Once it was shadow IT, and now it’s shadow AI, a topic that will get introverted IT and security leaders speaking their minds. The other topic is technical debt, with data debt, AI debt, and AI cost debt as the mic-droppers.
I’ve always viewed shadow IT as an opportunity and wrote seven steps for turning shadow IT into a competitive edge. In another article, I advised citizen developers on how to respond to shadow IT allegations.
Shadow AI can point to opportunities, but…
I recently hosted a discussion on shadow AI at Coffee With Digital Trailblazers. I was hoping the conversation would lean toward finding hidden innovations.
Joe Puglisi, growth strategist and fractional CIO at 10xnewco, shared similar sentiments, saying, “There’s innocent use of [shadow] AI, and that’s derived from people who are curious to see what they can do with it. And to some degree, I think that should be encouraged.”
Joe encourages Digital Trailblazers to run the department of know, rather than the office saying, “no,” which discourages experimentation.
Other leaders have similar sentiments. “Punishing people for using AI just pushes the behavior further underground,” says Eric Johnson, CIO at PagerDuty. “The better move is treating it as a signal: understand what employees need, then give them a sanctioned way to get it without putting company data at risk.”
Cássio Christianini, head of AI go-to-market, Americas at SAP, adds, “Treat shadow usage as a roadmap by identifying the highest-value use cases and moving them onto governed tools grounded in company data, so governance becomes the path to scale rather than a brake.”
5 very critical risks of shadow AI
All that said, shadow AI poses significantly more risks than shadow IT.
Derrick Butts, enterprise AI cyber-resilience and risk advisor and founder of Continuums Strategy, said during the shadow AI episode, “We now have the point where data, intellectual property, other things that may be confidential are now being seeped into the ether without the business knowing what’s taking place or how much is being put out there. And once you put it out there, you can’t get it back.”
Several experts shared with me some of the risks of shadow AI and what to do about them.
1. Confidential data pasted into public AI tools
Your company procured one LLM, and likely more than one. But did IT configure monitoring tools to detect employees using other tools? If not, people are likely using their preferred tools, and they may be exposing confidential data.
“A shadow AI risk I see often is an employee, often innocently, inputting confidential data into a public AI tool,” says Andy Boettcher, chief innovation officer (CIO) at DoubleTrack. Often, employees do this unknowingly, not realizing they risk exposing the data in a breach or, in even worse cases, having the data published online. “
Recommendation: Boettcher’s suggests giving employees a secure and sanctioned tool, and then following up by
- Developing clear rules for what data is considered confidential.
- Providing clear instructions on how to use the tool with confidential data.
- Creating a frictionless process that lets employees ask IT for permission when a use case falls outside sanctioned AI use.
2. Over-privileged AI agents running on borrowed access
AI runs in many places. The tool may be sanctioned, but how people use it, and the default entitlements across many access points, may be a risk that looks like shadow AI.
“AI now lives in SaaS apps, internally built AI apps, copilots and browser extensions, and in the agent skills and MCP servers employees download or configure on their own,” says Anusha Vaidyanathan, senior director, product management for AI security, security services, and technology integrations at Versa Networks. “The most immediate risk is access, because no one can say whether a user or an AI agent is doing exactly what it’s authorized to do and nothing more, or whether it has drifted into privilege escalation, unintentionally or maliciously.”
Chris Lentricchia, director of cloud and AI security strategy at Sweet Security, says a second concern is that the risk extends beyond data leakage. “The agent can turn a legitimate identity into a persistent path for privilege abuse, untraceable changes, and bad decisions at machine speed.”
Recommendation: Lentricchia suggests, “Address it by discovering agent activity at the identity and API layers, then enforcing least privilege, scoped credentials, action-level logging, and a kill switch before the workflow reaches production.”
3. AI spending limits that push work to personal accounts
What happens when employees run out of tokens in sanctioned AI tools? Employees under pressure to get work done may shift their work to other, unsanctioned AI tools.
“AI employee spending limits are emerging as a solution for rising AI costs but have the potential to fuel shadow AI,” says Conal Gallagher, CIO and CISO at Flexera. “When employees hit their AI usage limits, what’s stopping them from moving their work to personal accounts or unapproved tools?”
Recommendation: Gallagher has two suggestions for CIOs:
- CIOs need visibility into what AI tools are being used, by whom, and at what cost, and then put governance around that usage.
- Organizations can set policies and spending controls based on business need and risk rather than relying on blanket limits.
4. Workflows that no one knows about
Another issue arises when IT and security’s AI governance tracks only sanctioned tools. When an employee uses an unsanctioned AI tool to do work and then leaves the company, no one has knowledge or access to pick up the responsibilities.
“Shadow AI presents real threats to organizations when in use, but can also create trouble when user employees depart for new jobs,” says Matt Wallace, co-founder & CTO at Kamiwaza AI. “Those organizations lose the provenance and context for the outputs the shadow AI had produced on the job.”
Recommendation: Handoffs and knowledge transfer around a departing employee’s AI agents is challenging enough, even when using sanctioned tools. Monitoring for shadow AI only exposes the problem. CIOs will need the CHRO’s partnership so that persistently breaking AI governance policies has consequences.
5. Changing systems without knowing the dependencies
Organizations face greater risks as they deploy more AI agents.
“An agent might have access to one system without any visibility into what depends on it, so a change that looks harmless can have consequences somewhere else like an outage, a security issue, or a compliance problem,” says Shannon Weyrick, co-founder and CTO at NetBox Labs. “Organizations need to understand those dependencies and set clear boundaries around what AI can see, change, and trigger before giving agents greater autonomy.”
Recommendations: Dependencies become a bigger issue as organizations scale from dozens to thousands of AI agents. It’s a topic I’ve been writing about:
- Ensure the AI strategy clearly defines business value to reduce the risk of employees pursuing misaligned shadow AI experiments.
- Document C-level responsibilities because multiple leaders share accountability for shadow AI.
- Define who’s accountable, what AI agent accountability means, and how responsibilities are audited before deployment.
- Establish a citizen developer checklist for building, testing, deploying, and monitoring AI-generated apps.
- Deploy an AI orchestration platform to manage A2A and MCP connections that comply with requirements and standards.
- Govern the AI context layer so AI agents have a clear set of entitlements from the AI brain the organization develops.
- Define and enforce AI agent observability standards so DevOps teams have consistent telemetry for troubleshooting and incident management.
- Mature key agenticops practices beyond observability include developing agent tracking models for drift and capturing employee feedback.
- Budget, measure, and manage spending and establish FinAI – FinOps for AI – to fund the most promising AI use cases and experiments.
- Encourage critical thinking with your teams, so AI does more than reshape the business; it drives the next wave of AI reinvention.
Leave a Reply