Shadow AI vs Shadow AI IT: The Next Enterprise Risk

Unsanctioned use of AI in the enterprise is already a thing.  What’s coming next is even more insidious.

• Shadow AI occurs when employees use artificial intelligence tools without enterprise approval.
• The bigger emerging risk is Shadow AI IT—employees using AI to build unsanctioned software systems that interact directly with corporate applications.
• Without a data-centric architecture, this can create security risks, technical debt, and hidden system dependencies.

What is Shadow AI in the Enterprise?

Shadow AI is already occurring in virtually all firms.  But it’s occurring differently wherever it shows up. 

Basically, Shadow AI is the use of AI within an enterprise that isn’t sanctioned by the “powers that be.” A few traditional firms are encouraging and supporting the use of AI in their employees’ daily work with paid subscriptions and training.  But most are still very ad hoc. 

Shadow AI is when employees take the initiative and start using AI for the enterprise on their own.  This takes several forms:

• Dialing it in – Some employees use AI to get their work done in a fraction of the time and use the rest of their time for other pursuits (side gigs, other jobs, or leisure).
• Improving the quantity and quality of their deliverables – Many employees are using AI like steroids, to improve their performance.  They take the PowerPoint, Excel spreadsheet, or marketing plan that they would have built on their own, slowly and with many bugs, and instead deliver something much more robust in less time. 
• Showing the way – Some employees are using AI to show the company how it should deliver value in the future.  This is laudable but should be taken with a grain of salt.  These are usually well-intentioned point solutions, but, being on the fringe of the enterprise, they are unlikely to get much traction. 

Most enterprises are in a very weird place here.  For the most part, their employees are front-running them.  They (at least some of them) are ahead in the adoption of AI.  Companies and agencies are struggling to catch up and trying not to kill a goose that might lay a golden egg. They want to control this: There are huge risks, a few of which we’ll get to in a moment, and there is great opportunity for those who get it right, but there is no proven game plan. 

Most companies would like to seize the productivity gains of AI for themselves.  So would most employees.  So far, many companies are disemploying people and promising Wall Street that they will capture the productivity gains, but the jury is still out.

Meanwhile, I want to talk about a related problem that I know slightly more about: What happens when these newly empowered employees turn their new superpower toward the enterprise information systems?  I’m talking about Shadow AI IT.

What is Shadow AI IT?

We’ve had shadow IT for a long time.  Every time some line manager got frustrated with central IT, they would contract with their brother-in-law to build a system that solved a specific problem.  These “systems” took downloads from the firm’s legacy systems and built Excel- or Access-based “systems” to solve an unmet need.

This is “shadow IT,” partly because the central group may not be aware of its existence.  This creates a problem, because there is now an unknown and undocumented dependency between the corporate systems and these shadow systems. 

The Risks of AI-Generated Software

Now we’ve taken the brother-in-law out of the equation.  Any line manager with a Claude Code account (which is like half of them) can now gin up their own system out of whole cloth with no one’s permission. 

What could go wrong?  Well, lots.  Let’s start with the more obvious and move to the subtle.  

1. There are security issues aplenty.  Vibe code contains a host of vulnerabilities.  More code means a larger surface area/attack vector.  Plus, employees are inadvertently sharing trade secrets and PII and providing LLMs with open access to company repositories.
2. Most AI-generated systems are technically obsolescent.  Sounds odd that something you just built yesterday could be obsolete so soon, but consider this:  Most of the AI-generated code produced last year was made obsolete by this year’s versions, especially Opus 4.6 and GPT-5.3.  The quality and security of this year’s crop is so far superior, you would be crazy not to redo anything you had done last year. The same will be true next year. 
3. A more pernicious problem exists when the new code accesses corporate systems.  This new generation doesn’t even ask for a download (which might leave a trace).  The generated code just accesses the same user interface you would have used.  This is the basis of the so-called SaaSpocalypse, where the automation of SaaS systems led to a dramatic drop in the number of seats (licenses) needed. 

But whether your employees train their newly minted assistants on a SaaS product or your in-house systems, you’ve created the same problem. You now have an unknown number of “systems” dependent on your current software configuration.  It is essentially hardening the status quo in place. 

Why Data-Centric Architecture is the Long-Term Solution

There is a better path.  It focuses on the enduring themes.  For any given business there are a small number of conceptual entities that endure for decades.  Not because everyone is afraid to touch the legacy system that houses them, but because they are still relevant.  These include employees, products, customers, vendors, orders, workflows, projects, and addresses.  These core concepts will outlive most companies.  And it is possible to build your APIs on these concepts.

We call the approach that starts with these the “Data-Centric Approach.”  This is more than the Semantic Layer that has been mooted lately.  The Semantic Layer being offered by software vendors now is primarily surfacing the as-is datascape through a unifying lens.  It’s a noble endeavor, but it preserves rather than reduces complexity. 

The Data-Centric Approach focuses on the core enduring themes and maps the as-is data landscape to them.  This provides a true isolation layer.  Encouraging your AI developers to build on this layer ensures they are building on a solid foundation.

Mitigate Your Shadow AI IT Risks

The real danger is shadow AI IT, where employees use AI to create unsanctioned systems that interact directly with enterprise software.

In the current state, organizations cannot stop this trend. The best way to buck this trend is to lay a stable architectural foundation that eliminates the need for this hidden activity. A data-centric architecture provides that foundation.

Curious how it works? Learn more about the data-centric approach to enterprise data transformation.