Quick Answer
Shadow AI is the use of AI tools, models, agents, and AI features inside a company without IT approval, security review, or spend governance. Currently, it is the norm, not the exception: 66% of office professionals at large enterprises admit using AI tools they believed weren't permitted, per PagerDuty's June 2026 survey.
In February 2026, Meta told employees that installing an AI agent called OpenClaw on work devices reportedly put their jobs at risk. By March, Chinese authorities had restricted it across government agencies and state-owned enterprises, including the country’s largest bank. By May 1, Microsoft had shipped Agent 365, which governs local agents starting with OpenClaw by name and lets administrators block it on managed devices. Microsoft’s security team said OpenClaw “should be treated as untrusted code execution with persistent credentials.”
Here’s the part those headlines skip: OpenClaw wasn’t malware. It was the fastest-growing project in GitHub’s history, past 390,000 stars by October 2026, installed by employees because it saved them hours, running on their personal API keys, invisible to every system the company used to watch its software. Token Security found the tool, which it tracks under the earlier name Clawdbot, running inside 22% of the customer environments it monitors.
That’s shadow AI in one story: enormously useful, individually rational, collectively ungoverned. The security industry has written its half of the response thoroughly. The half nobody has written is the one finance keeps asking: this stuff isn’t just risky, it’s being paid for, on cards and expense reports, through free tiers that charge in data, in duplicate across teams that each bought their own answer.
CloudZero’s entire product exists on that second half, turning every AI dollar, sanctioned or not, into visible, allocated cost.
What is shadow AI?
Shadow AI is the use of artificial intelligence tools, models, agents, or AI features inside approved software without the knowledge, approval, or governance of IT, security, or finance. The term covers personal ChatGPT accounts doing work tasks, developer API keys on personal cards, AI features toggled on inside sanctioned SaaS, and locally installed agents like OpenClaw.
What is shadow AI versus shadow IT is mostly a question of speed and shape. Shadow IT took years to accumulate; shadow AI arrived at scale almost immediately. In a November 2025 Gartner survey of 302 cybersecurity leaders, 69% suspected or had evidence that employees were using prohibited public generative AI tools.
The shadow AI meaning also stretches further than software: an AI feature inside an approved tool can be shadow AI when nobody reviewed what data it touches or what it costs when usage-based billing kicks in.
One more definitional edge, because the searches ask: the shadow AI definition does not require bad intent. Nearly every study on the subject finds employees adopt these tools to work faster, not to exfiltrate data. That’s precisely what makes bans fail, and we’ll show the failure numbers below.
2026 State of AI Spend Report
Why Finance Caps What It Can’t See
Learn how finance leaders are managing AI spend, and how they’re approaching FY2027
What does shadow AI look like in practice?
Shadow AI examples cluster into five patterns, ordered by how invisible they are to a finance team:
| Pattern | What it looks like | Where the money hides |
|---|---|---|
| Personal subscriptions | An employee expenses ChatGPT Plus or pays personally | Expense categories, personal cards |
| Personal API keys | A developer bills OpenAI or Anthropic usage to a card | Card statements labeled as software |
| Free tiers | Free ChatGPT, Claude, or Gemini accounts doing work tasks | No invoice at all; the payment is company data |
| Features inside approved SaaS | AI add-ons toggled on in sanctioned tools | Usage-based line items inside trusted bills |
| Local agents | OpenClaw-class software on work machines, personal keys attached | Nowhere; it defeats SaaS inventories entirely |
The free-tier row deserves a sentence, because “free” is doing heavy lifting: BlackFog’s study, published January 2026 and fielded that previous November, found 34% of employees use free versions of company-approved AI tools, which means the enterprise paid for governance and the employee opted out of it at a price of zero dollars and unknowable data.
This whole phenomenon is the AI shadow economy: a parallel software estate, adopted bottom-up, financed through expense reports and data, running alongside the official one.
How widespread is shadow AI in 2026?
The honest answer is that shadow AI is the majority behavior, documented by every named survey that has looked:
| Finding | Figure | Source (year) |
|---|---|---|
| Office professionals at $500M+ enterprises who used AI tools they believed weren’t permitted | 66% | PagerDuty Shadow AI Survey (June 2026) |
| Employees using unapproved AI tools | 81% | UpGuard State of Shadow AI (2025) |
| Security leaders themselves using unapproved AI | 88% | UpGuard (2025) |
| Organizations that suspect or have evidence of prohibited genAI use | 69% | Gartner survey of 302 security leaders (2025) |
| Shadow AI tools active at the average enterprise | 158+ | Gartner AI Governance Survey (2025) |
| AI users who brought their own tools to work | 78% | Microsoft/LinkedIn Work Trend Index (2024) |
| Companies with shadow IT, with AI adopted without security review | 70% | Vanta (2026) |
| Workers who would find workarounds if tools were blocked | 45% | UpGuard (2025) |
Vanta data reported by Fortune in April 2026 adds the detail that should end any remaining complacency: the average company sees staff reinstall an AI tool 1,000 times a year after security has revoked it, with Claude, ChatGPT, and Cursor the most reinstalled. The workforce has voted, repeatedly, after losing.
One trend line points the right direction, with a catch. Netskope’s 2026 Cloud and Threat Report found personal-app usage among genAI users fell to 47% from 78% a year earlier as sanctioned tools rolled out, yet the share of users switching between personal and company AI accounts more than doubled.
Adoption migrates when the approved path exists; it doesn’t fully consolidate on its own. The wider adoption numbers, sanctioned and not, live in our AI statistics library.
What does shadow AI actually cost?
Shadow AI risks to the budget come in three layers, direct invisible spend, duplicate spend, and priced breach risk, and two of the three bill monthly whether or not anything ever goes wrong. Every ranking page prices only the third. That’s real, and it’s also not the whole invoice.
Layer one is direct invisible spend. Take a 1,000-employee company at the PagerDuty rate: 660 people using unapproved AI. If just a quarter of them carry a paid tool at $20 a month, expensed or reimbursed, that’s 165 subscriptions, $3,300 a month, $39,600 a year of AI spend filed under travel-and-expense noise where no AI cost management process will ever review it.
Scale the same arithmetic to API keys, where a single busy developer’s personal key can bill hundreds a month, and the invisible line item grows fast. What a legitimate seat costs is public arithmetic (ChatGPT’s price card is one click away); what makes this spend expensive is that nobody can see it to manage it.
Price the free tier into this layer too, because its cost is denominated in something finance never audits. Free consumer accounts process work content under consumer terms: no enterprise data processing agreement, no retention guarantees, training-use defaults that vary by vendor. The cost is real, deferred, and paid in negotiating position and exposure rather than dollars, which is exactly why it never appears in a budget review.
Layer two is duplicate spend. That same 1,000-employee company almost certainly also bought an approved assistant, and the unapproved estate sprawls regardless, so the same capability gets purchased two, three, or ten times across teams.
Duplication is the quiet majority of waste in how much AI really costs an organization, and shadow adoption guarantees it, because buyers can’t coordinate on tools they can’t see. Tool-count proliferation is its own discipline, covered in our AI sprawl analysis.
Layer three is the priced risk the security pages do cover, so we’ll keep it to the two numbers finance needs for the model: IBM’s 2026 breach research puts shadow AI in 43% of security incidents, more than double the prior year, against a global average breach cost of $4.99 million. Multiply by your own probability; the expected value isn’t zero.
Add the layers and the pattern is familiar to anyone who lived through shadow IT: the visible risk gets the headlines while the invisible spend compounds. The difference is speed. Shadow IT took a decade to build its estate; shadow AI reached majority adoption in roughly three.
Why do employees use shadow AI?
Because it works, and because waiting doesn’t. BlackFog’s 2,000-respondent study found 60% of employees would accept security risks to meet deadlines, and 86% already use AI at least weekly. Software AG’s 6,000-worker study, from October 2024, found 46% would keep using personal AI tools even if their employer banned them outright. UpGuard’s 2025 State of Shadow AI measured the ban’s half-life directly: 45% of workers find workarounds to blocked applications.
The OpenClaw episode is the controlled experiment: enforcement escalated from employer bans to national restrictions to a dedicated Microsoft blocking policy, and the tool kept spreading through personal devices and personal keys anyway. When the world’s most resourced enforcement can’t hold the line, the lesson isn’t “enforce harder.”
The lesson is that shadow AI is a demand signal wearing a compliance costume. Employees are telling you which capabilities the approved stack lacks, at zero survey cost, with their own money. Governance that reads the signal beats governance that burns it.
How do you find shadow AI spend?
Follow the money: most shadow AI eventually touches an expense report, a card, or a bill, and those leave records finance already owns. Security teams hunt it in network logs and OAuth grants, a technical discipline of its own; the spend trail is the shorter path.
Four places it surfaces: expense reports (search reimbursement categories for AI vendor names and app-store charges), corporate and personal card statements (recurring $20-to-$200 software charges that match no approved vendor), SaaS bills (usage-based AI line items appearing inside sanctioned products), and cloud bills (API spend from token-metered providers nobody onboarded). The free-tier pattern leaves no money trail at all, which is why discovery through spend is necessary but never claimed sufficient.
The operational version of this is continuous rather than quarterly: every bill and expense feed flowing into one allocated view, so a new AI vendor appearing anywhere in the estate shows up as a new line with an owner’s name on it, not as a surprise in the annual audit.
Who owns the shadow AI problem?
Nobody, and that’s the cause, not a symptom. Security owns the data risk, IT owns the tool inventory, finance owns the spend, and shadow AI lives precisely in the seams between those three mandates: a personal subscription is too small for security review, invisible to the SaaS inventory, and categorized as expense noise by finance.
The governance gap is structural rather than accidental: adoption reached majority scale through thousands of individual choices, while ownership of the problem stayed unassigned. The estate grew majority-scale before most org charts assigned it an owner.
The companies handling this well give the mandate a single home, usually a cross-functional AI governance group with finance holding the spend ledger, because every shadow AI pattern eventually touches money, and money is the one trail that crosses all three seams.
How do you govern shadow AI without killing adoption?
The failure numbers rule out the ban as a strategy, so governance has to make the approved path the easier path. Four moves, in order:
First, offer the capability. Most shadow AI tools are solving real workflow gaps; close the gap with sanctioned equivalents and the demand migrates on its own.
Second, make intake fast: a one-week AI tool review beats a three-month one, because the three-month version is what created the shadow estate.
Third, put every AI dollar in one place with owners attached, so adoption is visible the day it starts rather than the quarter it explodes; that’s allocation work, and it’s covered across our AI cost governance guide.
Fourth, review vendors on a standard rubric before they’re approved.
A written policy gives the four moves teeth, and the useful ones are short. Five elements cover it: which tool categories are pre-approved, which data classes never enter external AI, how to request a tool and how fast an answer arrives, who pays so personal cards exit the picture, and what happens on violation.
A policy employees can hold in their heads beats a 40-page document they’ll never open.
For the vendor-review move, here’s The AI Vendor Due-Diligence Checklist, the eight questions that catch what pricing pages hide: What unit does it bill in, and what behavior moves that unit? Where does our data go, and does the free tier differ? Is there an enterprise agreement, or are we on consumer terms? What happens to spend at 10x usage?
Then the portfolio four:
- Does it overlap a tool we already pay for?
- Who owns the budget line?
- How do we export our data out?
- And what’s the retirement story if the vendor reprices or folds a model?
Run those eight before approval and most of the shadow estate’s failure modes never ship.
How CloudZero makes shadow AI spend visible
That governance model has one operational dependency: seeing every AI dollar in one place. That’s the product. CloudZero pulls spend from every source through its integrations, cloud bills, AI providers (it was the first cloud cost platform to integrate directly with Anthropic), SaaS, and card or expense exports through the AnyCost pipeline.
From there, Dimensions allocates every AI line to a team and product, and hourly anomaly detection inside the AI Hub surfaces a new vendor or a doubling subscription as a flagged change in days, not a year-end confession.
Want to see the allocated view before talking to anyone? Take the self-guided tour and open the AI Hub first. And when you’re ready to put owners on every AI dollar, sanctioned or not, schedule a demo and ask to see spend discovery across connected sources.