Shadow AI is when employees use neural networks, AI assistants, and AI browser extensions for work tasks without the company’s knowledge or oversight: through personal accounts, free service tiers, and browser plugins. The danger is not the use of AI itself, but the fact that the company does not know what data is being sent to external services, who is checking the output, or where to look for mistakes. Shadow AI can be brought under control in just a few weeks: identify actual usage, classify data by sensitivity, give employees an approved alternative, and measure how much AI-enabled work has moved into a managed environment.
Material prepared by AIrassvet using open sources, verified October 10, 2026. It is intended for business owners, CIOs, information security leaders, and department heads at Russian small and midsize companies. This is a practical guide: it includes a detection methodology, a decision matrix, and a 30-day plan. This article is not legal advice: coordinate personal data processing procedures and employee monitoring measures with legal counsel. AIrassvet is not affiliated with the companies or researchers mentioned, and the section at the end describes the publisher’s services.
Contents
- Why Shadow AI Became a Hot Topic in October 2026
- What Shadow AI Is and What It Is Not
- In Brief: Five Steps to Control
- Why Shadow AI Is Dangerous
- Why Bans Don’t Work
- How to Detect Shadow AI
- What to Allow, What to Restrict, and What to Ban
- An Approved Alternative Instead of a Ban
- 30-Day Plan
- Metrics: How to Know Whether Control Is Working
- When This Approach Is Not a Fit
- Frequently Asked Questions
- Primary Sources and Verification Method
- How AIrassvet Helps Bring AI Out of the Shadows
- Bottom Line: Control Starts with Visibility
Why Shadow AI Became a Hot Topic in October 2026
On October 9, 2026, Arenadata Group published the results of a survey of participants in the Digital Solutions forum on AI adoption and governance. The numbers show that AI use in Russian organizations has outpaced the rules:
- AI is already used by 76% of organizations;
- in more than half of organizations, employees use AI without any oversight;
- in 78% there is no working AI usage policy;
- almost 70% of respondents noticed colleagues sharing confidential company information with AI;
- more than half admitted that they themselves had sent data to AI, most often believing it was not confidential;
- DLP protection had been implemented by 26%, a full ban on public models was chosen by 21%, and most organizations use no restrictions at all.
Arenadata Group’s Data Strategy and AI Director, Pyotr Emelyanov, described the problem this way: employees are effectively “hiring” AI assistants themselves, giving them access to email and calendars, files, and corporate systems, and all of this is uncontrollably expanding the attack surface. The sample size was not disclosed in the publication, and participants in an industry forum are not the same as the entire market. But the picture matches foreign research, which we cite below.
What Shadow AI Is and What It Is Not
Shadow AI is a subset of shadow IT, where employees use software and cloud services without approval from IT and security teams. The difference is that an AI service does not just store a file — it processes the content: it reads a contract, summarizes correspondence, or writes a reply to a customer.
Shadow AI includes:
- work prompts sent to public chatbots through a personal account;
- browser extensions that can “read” pages and documents;
- meeting transcription bots that join calls;
- AI assistants that an employee has personally granted access to work email, calendar, or CRM;
- personal accounts in a service the company has officially purchased but does not manage.
The last item is often overlooked. The Teramind The Shadow AI Behavior Report from June 17, 2026 says that 67% of enterprise AI usage goes through unmanaged personal accounts on platforms that companies have licensed and pay for themselves.
Shadow AI is not an officially deployed use case with an owner, nor is it simply employees using AI. If the service is approved, the data is classified, and there is a responsible owner, then it is managed AI, even if thousands of people use it.
In Brief: Five Steps to Control
- Find actual usage: anonymous survey, network logs, spending, extensions, meeting bots.
- Separate data into classes and decide what can be sent to which service.
- Provide an alternative — an approved AI environment that is as convenient as public services.
- Codify the rules in a short policy and assign an owner.
- Measure the share of AI work inside the managed environment and data incidents.
Why Shadow AI Is Dangerous
| Risk | How It Shows Up | Consequence |
|---|---|---|
| Leak of trade secrets | A financial report or strategy is uploaded to a public service for “analysis” | Data leaves the perimeter; retention terms depend on the service |
| Personal data | HR uploads resumes, or legal uploads contracts with clients’ full names | Risk of violating Federal Law No. 152-FZ, including requirements for cross-border data transfer |
| Unreviewed Decisions | The model’s response is sent to the client or included in a document without verification | Errors in code, calculations, and legal wording |
| Access for Assistants | An employee gave an AI service the password to email or CRM | An external service acts on behalf of the employee without a log |
| No Audit Trail | The company does not know which service provided the answer | The incident cannot be analyzed or the root cause fixed |
Separate note on decisions about people. Article 16 of Federal Law No. 152-FZ prohibits decisions that create legal consequences for a person based solely on automated processing of their personal data, except in cases of written consent or a direct provision of federal law. If a hiring manager filters out candidates using a chatbot response through a personal account, the company will not even know that such a decision was made.
The scale of data sharing is also confirmed by international surveys. In a PagerDuty study published on June 11, 2026 (1,250 office workers at companies with revenue of at least $500 million in Australia, Japan, the UK, and the US), 88% admitted sharing work information with public AI tools: 43% shared emails and messages, 34% shared customer data, and 31% shared financial information or confidential documents.
Why bans do not work
Many leaders’ first reaction is to ban ChatGPT and similar tools. The data show that does not solve the problem:
- Employees work around restrictions. According to Teramind, 45% of employees find workarounds when AI tools are restricted, and 48% say they will keep using AI even under a direct ban.
- They violate rules even when they know them. In a PagerDuty survey, 66% of office workers used AI at work while believing company policy did not allow it.
- A ban does not eliminate the need. An employee who has gotten used to drafting emails and summarizing documents with AI faster than manually will not go back to manual work. They will switch to a phone or home computer, and the company will lose even indirect visibility.
That is why we recommend not “ban or allow,” but make the approved path easier than the shadow one. A ban remains a tool for specific data classes and services, not for AI as a whole. The cost of this approach is investment in a corporate environment and training. In return, the company gets a request log, approved models, and the ability to replace personal subscriptions if employees are paying for them.
How to detect shadow AI
Start with methods that do not require monitoring specific people, and move to technical ones only after coordinating them with legal counsel.
1. Anonymous, no-penalty survey
Ask employees which AI services they use, for what tasks, what data they paste in, and how much they pay out of pocket. Make it clear in advance that answers will not lead to penalties. This step gives the most information about the tasks people are trying to solve with AI — and those are exactly the needs the approved alternative should cover.
2. Network logs and DNS
If the company has a proxy, firewall, or corporate DNS, export requests to the domains of popular AI services for the last 30 days. Look at aggregates by department, not by individual. Limitation: this method does not see mobile internet or personal devices.
3. Expenses and subscriptions
Check expense reports and corporate cards for payments to AI services. A common finding is several personal subscriptions in one department that are easier to replace with a single managed one.
4. Browser extensions and apps
If workstations are centrally managed, get a list of installed extensions and applications. AI extensions with access to all pages are one of the least visible data-sharing channels.
5. Meeting bots and email access
Check video meeting participants for automatic transcription services and the list of third-party apps that have been granted access to corporate email and calendar. This is a direct response to the risk described by Pyotr Emelyanov: assistants that employees themselves granted access to.
6. Consolidate into a register
Compile the results into an AI use-case register: task, service, department, data class, and whether an approved alternative exists. The template and register maintenance rules are in the article on the company AI usage policy.
What to allow, restrict, and prohibit
After detection, you need a simple matrix that an employee can understand without a legal background.
| Data class | Examples | Public AI through a personal account | Corporate environment with AI | Model in the company’s environment |
|---|---|---|---|---|
| Public | Website copy, general questions, public documents | Allowed | Allowed | Allowed |
| Internal | Policies, internal correspondence without clients | Not recommended | Allowed | Allowed |
| Confidential | Finance, contracts, strategy, source code | Prohibited | At the discretion of security | Allowed |
| Personal data | Customers, candidates, employees | Prohibited | Only if 152-FZ requirements are met | Allowed if 152-FZ requirements are met |
The table is a starting point, not a finished policy. Specific decisions depend on your contracts with clients, industry requirements, and where requests are physically processed in the selected service. For services that do not store requests, see the article on Zero Data Retention.
An approved alternative instead of a ban
Arenadata’s survey shows that participants agree on one thing: the organization must provide a technically secure environment for working with AI — local models, data anonymization, and revised rules for handling confidential information. In practice, the alternative has three parts:
- Corporate AI assistant. A single entry point with approved models, corporate authentication, and a request log. How to choose and implement it — in the article on a corporate AI assistant for employees.
- Model in the company’s environment for sensitive data. An open model on your own servers or in a Russian cloud covers use cases involving personal and confidential data. Deployment options are covered in the article on open-source LLM inference in a Russian cloud.
- Ready-made workflows for common tasks. An employee survey will reveal 5–10 repetitive tasks: emails, document summaries, meeting notes, formulas. If the company environment has ready-made templates for them, people won’t have any reason to switch to a personal service.
The main success factor for an alternative is convenience. If the corporate tool is slower and requires three approvals, shadow AI will come back.
30-Day Plan
| Week | Actions | Result |
|---|---|---|
| 1 | Anonymous survey, export of network logs and spending, list of extensions and third-party apps | Map of actual AI usage |
| 2 | Use-case register, data classes, allow / restrict / prohibit matrix | A short 2–3 page policy and an AI owner |
| 3 | Launch of an approved environment for 1–2 departments with ready-made workflows for their tasks | Pilot of the alternative with a request log |
| 4 | Employee training, technical restrictions for prohibited data classes, first metrics measurement | Baseline for monthly monitoring |
For a small company, the plan can be condensed: survey and register in one week, then a pilot for everyone at once.
Metrics: How to Tell Whether the Controls Are Working
| Metric | How to Measure | Direction |
|---|---|---|
| Share of employees in the approved environment | Active users of corporate AI / employees using AI according to the survey | Goes up |
| Requests to public AI services | Requests to domains on the list in network logs per month | Goes down |
| Personal AI subscriptions | Number of reimbursements and payments to AI services in expense reports | Falls to zero |
| Data incidents | Recorded cases of sharing prohibited data classes | Declines, each one investigated |
| Register coverage | Use cases with an owner and data class / all identified use cases | Aims for 100% |
Review the metrics monthly for the first three months, then quarterly. A repeat anonymous survey after 90 days will show whether use outside the approved environment has persisted.
When This Approach Is Not a Fit
- A regulated organization with a complete ban on external services. If industry requirements or contracts prohibit any data transfer outside the company, the only approved alternative is a model running within the company environment, and public services are fully blocked.
- A very small team. In a company of 5–10 people, one conversation, a shared corporate subscription, and the rule “no personal data goes into AI” are enough. A full process with network logs is overkill.
- Covert employee monitoring. Technical detection methods need to be reviewed with a lawyer and reflected in internal policies. Secret monitoring of communications creates labor and legal risks that outweigh the benefits of detection.
Frequently Asked Questions
What is shadow AI?
It’s when employees use neural networks and AI assistants for work tasks without the company’s knowledge or oversight: through personal accounts, free service tiers, browser extensions, and meeting bots.
How widespread is shadow AI in Russia?
According to an Arenadata survey published on October 9, 2026, employees in more than half of organizations use AI without oversight, and 78% have no working AI use policy. Nearly 70% of respondents have seen colleagues share confidential information with AI.
Can you just ban employees from using ChatGPT?
A ban without an alternative usually just pushes usage onto personal devices. According to Teramind, 48% of employees will continue using AI even with a direct ban. It’s more effective to prohibit sharing specific data classes and provide a convenient approved environment.
How can you find out which AI services employees are using?
Start with an anonymous, non-punitive survey, then review network logs and DNS, subscription spending, installed browser extensions, meeting bots, and third-party apps with access to email. Coordinate technical methods with legal counsel.
What data should not be sent to public neural networks?
Personal data of customers, candidates, and employees, confidential information—financials, contracts, strategy, source code. These require a corporate environment that meets 152-FZ requirements or a model running within the company environment.
Who in the company should be responsible for shadow AI?
You need one person responsible for AI who maintains the use-case register and approves services, with support from security and legal. Department heads are responsible for making sure their teams work in the approved environment.
Primary Sources and Verification Method
We preserved the publication texts as of October 10, 2026, and checked every figure against the primary source or the author’s official press release. For each survey, the sample and market are listed if disclosed. The detection methodology and metrics are editorial recommendations, not the result of our own research.
- CNews, 09.10.2026: Arenadata survey, forum participants of “Digital Solutions”; sample size not disclosed.
- PagerDuty, 11.06.2026: Shadow AI Survey, 1,250 office employees, Australia, Japan, the United Kingdom, and the United States.
- Teramind, 17.06.2026: The Shadow AI Behavior Report as summarized by distributor Softprom; a survey of 300 security leaders and behavioral telemetry.
- ConsultantPlus: 152-FZ “On Personal Data”, Article 16.
Limitations: overseas surveys describe large companies in other countries and do not directly translate to Russian small businesses; the PagerDuty and Teramind data were published by companies that sell monitoring solutions and may reflect their interests.
How AIrassvet helps bring AI out of the shadows
If employees are already using neural networks, and the company doesn’t know which ones or what data they’re using, the problem can be solved without bans that nobody follows. On the AIrassvet website the areas directly related to it are outlined below:
- Process audit: where and for which tasks employees are already using AI, what data is being used, and where an owner is needed.
- A corporate AI workspace where employees work with approved models instead of personal accounts.
- Deploying a language model on the company’s servers for scenarios involving personal and confidential data.
- Training the team to work in the approved environment.
The first step is to choose one department where AI is already being used informally, describe its current tasks, data sources, constraints, and the criterion you will use to determine that the move to the approved environment has been successful.
Bottom line: control starts with visibility
Shadow AI in a company is not a discipline problem among individual employees, but a sign that the need for AI has outpaced the tools and rules. The Arenadata survey shows that in Russian organizations this is a widespread situation: three-quarters use AI, while only a minority have a working policy in place.
Shadow AI can be brought under control in a month: identify actual usage without hunting for people, classify data into tiers, give employees a convenient approved environment, and measure every month how much AI work has moved into a managed framework. Start with an anonymous survey — it will reveal both the risks and the tasks people need AI for.