Shadow AI: What Happens When Staff Paste Company Data Into AI

A project manager at a 70-person firm has a client contract open and a deadline in an hour. She pastes the full text into a free AI chatbot on her personal account and asks for a summary. It works, she saves forty minutes, and nobody in the company ever hears about it.

Multiply that by every department and every week, and a company has a steady flow of contracts, customer lists, financial figures and source code moving into tools that IT never approved and nobody monitors. This is shadow AI. This post explains where that data goes, what it can cost, and how to put guardrails in place without taking useful tools away from your team.

What shadow AI actually looks like

Shadow AI is not sabotage. It is ordinary employees trying to get work done faster with tools they already use at home. The common patterns are consistent:

• Pasting contracts, proposals or meeting notes into a chatbot to summarize or rewrite them
• Uploading spreadsheets with customer or payroll data to get help with formulas or analysis
• Pasting code, credentials or configuration files into an assistant to debug a problem
• Installing browser extensions or meeting-note apps that record and transcribe calls
• Using AI features inside tools the company already pays for, but on a personal login

The scale is hard to dismiss. A National Cybersecurity Alliance and CybSafe survey found that 43% of workers reported sharing sensitive work information with AI tools without their employer's knowledge. Different surveys use different samples and definitions, so treat the exact number as a range. The direction is the same everywhere: it is common, and most companies cannot see it.

Where the data goes

The answer depends on the tool and the plan, which is exactly why a blanket assumption does not work.

Consumer accounts. On free and personal plans, the vendor's terms usually govern. Some vendors use conversations to improve their models unless the user turns that setting off. OpenAI, for example, documents a "Improve the model for everyone" toggle in its data controls, which the user controls per account. The company has no say over whether each employee has changed it.

Business and enterprise plans. Paid business plans from major vendors generally state that customer content is not used for training by default, and they add admin controls, logging and contractual terms. The protection comes from the plan and the contract, not from the brand name on the chatbot.

Retention. Even when a vendor does not train on your data, it may keep conversations for a period for abuse monitoring or to give users a history. Deleting a chat from your screen does not always mean it is gone from vendor systems on the same day. Read the retention terms for each approved tool.

Third-party add-ons. Browser extensions, plug-ins and small AI apps often have their own terms, and some have thin security practices. A tool with a handful of users and no published security documentation is a different risk than a major vendor's business plan.

The practical point: once data leaves your environment through a personal account, you cannot recall it, audit it or prove to a client or regulator where it went.

What it costs when it goes wrong

The IBM 2025 Cost of a Data Breach Report found that breaches involving shadow AI added roughly $670,000 to the average breach cost, and that about one in five organizations studied reported a breach tied to shadow AI. The same research found that 63% of organizations had no AI governance policy.

Put that in terms of a growing company. If a breach at a 60-person company would otherwise cost a few hundred thousand dollars in response, legal and downtime costs, the IBM figure suggests that unmanaged AI use can more than double the exposure. Those are averages across many organizations, not a prediction for any one company. The cost drivers are familiar: longer detection time because nobody knew the data was there, notification duties to customers, and contract penalties when client data was exposed.

There are also non-breach costs that never show up in a report:

• Client contract violations. Many client agreements restrict sending their data to third parties. A pasted contract or dataset can breach that clause without any hacker involved.
• Failed security questionnaires. Clients and insurers now ask whether you have an AI policy and how you control AI tools. "We don't know" is a poor answer.
• Loss of confidentiality. Pricing, deal terms and unreleased plans lose their value once copied somewhere you do not control.

Why banning AI does not work

The instinct is to block everything. In practice, blocking the well-known tools sends staff to lesser-known ones, to personal phones, or to home laptops. The data still moves, and now there is even less visibility.

A better approach treats AI the way you treat other business software: approve what is safe, give people a clear path to use it, and watch for the rest.

A practical guardrail plan

The NIST AI Risk Management Framework asks organizations to govern, map, measure and manage AI risk. For a company with no security team, that translates into five concrete steps.

1. Find out what is already in use. Review web filtering and sign-in logs, expense reports and browser extensions to see which AI tools staff actually use. Ask staff directly and make it safe to answer honestly.
2. Approve a short list of tools. Pick one or two business-grade AI tools with training on your data turned off, admin controls and clear retention terms. Give staff accounts on them so the safe choice is also the easy one.
3. Write a one-page acceptable use policy. It should say which tools are approved, which data types never go into any AI tool (client confidential data, health records, credentials, payroll, source code), and who to ask about edge cases. Our AI acceptable use policy template is a starting point.
4. Add technical controls. Use web filtering to block unapproved AI sites and risky extensions, require company sign-in on approved tools, and use data loss controls where your email and file platform support them.
5. Train and review. Cover AI in security awareness training with realistic examples, and review the approved list every quarter because vendors change their terms and features often.

A quick self-check: if a client asked tomorrow which AI tools touch their data, could you answer in one email? If not, steps one and two are where to start. The free Security & AI Risk Scorecard gives a fast read on where you stand.

How managed security handles this

Most small companies do not have anyone whose job is to track AI vendor terms, tune web filters and review extension lists. That work falls between the owner, the office manager and the IT provider, so it does not get done.

Elevaire Managed Security takes that work on as part of running a company's security: web filtering, device and browser controls, security training and a quarterly security review that includes AI tool use. A virtual CISO can write the acceptable use policy, make the risk decisions about which tools to approve, and report the results to leadership.

Example engagement: a 60-person professional services firm learns during a client security questionnaire that it cannot say which AI tools staff use. Elevaire would inventory the AI tools in use, help leadership approve one business-grade assistant, set web filtering to block the rest, draft a one-page policy, and add AI scenarios to the next phishing and training cycle. Elevaire works alongside the firm's existing IT provider, which keeps handling devices and support.

Frequently Asked Questions

What is shadow AI?

Shadow AI is the use of AI tools at work without the company's knowledge or approval. The most common form is an employee pasting company or client information into a public chatbot using a personal account. It is the AI version of shadow IT.

Is it safe to put company data into ChatGPT or similar tools?

It depends on the plan and settings. Free and personal accounts are governed by the vendor's consumer terms, and some use conversations for model improvement unless the user opts out. Business plans typically offer stronger commitments and admin controls. Never put client confidential data, health information, credentials or payroll data into any tool the company has not approved.

How much does it cost to manage AI risk at a small company?

Managed security from Elevaire uses predictable monthly pricing with no hourly billing, and AI tool controls are part of the security work rather than a separate project. For market context, IBM's 2025 research puts the added cost of a shadow AI breach at about $670,000 on average, so prevention is cheap by comparison.

How does this work alongside our existing IT provider?

Elevaire adds the security layer and does not replace your IT provider. Your provider keeps running devices, networks and support, while Elevaire sets the AI policy and security controls, monitors for risky use and reports on it. Both teams work from the same approved tool list.

How do we get started?

Start with a short inventory: ask staff which AI tools they use and check your web filtering logs. Then approve one business-grade tool, publish a one-page policy, and schedule a review for next quarter. Most companies can do this in a few weeks.