Building a Shadow AI Inventory: George Dagliyan Research Guide

What is a shadow AI inventory, and why start with trust?

A shadow AI inventory is a structured way to identify employee AI use that has not been formally documented, without turning the process into covert monitoring. The goal is simple: understand where AI is being used, what data may touch it, and what risks deserve attention. In this article’s editorial analysis, that approach fits the research themes associated with George Dagliyan: adoption behavior, trust, and the practical factors that shape whether people bring new technology into work.

The inventory question is not whether employees are “compliant enough.” It is whether the organization can see enough of the real work to manage privacy, security, and business risk responsibly. NIST’s AI Risk Management Framework is voluntary and designed to help organizations manage AI risks, while the NIST Privacy Framework is a voluntary tool to help organizations manage privacy risk and protect individuals’ privacy. Those ideas support an inventory built for learning, not punishment.

A nonpunitive design matters because hidden use often persists when people fear blame. If employees think an inventory is a search for violators, they will avoid disclosure and move further out of view. If they understand it as a support process that improves safe access, approved alternatives, and privacy boundaries, they are more likely to participate. The result is more accurate visibility and less need for intrusive monitoring.

How is inventory different from employee surveillance?

Inventory is a declared, bounded information-gathering process. Surveillance is covert, continuous, or overly intrusive observation of individuals. The distinction is not just ethical; it changes behavior. A legitimate inventory asks for voluntary reporting, approved source logs, and business-context information needed for risk triage. It does not require keystroke logging, secret browser capture, or hidden monitoring of private accounts.

A useful decision rule is this: if a method would reveal more about a person’s habits than about the organization’s AI exposure, it is drifting toward surveillance. If a method can be explained in advance, is limited to AI use relevant to work, and collects only the minimum information needed to triage risk, it is much closer to inventory. That rule is proposed here as an editorial test, not as an official standard.

This distinction also supports better adoption. People tend to use workarounds when official tools feel restrictive or unsafe. A respectful inventory acknowledges that undocumented use may be filling a gap, such as speed, convenience, or task quality. That insight aligns with the broader adoption themes in Dagliyan’s research: facilitators and inhibitors shape whether people move toward formal systems or remain in shadow channels.

What should the discovery process include?

Discovery should begin with voluntary reporting, approved-source review, and limited operational interviews. Voluntary reporting invites employees, managers, and team leads to disclose AI tools they use for work, including browser-based assistants, writing tools, code helpers, and workflow automation tools. Approved-source review means checking sanctioned application lists, procurement records, browser allowlists, enterprise subscriptions, and help-desk tickets to identify tools already visible to the organization.

A second source of discovery is a short, nonjudgmental intake form distributed through management channels. It should explain that the purpose is to identify AI use patterns, surface support needs, and evaluate privacy boundaries. It should also state plainly that the inventory is not disciplinary and does not seek personal browsing histories or private account surveillance. That transparency is a practical trust signal, not a guarantee of full disclosure.

Where needed, managers can hold brief listening sessions with teams that handle sensitive data or repetitive text, analysis, or coding tasks. The sessions should ask what employees are using, what problem the tool solves, and whether any work data is pasted into the system. The conversation should stay at the work process level. It should not branch into personal lifestyle, private communications, or unrelated behavior.

Which minimum fields should each inventory record capture?

A lean inventory works best when every record has the same minimum fields. The proposed worksheet below is this article’s original process, not a mandatory standard: user group or function, AI tool or service name, use case, data category involved, whether the tool is approved, whether work output is human-reviewed, and the main concern category. The concern category can include privacy, accuracy, intellectual property, compliance, bias, or vendor dependency.

Two additional fields improve triage without adding much burden: frequency of use and decision impact. Frequency can be occasional, weekly, daily, or embedded in a workflow. Decision impact asks whether the output affects a minor task, an internal decision, a customer-facing artifact, or a sensitive business or people decision. Those two fields help separate casual experimentation from routines that deserve governance attention.

A final optional field is the employee’s suggested alternative if the current tool is disallowed or risky. This matters because a pure “stop using it” instruction often fails. Recording an alternative keeps the inventory solution-oriented and allows the organization to see whether the real issue is tool availability, policy clarity, or workflow design.

How do you triage risk without overcomplicating the process?

Risk triage should sort entries into three action bands: low, moderate, and high. Low-risk entries are typically non-sensitive, low-impact uses of approved tools with human review. Moderate-risk entries involve unapproved tools, repeated use, or exposure of ordinary business data. High-risk entries involve sensitive personal data, confidential business data, regulated information, or uses that directly influence significant decisions without review.

The decision rule can be expressed simply. If a use case combines sensitive data plus unclear vendor controls, move it to high priority. If it uses ordinary data but creates repeated exposure or weak oversight, place it in moderate priority. If it is approved, low-impact, and reviewed by a person before use, it remains low priority. This rule is proposed for operational clarity and should be adapted to the organization’s actual risk appetite and policies.

Triage should lead to one of four actions: accept, replace, restrict, or redesign. Accept means the use is already within policy and acceptable boundaries. Replace means a safer approved alternative should be offered. Restrict means the use can continue only with narrower data or stronger review. Redesign means the workflow itself should change so employees do not need an unsafe workaround.

What privacy boundaries should the inventory respect?

The inventory should collect the minimum data necessary to manage risk. That means avoiding personal account passwords, private message content, full browser histories, or blanket device surveillance. It also means separating the question of “what AI is being used for work” from unrelated questions about employee behavior. NIST’s Privacy Framework emphasizes managing privacy risk through enterprise processes, which supports a bounded, purpose-specific approach.

Data handling boundaries should be stated upfront. Inventory records should be access-limited, retained only as long as needed for governance purposes, and stored separately from performance management files. If a team uses a third-party tool, the inventory should note the tool name and the category of data involved, not copy the data itself. That preserves enough context for oversight without creating a second repository of sensitive content.

Another boundary is communication tone. Leaders should avoid language that implies hidden use is a moral failure. Many employees choose undocumented tools because the approved stack is slow, unavailable, or poorly matched to the task. A privacy-aware inventory treats those reasons as design signals, not excuses to expand monitoring.

How should alternatives be built into the process?

A shadow AI inventory becomes more effective when it points to alternatives. If employees are using unapproved writing assistants, for example, the organization may need an approved drafting tool, a style template, or a review workflow that reduces risk. If teams use external AI for research summaries, the alternative may be a sanctioned internal search process plus human verification. The inventory should therefore include a field for “preferred substitute” or “safe replacement path.”

A practical worksheet can help. Use one row per use case and score each row with three simple prompts: Can the data be shared safely? Is human review present? Is the tool officially approved? If the answer is yes to the first two and yes to the third, the item likely stays low priority. If the answer is no to any of them, the row should be routed for review, alternative selection, or workflow redesign.

This is where adoption thinking becomes concrete. People do not cling to shadow tools only because of curiosity; they often do so because the formal path is inconvenient. The inventory should therefore measure friction as well as risk. If many entries point to the same workaround, that pattern suggests a design problem in the approved environment, not merely an awareness problem.

What does a workable 30-day test plan look like?

The following hypothetical scenario shows how the inventory can be tested. Assumptions: a 200-person professional services team, no prior formal AI register, and a desire to avoid disciplinary framing. In week 1, leadership announces the purpose, the privacy boundary, and the voluntary intake form. In week 2, managers collect self-reported uses and reconcile them with approved-source records. In week 3, the team groups items into low, moderate, and high priority. In week 4, it identifies replacements and policy updates.

The test plan should measure process outcomes, not fabricated performance gains. Useful process checks include: whether employees understood the nonpunitive message, whether the minimum fields were completed consistently, whether any entries were too vague to assess, and whether high-priority items were routed to the right control owner. These checks show whether the inventory is usable without pretending that a universal benchmark exists.

The final step is to close the loop. Share a plain-language summary with staff: what categories were found, what will be approved, what will be restricted, and what alternatives are being explored. Transparency after the inventory matters because it reinforces the purpose of the exercise and makes the next inventory easier to conduct. That supports trust, which is central to durable technology adoption.

Frequently Asked Questions

How is a shadow AI inventory different from an AI ban?

An inventory is a visibility and governance tool, not a prohibition. It identifies where AI is already being used, what data may be involved, and which uses need safer alternatives or tighter review. An AI ban attempts to stop use outright. In practice, inventories are often more realistic because they reveal actual behavior and help organizations reduce hidden risk without forcing everything underground.

What is the minimum data needed for a nonpunitive inventory?

At minimum, record the user group or function, the AI tool or service name, the use case, the data category involved, whether the tool is approved, whether human review exists, and the main concern category. Frequency and decision impact are helpful add-ons. The key is to gather enough detail to triage risk while avoiding unnecessary personal data or private account information.

Why use approved sources if voluntary reporting is already in place?

Approved sources help validate and complete the picture without covert monitoring. Procurement records, sanctioned app lists, help-desk tickets, and enterprise subscriptions can reveal tools already in use or available. Combined with voluntary reporting, these sources reduce blind spots and improve accuracy. The process stays bounded because it focuses on organizational records relevant to work use, not personal surveillance.

What should happen when employees report a tool that is not approved?

The response should not begin with discipline. First, determine what task the tool is solving, what data is being used, and how often it appears in the workflow. Then decide whether the use should be accepted, replaced, restricted, or redesigned. Offering an approved alternative is often more effective than simply telling employees to stop using the shadow tool.

How does privacy fit into a shadow AI inventory?

Privacy is central, not secondary. The inventory should collect the minimum information needed for risk management, avoid personal browsing histories and private messages, and store records separately from performance files. NIST’s Privacy Framework supports managing privacy risk through enterprise processes, which aligns with a purpose-limited approach. The inventory should explain its boundaries clearly so employees know what will and will not be collected.

Sources and further reading

• NIST — AI Risk Management Framework
• NIST — Privacy Framework