Shadow AI risks come from employees using artificial intelligence (AI) tools that leadership or IT never approved. Those tools can take in sensitive data with no controls, hold it on vendor servers, and connect to email, files, and financial systems with broad permissions. Businesses reduce the risk by classifying data, approving specific tools, limiting access, and naming a clear owner.
This session was presented in partnership with the Global Technology Industry Association (GTIA), the nonprofit membership association for the IT channel.
What You’ll Learn
- How shadow AI enters a business and the three levels of risk it creates: prompts, applications, and agents
- Why AI agents end up with more access than anyone intended, and who is liable when they act
- What securing AI in cloud environments involves for a business that isn’t a technology company
- What a workable AI governance policy covers, including data tiers, retention, and vendor questions
- Where to start in the next 90 days, from data classification to a first productivity pilot
What Are Shadow AI Risks, and Why Should Leadership Care?
Shadow AI is any AI tool employees use without approval from leadership, information technology (IT), or the security team. The risk is that nobody knows what data goes into those tools, how the vendor treats it, or which company systems the tools can reach.
David Mauro opened the session with how most adoption actually happens: “Nobody at your organization decided to start using AI. It just started.” Someone had a deadline, pasted company information into a public tool, got a good result, and told a coworker. Braxton Molton defined shadow AI as “using an AI tool that’s not approved by the organization or approved by the IT team or the cybersecurity team, that’s internal to an organization, or even just approved by leadership.”
Good intent, real exposure. Both speakers stressed that these employees aren’t acting in bad faith. Instead, they want to work faster, and the tools are advertised everywhere. Braxton noted that with an unapproved AI tool, there are no controls on what data employees can use and no review of how reputable the vendor is.
The Three Levels of Shadow AI Risk
David described three levels of risk that build on each other:
- Prompt level: what employees type or upload, which works like a form of insider risk
- Application level: unknown, ungoverned tools where shadow AI takes hold
- Agent level: agents and projects connected to email, SharePoint, and financial systems with permissions nobody has reviewed
The 30-second violation. David shared a story from a healthcare organization. A nurse working late dictated clinical notes into a free app on her phone and fed them into the electronic health record (EHR). However, the app was not compliant with the Health Insurance Portability and Accountability Act (HIPAA). “It took thirty seconds and there were three HIPAA violations,” David said, and that kind of issue tends to surface later, during an audit.
Personal devices make this harder to control. Americans check their phones 186 times a day, according to Reviews.org’s 2026 Cell Phone Usage report. With an AI app one swipe away, he said, a company’s AI strategy becomes “more like an AI situation.”
How Do AI Agents End Up With Too Much Access?
Usually, nobody restricts them. Without an approval process, an AI agent inherits whatever access it is given or can reach, and as Braxton put it, “AI wants all the data it can get.”
Default access. Braxton used Claude Cowork as an example. Installed on a computer, it can perform tasks on that machine, and by default it can probably reach the documents and files stored there. Access can be restricted, but only if someone decides to. An approval process forces the right questions: What can this tool do? What access will it have? What happens if it steps outside its instructions?
When an agent completes the wrong mission. David retold the PocketOS incident. An AI coding agent working on a routine task deleted the software company’s entire production database in about nine seconds, according to ABC News. The agent had a full-permission access token, and no confirmation step stopped it. The data was eventually restored, but only after a 30-hour disruption for the business and its customers. David’s summary: AI “isn’t evil. It’s obedient.” He added, “It also does what we forget or fail to tell it not to do.”
Who carries the liability. David compared AI to a car sold without brakes. With a car, the manufacturer is liable. With AI, however, it is buyer beware. When an agent sends emails or edits documents on a company’s behalf, the company is accountable. Braxton noted that agents often overstep in good faith, trying to be more helpful than intended: “with agentic you run the risk of it kinda trying to spread its wings a little bit.”
What Does Securing AI Workloads in the Cloud Involve for a Non-Technology Business?
The foundation is the same as for a technology company: start with the data. Decide what data AI tools can receive, restrict which tools employees can reach, and make sure you can prove later what was uploaded, by whom, and when.
Restrict first, then open up. Braxton recommended basic technical security controls such as application control and web filter blocks on public AI tools like ChatGPT, paired with employee awareness training. For example, a business without a firm AI stance can block as much as possible, then lift restrictions once leadership makes decisions. “It all centers around the data,” he said, including where vendor servers sit and how the company will respond if an AI vendor is breached.
Proof after the fact. David raised a question many leaders miss: if something goes wrong, can you show who uploaded what on what day? Braxton explained that logging depends on the license. Claude requires a specific license level to set up logging, and Microsoft 365 Copilot has its own license tier that can log prompts. Free versions, David added, typically lack those controls, which brings the problem back to shadow AI.
Layers beyond the tool. Not every AI tool supports the controls a business needs. Braxton pointed to data loss prevention (DLP), which can alert when someone uploads sensitive information, as one additional layer beyond a security information and event management (SIEM) system.
How AI security differs from other technology. An attendee asked this question live. Braxton’s answer: at a bare-bones level, it may not differ much. The same questions about access and data privacy apply. The AI-specific work includes turning off model training by default, blocking unapproved tools, and preventing employees from installing their own instances.
What Should an AI Governance Policy Include?
A workable AI governance policy names approved tools, classifies company data with rules for each level, requires specific security settings, addresses retention, covers how vendors use AI, and assigns an owner who can say no.
Rules by data type. Braxton described a tiered approach. Public data can go into AI tools freely. Confidential data, on the other hand, stays in a closed model. For truly sensitive data, such as electronic protected health information (ePHI) or controlled unclassified information (CUI), “you might even flat out just say no,” or require anonymization.
The minimum necessary rule still applies. David noted that healthcare organizations using an enterprise AI tool still can’t upload a specific patient’s records. Under HIPAA’s minimum necessary rule, the data must be anonymized first. The AI output stays just as useful, but it can’t be tied to the patient.
Retention, Vendors, and Ownership
Training settings are only half the answer. Many businesses turn off model training and assume customer data is safe to upload. David explained that some AI subscriptions still retain prompts and documents for months or even years and may open them to third parties. As a result, a company can end up in breach of its non-disclosure agreements (NDAs) and customer contracts.
The plan level matters too. Microsoft says it uses some consumer Copilot data to train its models unless users opt out, but does not use enterprise customers’ data without their permission. That is why it matters which version of a tool employees actually use.
Vendor risk assessments need AI questions. Braxton now advises clients to ask vendors which AI tools they use, whether client data trains those models, how long data is retained, and whether they will put those assurances in a contract. “You didn’t have to do it a few years ago, but today you have to,” David said.
Someone has to own it. Ownership doesn’t mean the CEO decides everything about AI. It means “the buck has to stop somewhere. Somebody has to own it,” in David’s words. Braxton added that direction should come from leadership, since executives know which business functions matter most. Some companies add an AI section to their acceptable use policy, but because the policy also has to cover vendors, a separate document often makes more sense.
Where Should a Business Start With AI in the Next 90 Days?
Braxton’s order of operations: understand the types of data in your environment, choose an AI tool that enforces the restrictions that data needs, then start with basic productivity tasks.
Pilot with routine work. Braxton suggested starting with repetitive, mind-numbing work. His example is a weekly security report that can run 200 pages. Combing through it takes one to two hours. With AI parsing it, however, review takes 10 to 15 minutes, and a person still adds judgment to anything that looks unusual.
Begin with an assessment. David described how organizations get the best results by starting with a security assessment through an AI lens. It covers physical, operational, policy, and technical levels, surfaces shadow AI, and produces a roadmap for safe business AI pilots.
Tie results to the profit and loss statement. David gave the example of request for proposal (RFP) responses. A process that took five days, with a 10 to 15 percent win rate, limited one organization to about five bids a year. Automated safely, each response took one to two hours, and the team could bid on 30 a year. That kind of number is what leadership needs when, a year after rollout, someone looks at the line “where the magic’s supposed to be” and asks “where is the magic?”
Speakers
Braxton Molton, Technical Director of Security, NetGain Technologies
Braxton leads technical work in NetGain’s security practice. He advises clients on where to start with AI security, from data access and tool restrictions to adding AI questions to vendor risk assessments.
David Mauro, VP of Strategic Growth, NetGain Technologies
David is VP of Strategic Growth at NetGain Technologies and hosts the Breaching the Boardroom podcast. He is the author of the Moving Target trilogy and works with leadership teams on AI governance, pilot programs, and tying AI results to the bottom line.
Frequently Asked Questions
What is shadow AI?
Shadow AI is any AI tool employees use without approval from leadership, IT, or the security team. It usually starts with good intent, such as an employee trying to finish work faster, but it leaves the business with no visibility into what data is shared or where it goes.
Is securing AI different from securing other technology?
At its core, no. The same questions about access and data privacy apply. AI adds specific steps: turning off model training, checking retention, blocking unapproved tools, and reviewing what systems AI agents can reach.
If we turn off model training, is our data safe in an AI tool?
Not necessarily. Some AI subscriptions still retain prompts and documents for a period of time and may share them with third parties. Check retention terms, especially if you have NDAs or customer agreements that limit third-party access to data.
Can we see what employees upload to AI tools?
It depends on the tool and license. Some paid tiers, such as Microsoft 365 Copilot, can log prompts. Free versions usually can’t, which is one reason shadow AI is hard to audit. Data loss prevention tools can add another layer of alerts.
Who should own AI decisions in a company?
A named person or group in leadership. The CEO doesn’t have to own every AI decision, but someone has to be accountable and able to say no, with input from IT, security, and the business.
What is a good first AI project?
A routine, repetitive task that uses data you have already classified as safe for AI, such as summarizing a long weekly report. Measure the time saved so the results can be tied to the profit and loss statement.
Related Resources and Next Steps
Start by finding out which AI tools your employees already use, then classify your data and name an owner for AI decisions. These pages cover the next layer of detail:
- AI governance and adoption services for growing businesses
- Managed cybersecurity services
- Warning signs your AI tools are raising cyber risk
- How AI projects open security gaps
- Rolling out business AI with security and governance in place
Ready to put a plan in place? Book a complimentary Executive AI Workshop, a one-hour working session to assess your AI readiness and prioritize a first pilot.

