Shadow AI is any AI tool, AI feature or AI agent that employees use for work without the organisation’s approval or oversight, such as a personal ChatGPT account, an AI browser extension or an AI note-taker. To find it, combine network, SaaS, browser and endpoint telemetry with a no-blame employee survey. To stop data leaking into it, give people an approved alternative, set clear data red lines in an AI acceptable use policy, enforce them with prompt-level data loss prevention (DLP), and train every employee. Banning AI on its own usually pushes the same behaviour onto personal devices.
This guide is for CISOs, CIOs and risk heads in Indian enterprises and GCCs who need a practical way to get shadow AI under control without slowing the business down.
What counts as shadow AI?
Shadow AI is broader than “people using ChatGPT”. In most organisations it includes:
- Public chatbots on personal accounts: ChatGPT, Gemini, Claude, Perplexity, DeepSeek and others, used on free or personal plans.
- Personal accounts on approved tools: an employee logged into a personal ChatGPT account on a company laptop, even though the company has an enterprise licence.
- AI browser extensions: writing assistants, summarisers and “chat with this page” tools that can read everything in the browser.
- AI features inside approved SaaS: new AI assistants switched on inside CRM, HR, design or collaboration tools, often without a fresh security review.
- AI meeting note-takers that join calls and send transcripts to a third-party service.
- Coding assistants and local models used by developers outside the approved toolchain.
- AI agents that employees connect to email, calendars, drives or internal systems with their own credentials.
Shadow AI is a form of shadow IT, with one important difference: the risk is not only the tool itself but what people type, paste and upload into it, every day.
Why shadow AI is a board-level risk in 2026
The numbers from credible studies are hard to ignore:
- In IBM’s 2025 Cost of a Data Breach report, one in five organisations reported a breach due to shadow AI. Organisations with high levels of shadow AI saw an average of USD 670,000 in higher breach costs than those with low or no shadow AI. Of the organisations that reported a breach of AI models or applications, 97% said they lacked AI access controls, and 63% of breached organisations either had no AI governance policy or were still developing one.
- In India, IBM’s 2026 study found shadow AI was one of the top three factors that increased breach costs, adding an average of ₹1.79 crore where present, against a record average breach cost of ₹25.5 crore.
- Harmonic Security analysed one million generative AI prompts and 20,000 uploaded files from enterprise users in the second quarter of 2025. It found that 4.37% of prompts and 21.86% of uploaded files contained sensitive data. ChatGPT accounted for 72.6% of the sensitive prompts, and despite wider use of enterprise plans, 26.3% of sensitive exposure still went through the free version of ChatGPT.
- Usage is already mainstream. In Microsoft and LinkedIn’s 2024 Work Trend Index, 72% of Indian AI users said they bring their own AI tools to work.
The best-known example is still Samsung. In May 2023, Bloomberg reported that Samsung had restricted employee use of generative AI tools after discovering that staff had uploaded sensitive internal code to ChatGPT. It shows how a well-meaning employee, trying to work faster, can create a data incident in seconds.
For Indian companies there is also a regulatory angle. Under the DPDP Act 2023, a data fiduciary must protect personal data in its possession with reasonable security safeguards. An employee pasting a customer list or employee records into an unapproved AI tool is exactly the kind of exposure those safeguards are meant to prevent, and the DPDP Rules 2025 bring the core security and breach notification obligations into force from 13 May 2027.
The FIND-DECIDE-PROTECT-TEACH-RESPOND framework
We teach security teams a five-step approach. Each step has a short checklist you can use straight away.
1. Find: build an honest inventory
No single data source shows all shadow AI. Combine several:
| Discovery method | What it finds | Blind spots |
|---|---|---|
| DNS, proxy and firewall logs | Which AI services are being reached, how often and by which teams | Personal devices, home networks, what data was sent |
| SaaS and identity (SSO, OAuth) review | AI apps connected to corporate accounts; third-party apps with access to mail and drives | Tools used without corporate login |
| Browser extension inventory | AI extensions that can read page content | Unmanaged browsers |
| Endpoint telemetry | Desktop AI apps, coding assistants, local models | Unmanaged devices |
| Prompt-level monitoring | What data is actually going into AI tools | Needs a deployed control and a clear privacy notice to employees |
| Anonymous employee survey | Why people use each tool and what job it does for them | Self-reporting bias |
The survey matters more than most security teams expect. Shadow AI is usually a signal of unmet demand. If you know which jobs people are doing with AI, you can give them a safe way to do those jobs.
2. Decide: sort tools into approved, restricted and blocked
- Approved: enterprise tools with contractual data protection, SSO and admin controls. Make these easy to get.
- Restricted: allowed for non-sensitive work only, or for specific teams.
- Blocked: tools with unacceptable data terms, unclear data residency or no enterprise controls.
Write the decision rules down (data handling terms, data retention, training on customer data, residency, admin controls) so new tools can be assessed in days, not months.
3. Protect: put controls at the point of risk
- Block personal accounts on approved AI tools from corporate devices and allow the corporate tenant.
- Deploy prompt-level DLP that detects sensitive data (personal data, credentials, source code, client identifiers) and warns, redacts or blocks it before it reaches a model.
- Review and restrict AI browser extensions on managed browsers.
- Require a security review before AI features are switched on inside existing SaaS tools.
- Limit the permissions of any AI agent connected to company systems, and log what agents do.
Platforms such as AccuroAI’s shadow AI discovery and prompt DLP are built for this layer. Our hands-on labs use the AccuroAI platform, but the principles apply to whatever stack you run.
4. Teach: make the safe path the easy path
Controls catch mistakes; training prevents most of them. Every employee should know three things: which tools are approved, what data must never go into a prompt, and how to report a mistake without fear. Our Safe AI at Work programme covers this for the whole workforce in about 90 minutes. Back it with a short AI acceptable use policy that people can read in five minutes.
5. Respond: treat AI data exposure as an incident type
- Add “sensitive data entered into an unapproved AI tool” to your incident categories.
- Define the first steps: identify the data and the tool, request deletion where the provider allows it, revoke exposed credentials, and assess whether personal data was involved.
- Check whether the exposure is a personal data breach that needs notification under the DPDP framework, with legal and the DPO involved early.
- Feed lessons back into policy, controls and training.
A 30-day starter plan
- Week 1: pull 30 days of DNS and proxy logs for known AI domains; list the top 20 tools by users.
- Week 2: run an anonymous survey; review OAuth grants and browser extensions.
- Week 3: publish the approved, restricted and blocked list and an interim AI acceptable use policy; make sure at least one approved tool covers the most common use cases.
- Week 4: pilot prompt-level DLP with one high-risk function (for example, finance or customer support) and schedule company-wide training.
If your security and IT team would rather learn this hands-on, our one-day Shadow AI & Prompt Data Leakage Defense workshop takes them through discovery, prompt DLP policy and AI incident response in live labs. Teams building their own agents should look at Securing AI Agents.
How to measure progress
- Share of AI usage going through approved tools (target: rising every month)
- Number of unique unapproved AI tools in use (target: falling)
- Sensitive-data detections per 1,000 prompts, by type and team
- Training completion and assessment scores by business unit
- Time to triage an AI data exposure incident
Frequently asked questions
Should we just block ChatGPT and other AI tools?
Blocking unapproved tools has a place, but a blanket ban without an approved alternative usually fails. People move to personal phones and home laptops, and you lose visibility completely. Block what is unsafe, approve what is safe, and make the approved option easy to use.
Is an enterprise licence for ChatGPT, Copilot or Gemini enough?
It helps, because enterprise plans offer contractual data protection and admin controls. It does not stop employees using personal accounts, other tools or AI browser extensions, and it does not stop someone pasting data they should not share. You still need discovery, DLP and training.
What is the difference between shadow IT and shadow AI?
Shadow IT is any technology used without IT approval. Shadow AI is the AI subset, and it is riskier in one specific way: employees continuously send company data into it through prompts and uploads, and the output can flow straight into decisions, code and client work.
Is shadow AI a DPDP compliance issue?
It can be. If personal data of customers or employees is entered into an AI tool without proper safeguards, the organisation, as data fiduciary, is responsible. The DPDP Act sets penalties of up to ₹250 crore for failing to take reasonable security safeguards to prevent a personal data breach.
How long does it take to get shadow AI under control?
Most organisations can get a credible first inventory and interim policy in place within 30 days. Full control, with DLP, training and an incident process running, typically takes one to two quarters.
Get your teams ready
Product Innovation Academy runs shadow AI and AI security training for Indian enterprises and GCCs in partnership with AccuroAI: a one-day hands-on workshop for security and IT teams, and company-wide awareness training for every employee. Request a proposal and we will tailor it to your tools, policies and risk profile.
Sources
- IBM Newsroom, “IBM Report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls”, 30 July 2025. newsroom.ibm.com
- IBM India Newsroom, “India Records its Highest Average Cost of a Data Breach at INR 255 million (INR 25.5 crore) in 2026”, 3 August 2026. in.newsroom.ibm.com
- Harmonic Security, “GenAI Data Exposure: What GenAI Usage Is Really Costing Enterprises”, 31 July 2025. harmonic.security
- Microsoft News Centre India, “92% of Indian knowledge workers use AI in the workplace, finds Microsoft and LinkedIn 2024 Work Trend Index”, 16 May 2024. news.microsoft.com
- Bloomberg, “Samsung Bans Staff’s AI Use After Spotting ChatGPT Data Leak”, 2 May 2023. bloomberg.com
- Press Information Bureau, “DPDP Rules, 2025 Notified” (backgrounder, including penalty schedule), 17 November 2025. pib.gov.in
- MirvoLegal, “DPDP Rules, 2025: Compliance Timeline for Data Fiduciaries” (reviewed 20 September 2026). mirvolegal.com