Shadow AI: How to Find the AI Systems Nobody Put on a List

Published: July 2026

Most organizations that start an AI inventory make the same discovery: the official list is short, and the real list is not. The AI that was formally approved — the chatbot project, the procurement-reviewed vendor — is the easy part. The risk sits in everything else: the feature that arrived in a SaaS update, the pilot that quietly became production, the backlog ticket someone shipped eighteen months ago.

That unlisted layer is shadow AI. Here is how it gets in, why it carries more risk than the AI you approved, and a practical way to find it using data you already have.

How AI Gets In Without Anyone Deciding

Almost no shadow AI arrives through deliberate evasion. It arrives through three ordinary routes.

• Vendor features. SaaS products now ship AI by default. Atlassian, for example, shipped Rovo agents into Jira in 2026 — agents that can be assigned work items, mentioned in comments, and embedded in workflows. If your teams use Jira, AI agents may already be participating in your processes without any procurement review mentioning them. Gartner projects that roughly 40% of enterprise applications will include embedded AI agents by the end of 2026, up from under 5% in 2025.
• Pilots that became load-bearing. A hypothetical: a support team trials an LLM to draft ticket replies "for two weeks." Eight months later it handles first-pass responses for an entire queue. Nobody re-evaluated it, and it appears in no inventory because it was never a project — it was an experiment that never ended.
• Backlog items. A developer picks up a ticket titled "auto-categorize incoming requests," wires in a model API, and closes the ticket. The AI is now in production. The only record of its existence is a resolved issue in your work tracker.

Why Shadow AI Is the Riskiest AI You Run

Approved AI has an owner, a purpose statement, and usually some review. Shadow AI has none of that — and it faces the same laws.

The EU AI Act's general-purpose AI obligations have applied since August 2025, and after the 2026 simplification package, high-risk obligations take effect on 2 December 2027 for standalone systems and 2 August 2028 for AI embedded in products. Colorado's AI Act — with its duty of care around algorithmic discrimination, risk-management programs, and impact assessments — took effect on 30 June 2026. None of these laws distinguish between AI you approved and AI you never noticed.

Shadow AI is also operationally fragile. Gartner has warned that over 40% of agentic AI projects may be cancelled by 2027, citing cost, unclear value, and weak risk controls — and ungoverned systems are exactly the ones with the weakest controls. An unowned system has nobody watching for drift, nobody validating outputs, and nobody who can answer the first question a board or regulator will ask: what does this system do, and who is accountable for it?

The Detection Method: Mine Your Work-Tracking Data

You do not need network monitoring or an agent on every laptop to find shadow AI. You need the one dataset that records nearly every build decision your organization has made: your work tracker.

Every AI system somebody built or integrated left a trail in Jira, Azure DevOps, or whatever tool your teams plan in. Tickets to integrate a model API. Epics named "smart triage." Components called ml-service. A story that says "add recommendations to the dashboard." The AI itself may be invisible; the work to create it is not.

The method is simple: export the data (Jira and Azure DevOps both export to CSV) and scan it for AI signals. This is the discovery pass AI Governance Hub runs — upload a Jira, Azure DevOps, or CSV export, or install the Atlassian Forge app for Jira Cloud, which runs entirely inside your own Atlassian tenant with no data egress. If you want to see what the input looks like first, there are sample CSV files you can inspect.

What the Signals Look Like

Whether you scan manually or with tooling, look for four classes of signal:

• Model and vendor terms: LLM, GPT, copilot, embedding, fine-tune, inference, prompt — plus the names of any AI vendors your teams may have integrated.
• Capability verbs: summarize, classify, predict, score, recommend, auto-triage, detect, generate. These often flag AI features that were never described as AI.
• Structural signals: components, labels, and epics — ml-pipeline, recommendation-engine, or an ai-experiment label that stopped being an experiment.
• Integration work: tickets about API keys, rate limits, or usage costs for a model provider are strong evidence a system went live.

A hypothetical to calibrate expectations: a mid-size product organization scans three years of backlog and finds forty-odd AI-flagged items that collapse into nine distinct systems — two known, seven not. That ratio, more unknown than known, is the argument for doing discovery before writing your governance policy, not after.

Turning Findings Into a Governed Register

A pile of flagged tickets is not an inventory. To become governable, findings need four things: deduplication into distinct systems, a named owner for each, a risk score, and a mapping to whichever frameworks apply — the EU AI Act, ISO/IEC 42001 (the AI management-system standard published in December 2023), or the NIST AI RMF, whose Map function is precisely this exercise: establishing context and cataloguing what exists.

You can build this by hand — a free AI risk register template is a reasonable starting structure — or generate it from your export. AI Governance Hub turns the discovery pass into a risk-scored register, maps each system against EU AI Act, ISO/IEC 42001, and NIST AI RMF criteria, computes a governance score, and produces board-ready reports in PDF, Word, PowerPoint, or HTML. To be clear about what that is: governance evidence and executive analysis to support your compliance work — not legal advice, and not a certification of compliance.

The sequencing matters. With the EU high-risk dates now set for December 2027 and August 2028, there is a window to do this calmly — and the register you build now is the input everything else, from impact assessments to board reporting, depends on.

Frequently asked

What is shadow AI? AI systems in use inside an organization without formal approval, ownership, or governance — usually vendor-shipped features, pilots that never ended, or quietly shipped backlog items rather than deliberate rule-breaking.

Can I find shadow AI without installing monitoring software? Yes. Work-tracking exports from Jira or Azure DevOps record the tickets, epics, and components behind most AI systems, so scanning that data gives a fast first-pass inventory with no new infrastructure.

How often should AI discovery be repeated? Treat it as a recurring control rather than a one-off project — quarterly is a sensible starting cadence, because vendor updates and new backlog work introduce AI continuously.