Implementing AI is the Weakest Link – Shadow AI creates a Hidden Problem of Technical Debt
Employee-built AI can deliver quick wins. Without ownership, documentation and ongoing support, those useful experiments can quietly become technical debt.
An employee builds an AI assistant that saves time; creates an agent to prepare reports; connects a chatbot to a spreadsheet; uses AI to join up their go to market sales/prospect outreach, calendars, meeting bookings, email call follow ups. For the tech savvy, the list goes on and it’s getting easier. Each improvement is useful, saves them time and makes them more productive. What’s not to like?
Such employee initiative is invaluable. When it improves a day to day, practical workflow that people recognise as having an immediate positive impact on their ‘day job’, it is often the vanguard to wider adoption (see Article 1 – Theory to Practice… ‘most organisations struggle here’). But when it is a single virtuoso performance, a personal bespoke solution – not a repeatable orchestra play, when it doesn’t ‘connect to a larger ambition’, it will become a Shadow AI problem.
It creates a collection of business dependencies that nobody has designed, managed, built to integrate more widely, understands how to keep current or has a plan to maintain. That’s not mention governance, observability, audit and regulatory compliance. The problem starts when an experiment becomes essential to daily work without acquiring an owner or documentation.
It’s called technical debt - the future effort created by shortcuts and unplanned design decisions. Routine maintenance is ad hoc - debt accumulates when decisions that make something quick to build make it harder to understand, change or sustain.
Designing and Engineering as One, is the Beginning of the Commitment
Researchers described hidden technical debt in machine learning systems as far back as 2015, highlighting risks such as data dependencies, tightly connected components and changes in the external environment.[1] Today's agents introduce their own combinations of models, instructions, tools and business connections that require ongoing attention.
Consider an agent that prepares an operational report from several sources. A change to a data field may affect the result. A permission may expire. A replacement model may interpret an instruction differently. The process needs someone to notice the problem, understand the dependency and check the correction.
If employees build separate solutions to similar problems, the organisation may also inherit duplicated connections and inconsistent definitions. One department's version of a customer or an overdue order may differ from another's. Individual productivity can improve while coordination becomes harder.
This is where governance bites! Leaders need to know which tools are being used, what information they can access and what actions they can take. Without that knowledge, it is difficult to apply the organisation's policies or judge whether the controls are appropriate.
Test what Happens when the Creator Leaves
A ‘shadow’ agent can depend on its creator's account, personal connections and undocumented knowledge. When that person changes role or leaves, the agent may still exist while nobody can reliably operate or maintain it.
Ownership transfer is available in some platforms: Microsoft, for example, documents reassignment of orphaned Copilot Studio agents.[2] The wider question is whether the business has a workable handover for the process, including its access, dependencies and evaluation.
Karrik, the agentic platform built by Professor Andy Pardoe and his team, addresses the need to retain agents within the business when their creators move teams or leave. That capability supports continuity; organisations still need to assign responsibility and ensure an authorised colleague can use and maintain the work.
Building a bespoke solution can be the right choice when it delivers operational task value – reducing time, effort and cost. If it works for one, can it be made to work for all?. That decision should include the ongoing commitment: support, testing, security, documentation and adaptation as connected systems change. Buying a platform also creates responsibilities, including configuration, supplier oversight and integration.
A practical first step is to map the AI tools that people already rely on. Establish which are experiments and which support important operations. For each important workflow, name an owner and test whether another authorised person could run it tomorrow.
Give employees a clear route for turning useful experiments into supported business capabilities. Informed.AI can help assess those dependencies and build-or-buy choices; Karrik provides a platform for retaining the agents your people create.
Review your AI operating approach at informed.ai, or explore Karrik at karrik.com.
Sources
[1] D. Sculley and colleagues. [Hidden Technical Debt in Machine Learning Systems](https://proceedings.neurips.cc/paper_files/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html). NIPS 2015.
[2] Microsoft Learn. [Reassign ownership of orphaned agents with the Power Platform API](https://learn.microsoft.com/en-us/microsoft-copilot-studio/admin-api-reassign-ownership-orphaned-agent). Updated 11 September 2026.
About the Author
Professor Andy Pardoe is a Professor of AI and founder of Informed.AI. He brings more than 30 years of technology, business and management experience to his work with CEOs and boards on AI strategy, governance and adoption.