What Is Shadow AI? | TMA Glossary

Shadow AI is any use of AI tools, models, or agents inside a company that IT, security, or compliance has not approved or cannot see. An analyst pastes a client contract into a free chatbot. A team wires an agent to the shared drive with a personal API key. A browser extension summarizes every page an employee opens.

The people doing it usually mean well. They want to get work done faster.

How common it is

Microsoft’s 2024 Work Trend Index (8 May 2024) surveyed 31,000 knowledge workers across 31 markets. It found 75% were using AI at work, and 78% of AI users were bringing their own AI tools to work.

IBM’s 2025 Cost of a Data Breach report (30 July 2025) added the security angle. One in five organizations reported a breach due to shadow AI. Organizations with high levels of shadow AI saw an average of $670,000 higher breach costs than those with low levels or none. Of the organizations that reported an AI-related breach, 97% said they lacked AI access controls. IBM noted that the share of organizations with an AI-related breach was small, but said this was the first year the report studied AI security and governance.

Why it happens

Approved tools are slow to arrive, or do not fit the work.
Free tools take a minute to start. A procurement review takes months.
Nobody told employees what is allowed.
Leaders want AI results and do not ask how they were produced.

Banning AI does not fix this. It pushes the use out of sight. The Work Trend Index figures above suggest many workers will bring tools anyway.

What the risks are

Data leakage. Customer data, source code, or contracts go to a provider with no agreement in place.
No audit trail. You cannot answer “who used what data, and when?”
Unreviewed agents with real access. A homemade agent with a shared credential is a non-human identity nobody inventoried.
Compliance exposure. Rules on personal data, records, and automated decisions apply whether or not IT knew about the tool. The EU AI Act guide covers what to track.
Inconsistent output. Teams make decisions from tools nobody tested.
Leaked credentials. Personal API keys end up in scripts and repos. OWASP lists secret leakage and long-lived secrets in its Non-Human Identities Top 10, and shadow tools produce both.

Shadow agents deserve the most attention. A chatbot leaks what you paste. An AI agent can also act. If someone connects one to email or a CRM and a prompt injection lands, the agent acts with that person’s access.

A practical response

I would work in this order:

Find it. Check network logs, SSO app grants, browser extension reports, expense reports, and API billing. Then ask people. A short, no-blame survey surfaces more than a scan.
Classify the data. Decide what must never leave: customer PII, health data, source code, unreleased financials.
Offer a better path. Approve a small set of tools, with data agreements, that cover the common jobs. If the sanctioned tool is worse than the free one, people will keep using the free one.
Set simple rules. One page. What is allowed, what is not, who to ask.
Add controls. Single sign-on for approved tools. Data loss prevention on the riskiest data. Egress rules for known AI endpoints.
Keep an inventory. Every agent has an owner, a purpose, a credential, and a review date.
Repeat. The tool list changes monthly.

For a wider checklist, see the enterprise AI security checklist.

A short example

Say a sales team builds an agent that reads call transcripts and drafts follow-ups. One rep set it up with a personal API key from a consumer account, and transcripts include customer names and pricing. A review finds it, and security does not shut it down. They move it to a company account with a data agreement, scope it to the sales team, and add logging. The team keeps the productivity. The company gains visibility.

Common mistakes

Blanket bans without alternatives.
A policy nobody reads.
Treating shadow AI as an employee discipline problem instead of a tooling gap.
Counting only chatbots and missing agents, plugins, and automations.
No owner. Governance that belongs to everyone belongs to no one.

Signs it is already happening

You likely have shadow AI if any of these are true: staff cannot name the approved AI tools; expense reports show AI subscriptions that IT did not buy; browser extension reports list AI assistants; developers have personal API keys in repos or scripts; or a team demos a useful automation nobody in IT has seen. None of these mean bad intent. They mean people found a faster way and had no sanctioned one to use.

AI guardrails limit what approved tools can do. An AI audit trail is the record that makes oversight possible. Non-human identity covers the credentials shadow agents depend on.