sanction, policy and training), a consumer-versus-business tier comparison with 2026 prices, and a live on-device checker that flags and redacts sensitive text before it is pasted into a public AI tool.

Frequently Asked Questions

What is shadow AI?: Shadow AI is any use of AI tools, models or agents at work that sits outside the tools your organisation has assessed and approved: a personal ChatGPT account used for client emails, a browser extension that summarises pages, an AI feature switched on by default inside a SaaS product, or an autonomous agent such as OpenClaw installed on a laptop. It is the AI equivalent of shadow IT, and in 2026 it is the norm rather than the exception.
How common is shadow AI in the UK?: Very. Microsoft UK's October 2025 survey of 2,003 employees found 71% had used unapproved consumer AI tools at work and 51% did so every week. Okta's March 2026 research put UK staff ignoring AI rules at 55%, while 96% of UK executives said they were confident they could see what tools were in use, the widest gap of any country surveyed. SAP found 68% of organisations report unapproved AI use and 60% of UK firms have not trained staff on AI.
Is it illegal for staff to use ChatGPT at work?: No. Using an AI tool is not illegal in itself. The exposure comes from what goes into it: pasting personal data into a consumer tool can breach UK GDPR and your contracts with customers, pasting confidential material can breach NDAs, and pasting credentials is a security incident. The ICO expects you to know where personal data goes, and since 2 August 2026 the EU AI Act's Article 50 rules apply to any AI that interacts with people in the EU.
Does ChatGPT, Claude or Gemini train on what my staff paste in?: On consumer plans, by default, yes for ChatGPT and Gemini unless the user opts out, and Claude's consumer plans ask users to choose. On business tiers (ChatGPT Business and Enterprise, Claude Team and Enterprise, Gemini inside Google Workspace, Microsoft 365 Copilot) prompts are not used for training by default and you get admin controls, retention settings and audit logs. The difference between a personal account and a business seat is the single cheapest risk reduction available, at roughly $20 to $25 per person per month.
What data should never go into a public AI tool?: Anything that identifies a person (names with contact details, dates of birth, National Insurance or NHS numbers, health or HR matters), payment and bank details, credentials and API keys, unpublished financial results, contract and pricing terms, and anything marked confidential or covered by an NDA. The demo in this article scans text for exactly these patterns on your device and produces a redacted version.
Should we ban AI tools at work?: Bans fail. Microsoft found 28% of UK employees use consumer AI because their employer offers no approved alternative, and staff report saving 7.75 hours a week with it, so the demand does not go away, it goes underground. The pattern that works is to provide sanctioned tools within 30 days, block the specific high-risk categories (browser extensions that intercept prompts, agents with system access), and train people on what not to paste.
What is a shadow AI audit and how long does it take?: A four-week exercise. Week one discovers what is in use from DNS and proxy logs, SSO and OAuth grants, expense claims for AI subscriptions, browser extension inventories and an amnesty survey. Week two classifies the findings by data class and by tool risk. Week three provisions sanctioned alternatives and puts a check in the path staff already use. Week four issues a one-page policy, an hour of training, and the metrics you will track. Cloud First runs these for UK mid-market firms at a fixed price.
What does the EU AI Act require of UK businesses right now?: Since 2 August 2026, Article 50 transparency applies: chatbots must tell people they are dealing with AI, deepfakes must be labelled and AI-generated content must carry machine-readable marks. The Digital Omnibus, in force since 27 July 2026, pushed the high-risk system obligations (recruitment, credit, education and similar uses) back to 2 December 2027, and to 2 August 2028 for AI embedded in regulated products. Any UK business whose AI is used by people in the EU is in scope. The UK itself has no AI statute; a Regulating for Growth Bill with regulatory sandboxes is expected around November 2026.
What about AI agents such as OpenClaw on staff laptops?: Treat them as a separate, higher category. OpenClaw became the most-starred project on GitHub in early 2026 and, by default, gives an agent broad access to files, shell and browser. A one-click remote code execution flaw (CVE-2026-25253) and more than 21,000 exposed instances followed. The NCSC's August 2026 guidance is explicit: always run agents in a sandbox that controls what they can reach, size the controls to the autonomy granted, log everything and keep an emergency stop. An agent on an unmanaged laptop with a corporate login is none of those things.
What does a business AI plan cost per person?: In 2026: ChatGPT Business is $20 per user per month billed annually ($25 monthly, two-seat minimum after OpenAI's April price cut). Claude Team is $20 for a standard seat annually or $25 monthly, with premium seats at $100. Microsoft 365 Copilot is £16.10 per user per month annually in the UK and is now bundled into Business Standard and Premium. Google Workspace includes Gemini in Business plans from roughly $7 to $22 per user per month. For a 50-person firm that is £10,000 to £15,000 a year, against an average shadow AI breach cost IBM puts at $5.39m.

Our Services

Contact Cloud First Consulting

Email: info@cloudfirstconsulting.com

Location: London, United Kingdom

Hours: Monday-Friday, 9:00 AM - 6:00 PM GMT

Book a Free 30-min Discovery Call