Shadow AI is the shadow IT problem again, according to two Airia leaders. In a Data Science Connect webinar, they explained why blanket bans fail, how to detect unsanctioned AI at the network, endpoint and application layers, and how to win employees over to approved tools.
Key takeaways
- Shadow AI is shadow IT again. Employees work around controls to be more productive, usually without malicious intent.
- Bans don't hold. A Fortune 50 bank's blanket ban and single approved tool wasn't seeing much success, Parwani said.
- Detect in layers. Combine network filters, endpoint browser extensions or agents, and application integrations such as Office 365 or Google Workspace.
- Ask employees. Surveys reveal where people see value in AI and set the bar a sanctioned tool must meet.
- Champions beat mandates. Power users inside business functions, even legal, drive adoption of approved AI tools.
Employees adopting AI tools faster than security teams can approve them is not a new problem, two Airia leaders told a Data Science Connect webinar audience. It is the same one companies faced with shadow IT. In "Out of the Shadows: A Step-by-Step Approach to AI Governance," a webinar sponsored by Airia and hosted by DSC co-founder and president Brian Mink, Dave Horton, VP of Solutions & Engineering at Airia, and Rahul Parwani, Head of Product, Security Solutions, laid out how enterprises can find unsanctioned AI use and replace it with tools people actually want.
"Shadow AI is nothing different from shadow IT," Horton said. "It's exactly the same problem that we're trying to solve."
Where is shadow AI showing up in the enterprise?
Parwani said the CISOs he talks to raise three themes: new attack vectors from LLMs and MCP servers, such as prompt injection, that traditional security tools can't address; AI sprawl and shadow AI; and the resiliency and ongoing evaluation of agentic systems.
Much of the sprawl comes from software companies already use, he said. Business applications that are approved and in use now let employees build agents on top of them, often without security seeing it. He described an HR team that enabled agents with access to all employee information, and a sales leader with admin rights building agents with access to the whole CRM.
The simpler cases are even more common, according to Parwani: employees uploading sensitive material to free tools whose terms allow training on user data, connecting a personal ChatGPT subscription to OneDrive, or using tools whose models are hosted outside their region, which can bring regulatory fines.
Why don't blanket AI bans work?
Parwani described a Fortune 50 bank that banned AI use and whitelisted a single tool but wasn't seeing much success. Employees see AI as a productivity multiplier and work around controls, usually without malicious intent, he said. Another company blocked AI on corporate machines and required full-device VPN, but nothing stops an employee from photographing a sensitive document with a phone and asking a chatbot to summarize it.
Horton compared the moment to the early smartphone era, when employees forwarded work email to personal iPhones because corporate policy lagged behind. Today the pull is stronger, Parwani added, because consumer AI tools often offer a better user experience than approved enterprise software.
How can companies detect shadow AI?
"There isn't a magic strategy to detecting shadow AI," Parwani said. Browser extensions alone aren't enough. He recommended working in layers:
- Network layer: existing network tools, which he said likely already include filters for known AI traffic
- Endpoint layer: browser extensions or agents on managed devices
- Application layer: integrations with Office 365 or Google Workspace to see which AI apps are connected to company data
Horton added the lowest-tech method: ask employees what they use AI for. If they don't feel threatened, they will say, and that sets the benchmark a sanctioned tool has to meet. "I also think it's a signal of where employees see value in AI," he said. Parwani said a Fortune 10 company ran such a survey and immediately picked use cases to build in a sanctioned way.
How should enterprises classify AI risk?
Horton outlined three tiers. High-risk uses touch customer data, production code or strategic decisions. Medium-risk uses carry concerns such as intellectual property leakage. Even low-risk tools like grammar checkers and calendar assistants need to see everything a user types or who they meet with. "The key is not to panic but to get ahead of it with some good processes," he said.
Parwani pointed to a recently disclosed case in which a calendar invite carrying an indirect prompt injection got past a copilot summarizing a user's day and sent sensitive data to a third party. Horton noted that frameworks including the EU AI Act and Colorado's AI law already require organizations to classify AI uses by risk.
How do you get employees to use sanctioned AI tools?
Top-down mandates rarely work, Horton said. Momentum comes from AI champions inside the business, power users who bridge technology and their function. At one multinational, the first champion came from the legal department, not IT. He recommended lunch and learns, hackathons, regular feedback loops, a gradual migration from shadow tools, and celebrating wins when an agent is widely adopted. The goal, he said, is tools that feel easier, safer and more rewarding than the unapproved ones.
Parwani said most organizations he works with have a cross-functional AI governance committee of legal, security and business stakeholders that approves models, MCP servers and starting use cases. Airia's platform lets that committee approve components and then supports no-code and low-code agent building, he said.
Asked whether Microsoft Copilot is safer, Parwani said security teams often start there because they trust the vendor, but organizations outgrow it quickly. "I do think it's a misnomer that those are more secure," he said.
His closing advice: don't launch a 12-month data readiness effort before starting, since many use cases don't need it. "Find the problem you're trying to solve," he said, then build an inventory of AI use and go after the quick wins.
Follow AI governance discussions on DSC webinars
Data Science Connect has 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the full lineup on the webinars page.
Frequently asked questions
What is shadow AI?
The use of AI tools, or agents built on approved business software, outside the oversight of IT and security teams. Airia's Dave Horton called it the same problem as shadow IT.
How do you detect shadow AI?
Rahul Parwani said no single method works; companies need network monitoring, endpoint browser extensions or agents, and application-level integrations, plus simply asking employees what they use.
Is Microsoft Copilot safer than other AI tools?
Parwani said security teams often start with Copilot because they trust the vendor, but he called it a misnomer that such copilots are more secure.
Why do AI bans fail?
Employees see AI as a productivity multiplier and find workarounds, such as photographing documents with a personal phone, so bans push usage out of sight instead of stopping it.
Watch: Out of the Shadows: A Step-by-Step Approach to AI Governance on YouTube
Event: Webinars · More coverage: Webinars insights
