Wiki · Concept · Last reviewed June 16, 2026

Shadow AI

Shadow AI is the use of AI systems, models, agents, plugins, or personal subscriptions for organizational work outside the institution's approved governance, security, privacy, procurement, and records processes.

Definition

Shadow AI is a workplace and organizational governance pattern: people use AI tools to do institutional work without those tools being approved, inventoried, secured, logged, contracted, or evaluated by the organization. It includes personal chatbot accounts, unapproved coding assistants, browser extensions, meeting bots, spreadsheet add-ons, third-party summarizers, unofficial API keys, local open-weight experiments, and agent workflows built before procurement or security review.

The term extends the older idea of shadow IT. The AI version is more than an unauthorized software problem because the tool may transform confidential data, generate advice, alter records, write code, summarize meetings, shape personnel decisions, or act through connected systems. A user may think they are only drafting text. The organization may actually be creating a vendor disclosure, a new record, a hidden automation, or an unreviewed decision aid.

Shadow AI overlaps with AI System Inventory, AI Governance, Data Minimization, and AI in Employment. Its core question is simple: where is AI already being used before the institution admits that it is being used?

How It Works

Shadow AI usually starts from practical pressure. Workers face inbox volume, reporting demands, coding tasks, sales copy, translation, meeting notes, policy summaries, customer responses, and performance targets. If official tools are absent, slow, or less capable than public tools, staff may paste material into consumer systems, connect personal accounts to workplace files, or build small automations that spread through a team.

Managers can create shadow AI indirectly. If productivity goals assume AI-like output but the organization provides no approved tool, training, data rules, or disclosure path, employees may hide their use. The result is not only a compliance gap. It is a broken feedback loop: the institution cannot learn which tasks need approved tools, which data is flowing to vendors, which outputs are unreliable, or which workers are quietly absorbing automation risk.

Current Context

Microsoft and LinkedIn's 2024 Work Trend Index reported that 75 percent of surveyed knowledge workers used generative AI at work and that 78 percent of AI users were bringing their own AI tools to work. The report framed this as Bring Your Own AI and warned that it could put company data at risk.

Public-sector rules also point toward the same issue. OMB Memorandum M-25-21, which rescinded and replaced M-24-10, tells U.S. federal agencies to set clear expectations for appropriate AI use, maintain annual AI use case inventories, identify Chief AI Officers, and apply minimum practices for high-impact AI. NIST's AI Risk Management Framework says organizations should have mechanisms to inventory AI systems, document roles, train personnel, and manage third-party AI risks.

Cybersecurity guidance treats unmanaged AI as an operational risk. The 2024 joint guidance Deploying AI Systems Securely recommends robust deployment governance, clear roles, security boundaries, threat models, and protection of proprietary data sources. The 2025 joint guidance on AI data security focuses on protecting data used to train and operate AI systems, including proprietary, sensitive, and mission-critical data. The EU AI Act's Article 4 adds a literacy duty: providers and deployers must take measures, to their best extent, to ensure sufficient AI literacy for staff and others using AI systems on their behalf.

Governance and Safety

Shadow AI should be treated first as discovery evidence, not only misconduct. It reveals where official processes are too slow, where employees need better tools, and where organizational work has already changed. A punitive-only response can drive use further underground and produce worse data leakage, worse records, and worse safety evidence.

The risks are concrete: confidential or personal data may leave approved systems; prompts and files may be retained under consumer terms; outputs may be copied into official records without review; code may be accepted without provenance; workers may rely on hallucinated or biased advice; and managers may evaluate labor through undisclosed automation. For agents, shadow AI can also create unapproved identities, credentials, tool permissions, and audit gaps.

Good governance sets boundaries that workers can actually follow. It names approved tools, prohibited data, allowed use cases, disclosure rules, review requirements, procurement paths, recordkeeping duties, and escalation channels. It also gives employees a way to request new AI support without having to hide the experiment.

Defense Pattern

Spiralist Reading

Shadow AI is the workplace making a private bargain with the machine before the institution has language for the bargain.

It is not a sign that workers are foolish. It is a sign that work has become porous: public models, private stress, corporate data, vendor clouds, and performance metrics all meet in the text box. The question is not whether to shame the user. The question is who made unofficial automation feel necessary, useful, or safer to hide than to name.

Open Questions

Sources

Return to Wiki