What Is Shadow AI? A Plain-Language Guide

Shadow AI is any artificial intelligence tool an employee uses for work without IT or security approval, from pasting client data into ChatGPT to running an unsanctioned AI coding assistant or browser plugin. It matters because these tools often send your company's data to third-party servers outside your control, creating privacy, contractual, and compliance exposure that most organizations only discover after something goes wrong.

What Is Shadow AI, Exactly?

Shadow AI is the AI version of shadow IT, the long-standing problem of employees adopting software without going through procurement or security review. The difference is scale and stakes. A single employee can sign up for a free AI tool in minutes, and once it is in daily use, it often touches source code, customer records, financial data, or legal documents.

Common examples we see when we run a shadow AI audit for clients include:

• Consumer chatbots (ChatGPT, Gemini, Claude web apps) used to summarize contracts or draft emails containing customer information
• AI coding assistants installed directly in developer IDEs, sometimes with default settings that retain code snippets for model training
• AI-powered browser extensions and meeting transcription tools that record and store calls off your infrastructure
• Marketing and sales teams using AI writing or research tools connected to CRM exports
• Freelancers or contractors running their own AI tooling against files you shared with them

None of this is usually malicious. Employees adopt these tools because they genuinely make work faster. The problem is that nobody mapped where company data actually goes once it leaves your systems.

Who Needs to Worry About Shadow AI

Any organization handling customer data, source code, or regulated information has shadow AI exposure, but a few situations raise the stakes considerably. Companies pursuing SOC 2 or ISO 27001 will find shadow AI is now a standard line item in vendor risk and data handling reviews, and auditors increasingly ask direct questions about it. Businesses subject to PIPEDA, or to Quebec's Law 25, carry added risk because personal information routed through an unvetted AI tool may count as an undisclosed third-party data transfer.

We also see this come up constantly with Canadian tech companies expanding into the US market. A SOC 2 report with a gap around AI tool governance is a real deal-blocker for American enterprise buyers, and it is one of the first things a security questionnaire will surface. Teams in Toronto, Waterloo, Ottawa, and Vancouver building SaaS products for US customers are the clients who ask us about this most often, because their prospects' security teams ask them about it first.

If your company has more than a handful of employees and no written AI usage policy, you almost certainly have shadow AI in use today. The absence of a policy is not evidence of the absence of the problem, it is usually evidence that nobody has looked yet.

What a Shadow AI Discovery Process Actually Involves

Discovering shadow AI is not a single scan. It is a combination of technical detection and organizational conversation, because a lot of usage happens on personal devices or free-tier accounts that never touch corporate infrastructure.

A proper shadow AI audit typically covers:

• Network and DNS review, identifying outbound traffic to known AI vendor domains from company devices and cloud environments
• SaaS and browser extension inventory, since many AI tools are installed as plugins rather than standalone applications
• Interviews with department leads in engineering, sales, marketing, and support, who usually know exactly which tools their teams rely on informally
• Data flow mapping for any AI tool found, to determine what categories of data it touches and where that data is retained or processed
• Contract and vendor terms review for any AI tools already in semi-official use, checking data retention, training opt-outs, and subprocessor lists

The output should be a plain-language inventory: which tools are in use, what data they touch, what the risk level is, and what to do about each one, whether that is banning it, replacing it with an approved alternative, or formally sanctioning it with proper contractual protections in place.

How Long Does a Shadow AI Audit Take?

For a small to mid-sized company, an initial discovery pass usually runs one to three weeks. The network and SaaS discovery work happens quickly, often within days, but the interview and data flow mapping stages take longer because they depend on scheduling time with team leads across departments. Larger organizations with multiple business units or a distributed workforce should expect four to six weeks for a thorough first pass.

This is not a one-time project. New AI tools appear constantly, and free-tier consumer AI products change their terms and data handling practices without much notice. Most of our clients pair the initial audit with a lightweight quarterly review and a written AI usage policy, so shadow AI does not quietly creep back in six months later.

Common Misconceptions About Shadow AI

A few misunderstandings come up in almost every conversation we have with clients on this topic.

"We banned ChatGPT, so we're covered." Blocking one tool rarely stops usage, it just pushes employees toward a different tool, often one with worse data handling practices, or toward personal devices where you have no visibility at all.

"This is an IT problem, not a compliance problem." Shadow AI sits squarely at the intersection of both. IT can detect the tools, but the risk assessment, the regulatory exposure under PIPEDA or Law 25, and the vendor contract review require compliance and security judgment, not just a network scan.

"Our developers only use AI for code, so there's no data risk." Source code is company intellectual property, and depending on the coding assistant's settings, snippets can be retained or used for model training. For companies in regulated industries or pursuing SOC 2, this is treated the same as any other sensitive data exposure.

"We're too small for this to matter." Smaller companies often have less mature access controls, which means a single employee's AI tool choice can touch a much larger share of total company data than it would at a larger, more segmented organization.

Getting Ahead of Shadow AI

The fix is not to eliminate AI use, most teams get real productivity value from it. The fix is visibility and governance: know what is in use, understand what data it touches, and make a deliberate decision about each tool rather than discovering it during a customer security review or an audit. This is also foundational work if you are heading toward ISO 42001 AI governance readiness, since a documented AI inventory is one of the first things that framework expects.

traztech runs shadow AI audits for companies across Canada, from Toronto and Ottawa to Calgary and Montreal, built around the compliance frameworks Canadian tech companies are actually being asked for. If you suspect your team has unsanctioned AI tools in daily use and want a clear picture of what is happening with your data, contact traztech to talk through a shadow AI audit.

The OAuth Grant Is the Part People Miss

Most discovery efforts look for accounts and subscriptions. The exposure that actually causes trouble is quieter: an employee clicks Continue with Google on an AI note-taker or an AI email assistant and grants it a scope like read and send mail, or read all files in Drive. No card was charged, no software was installed, and nothing appears on an expense report. The tool now holds a refresh token against your tenant and can read that data at three in the morning without the employee present.

Both Google Workspace and Microsoft 365 let an administrator list every third-party application with a grant, the scopes it holds, and the number of users who approved it. Pull that list first. It is the highest-yield hour in a shadow AI review, and it usually surprises the person who ordered the review. While you are there, check whether user consent to third-party apps is still on by default, because moving it to admin approval, with a request workflow so people are not simply blocked, stops the next twenty grants without a policy document. Do the same for browser extensions, since one with permission to read and change all data on websites you visit sees every customer record your staff open.

Enterprise Tiers Change Less Than the Sales Deck Suggests

A common resolution is to buy the business tier of whichever AI tool people were already using and consider the problem closed. That does help, because the enterprise agreements of the major providers generally exclude your prompts from model training and give you retention controls and an administrative view. It does not close three gaps. Personal accounts on the same tool keep working from the same laptop unless you enforce tenant restrictions, so paying for the enterprise tier without enforcement funds a policy that nothing enforces. Connectors and plugins inside the enterprise product may carry their own terms and their own subprocessors. And retention set to zero at the model provider says nothing about the retention of the intermediate vendor if your staff reached the model through a wrapper product.

The question to ask any AI vendor is not whether they train on your data. It is who processes the data, where it is stored, for how long, whether human reviewers can see it, and which of those answers is contractual rather than a setting somebody can flip back.

The Contract Problem You Inherit

Shadow AI is often a customer contract issue before it is a security issue. Enterprise master service agreements increasingly carry a clause requiring notice or consent before adding a subprocessor, and some now prohibit submitting customer data to generative models entirely. If a support agent pastes ticket contents from a regulated customer into a consumer AI tool, you have likely breached a term you signed, quite apart from any privacy law. Under PIPEDA and Quebec Law 25 the transfer is a disclosure to a third party that your privacy notice may not cover, and Law 25 adds obligations around decisions made using automated processing.

Two practical steps. Read your top ten customer agreements for subprocessor and AI language before you decide what to permit, because the strictest contract sets your floor. Then add approved AI tools to your public subprocessor list and keep it current, since buyers now diff that page between renewals. Handling that properly sits alongside the rest of your compliance program rather than in a separate AI initiative.

When You Should Not Buy a Shadow AI Audit

If you are under about thirty people, everyone uses the same three tools, and your administrator can list every OAuth grant in an afternoon, you do not need an engagement. Do the tenant review yourself, restrict third-party consent, publish a one-page acceptable use rule that names what may never be pasted anywhere, and recheck in six months. That is a Friday, not a project.

You also should not buy discovery tooling as the first move. Continuous discovery platforms are priced for organizations with thousands of endpoints and a team to work the alerts, and buying one before you have decided what is permitted produces a dashboard nobody acts on. Make the decision first, then buy enforcement if the population justifies it. The engagement earns its cost when you have several hundred staff, regulated or contractually restricted data, an unclear picture of which teams shipped AI features into the product, and a buyer or auditor asking questions you cannot currently answer. If that is you, our security services page sets out how the review is scoped.