Shadow AI: what it is, why it spreads, and how to govern it

Your legal team is drafting contracts with ChatGPT on a personal account. Your finance team is summarizing earnings reports with Claude on the free tier. Your engineers are pasting internal code into Cursor with no corporate agreement in place. You probably don't know about any of it.

That's shadow AI. And according to research from Cybersecurity Dive, roughly half of all employees are using unsanctioned AI tools - with executives leading the way.

This post covers what shadow AI actually costs, why banning it doesn't work, and how to build a governance program employees will follow.

What shadow AI is

Shadow AI is any AI tool used without authorization from IT, security, or whoever owns your technology governance. It's the AI equivalent of shadow IT from the 2010s, but it moves faster because the tools are free, incredibly useful, and available to anyone with a browser.

It includes:

• Consumer AI assistants (ChatGPT, Claude, Gemini) used through personal accounts
• Image generators (Midjourney, DALL-E) on free-tier plans
• AI-enhanced browser extensions without corporate review
• Third-party tools that silently pipe data to AI models (Grammarly, Notion AI, etc.)
• Custom GPTs or AI agents built by employees using company data

The last category is the riskiest. An employee who builds a personal GPT loaded with customer data to help with their job is creating a compliance exposure that a standard SaaS audit will miss entirely.

Why it spreads faster than you think

The speed is the point. An employee hits a bottleneck. They know an AI tool could solve it. They've used it personally. The corporate approved list has nothing similar, or the approval process takes six weeks, or nobody told them there was an approval process. So they use what works.

This is not malice. Research from Netwrix shows that 47% of employees use AI tools through personal accounts specifically to bypass detection - but most don't know that's what they're doing. They think they're being resourceful.

The irony: executives have the highest shadow AI adoption rates. The people setting technology policy are the most likely to be violating it.

Engineering teams follow close behind at 79% adoption of unauthorized tools. Developers have the strongest incentive (AI genuinely cuts coding time by 30-40% on well-defined tasks) and the highest ability to find and use new tools quickly.

What it actually costs

Second Talent's 2026 analysis puts the average enterprise cost at $412K per year. That number covers three categories.

Direct compliance costs. In healthcare, using unauthorized AI to process PHI can trigger HIPAA violations starting at $100 per incident. In finance, using a non-approved tool to summarize client data can violate GDPR or SOC 2 requirements. These aren't theoretical - regulators are catching up faster than most legal teams expect.

Data exposure costs. 38% of employees have shared sensitive company data with unauthorized AI tools. That includes customer PII, internal financial models, source code, and M&A documents. Once that data leaves your environment on a personal account, you have no control over how the vendor uses it for model training, who can access it, or what happens if that vendor gets breached.

Hidden productivity costs. When each team uses different AI tools with no standards, output quality varies and there's no audit trail. A sales team using three different AI writing tools will produce inconsistent messaging. A legal team using an unauthorized tool for contract review won't have version control on the AI's suggestions. The errors don't show up in a security audit - they show up in the quality of work six months later.

Why banning it doesn't work

Every company that has tried to solve shadow AI with a blanket ban has failed. Here's why.

The legitimate productivity gains are real. An employee who uses AI for first drafts works faster. An engineer who uses AI code completion ships more. Banning both the unauthorized tools and failing to provide approved alternatives doesn't eliminate AI use - it just pushes it further underground and ensures the best employees (who have options) resent the policy.

More specifically: employees who see leadership using ChatGPT in meetings while the official policy bans it will not take the policy seriously. The credibility gap is fatal.

What actually works is making the compliant path easier than the unauthorized one.

How to build a governance program that sticks

Step 1: Audit what exists

Before writing a policy, find out what tools people are actually using. Combine two approaches:

• SaaS monitoring tools (Zylo, Torii, BetterCloud) scan your network and identify unauthorized applications by traffic patterns
• Employee surveys with amnesty - tell employees you're building an approved tool list and need to know what's useful. Most will self-report honestly if they know they won't be punished for existing usage

Expect to find 250+ unauthorized tools if you have more than 1,000 employees. That's the Vectra AI estimate for typical enterprise environments.

Step 2: Build a tiered approval system

Not every tool needs the same level of review. A tier system makes approvals faster and less frustrating:

The goal: Tier 1 and Tier 2 approvals should be faster and easier than an employee setting up a personal account. If the bureaucratic path is longer than the workaround, employees will take the workaround every time.

Step 3: Data classification is the core rule

The most important thing employees need to understand is not which tools they can use - it's what data they can put into any AI tool. A simple four-level classification covers most cases:

• Public: Marketing copy, published documents, general industry information. Can go into any approved tool.
• Internal: Company memos, internal process docs, non-sensitive operational data. Approved enterprise tools only.
• Confidential: Customer PII, financial projections, source code, M&A materials. Approved enterprise tools with data handling agreements only.
• Regulated: PHI, PCI data, attorney-client privileged material. Requires specific approved tools with BAA or equivalent. Most AI tools are not approved for this tier.

Put this on one page. Make it a decision tree. If an employee has to read three paragraphs to figure out whether they can use a tool, they'll skip the policy entirely.

Step 4: Write a policy people will read

An effective AI usage policy is under two pages. It covers:

1. What tools are approved and where to find the list
2. How to request approval for a new tool (include a link to the form)
3. The four data classification levels and one example of each
4. What happens if someone violates the policy (be specific - "mandatory training" is clearer than "disciplinary action")

Send it to all employees. Get a signed acknowledgment. Update it quarterly - the tool landscape changes fast enough that a policy from six months ago is already outdated in meaningful ways.

Step 5: Run quarterly reviews

AI tools your company approved six months ago may have changed their data handling terms. New tools may have emerged that deserve approved status. The review cycle keeps the list accurate and maintains the credibility of the program.

The review should take less than a day for a team of two. Check each approved tool's current terms, run the SaaS audit again to find new unauthorized tools, and update the approved list.

The governance model that works in practice

The companies managing shadow AI well have one thing in common: they treat it as a productivity problem, not a security problem.

They start by asking which AI use cases are generating the most value and make sure there's an approved tool for each one. They make the approval process fast enough that employees don't feel blocked. They train on data classification rather than banning categories of tools.

The result: fewer unauthorized tools because employees don't need to go around the system. Better audit trails because approved tools have enterprise-grade logging. Lower compliance risk because data classification rules are understood.

If your current approach is "we banned ChatGPT," you've addressed about 5% of the actual problem. The other 95% is using approved tools to govern what data flows into AI - and that requires a policy employees understand and trust.

The cost of getting it wrong is $412K/year in direct and indirect losses. The cost of getting it right is a two-page policy document and a quarterly review meeting.

That's an easy ROI calculation.

---

Working on your company's AI governance framework? 1Raft helps engineering and ops teams build AI programs with proper data handling, approved tooling, and governance structures that don't slow development to a halt. Talk to us about AI consulting.