Shadow AI at Work and How to Control It Without Blocking Productivity

Shadow AI at Work and How to Control It Without Blocking Productivity

Managed IT16 min readReviewed November 17, 2026

Shadow AI at Work and How to Control It Without Blocking Productivity

Luis Garcia, CIO of On-Site Technology

By , CIO

On-Site Technology, Clifton, NJ. In IT since 2001: field tech, network engineering, managed security and CMMC compliance.

THE SHORT ANSWER

Shadow AI is the use of AI tools, accounts, or integrations for work without organisational knowledge, approval, or oversight. At On-Site Technology, we see this across clients in manufacturing, professional services, and education where approved tools and clear guidance simply haven’t kept pace with employee demand, with the actual risk driven by what data enters the tool and whether a human reviews the output.

Questions about Managed IT?

Tell us what you are working on. We typically respond within one business day. No obligation.

    Prefer to talk? Call (973) 777-7227
    Your info stays with us. No resale.

    Shadow AI is unapproved or unmanaged AI use at work, typically driven not by employee misconduct but by a gap between the productivity tools staff need and the ones their organisation has actually supplied. Understanding why it happens, what risks it creates, and how to draw clear boundaries is what separates businesses that enable safe AI adoption from those that discover a data problem after the fact.

    The goal here is not a blueprint for locking down every AI tool. Employees often reach for AI to get legitimate work done faster, and blanket prohibitions rarely hold. The practical response is a short, enforceable AI acceptable use policy, approved tools with managed accounts, clear data boundaries, real training, and monitoring that is transparent rather than hidden.

    The central decision every organisation needs to make is not simply “Is AI allowed?” It is: which account may be used for this task, what data may enter it, and who remains accountable for the result? That question distinguishes lower-risk productivity assistance from AI use involving sensitive data, consequential decisions, external publication, or unapproved integrations.

    KEY TAKEAWAYS

    • Shadow AI is unapproved or unmanaged AI use at work, and it can create privacy, IP, accuracy, compliance, spending, and reputational risks.
    • An AI acceptable use policy should permit useful low-risk work while setting firm rules for tools, accounts, data, review, and escalation.
    • Employees need practical examples that separate approved, restricted, and prohibited AI tasks instead of vague instructions to “use AI responsibly.”
    • Monitoring should identify unapproved tools and data-loss risks transparently, while giving employees a non-punitive way to request approval.

    Shadow AI Is a Management Gap Not Just a Policy Violation

    DEFINITION

    Shadow AI refers to the use of AI tools, accounts, integrations, or AI-enabled services for work without the organisation’s awareness, approval, or governance. It includes free public chatbots, paid personal subscriptions, and team-purchased tools bought outside IT, security, or procurement processes.

    Risk does not come simply from software being labelled “AI.” It depends on the specific tool, how it is configured, what account type is involved, what information gets entered, and how the output is used. A well-configured, organisation-managed AI workspace operated under documented data rules is a very different situation from an employee pasting a client contract into a free public chatbot during a deadline crunch.

    78%of AI users surveyed by Microsoft in 2024 brought their own AI tools to work

    That figure comes from Microsoft’s 2024 Work Trend Index, which also found notably higher reported AI use among small and medium-sized businesses. It is survey research from 2024, not a measurement of every workplace, but the scale of the gap it describes is real: employees are acquiring tools faster than governance programmes can accommodate them.

    Common motives are entirely familiar. Drafting and editing, research, meeting notes, summarising long documents, routine communications, coding help, and data analysis are all things people do every day. When no approved tool exists or procurement moves slowly, some staff will find their own solution. The right management response is to investigate the unmet need driving that behaviour, not to assume every employee is trying to evade controls.

    Recognise What Shadow AI Looks Like in Everyday Work

    The most common example is an employee using a public chatbot to rewrite a work document without knowing whether that account is approved. That alone may be low risk or high risk depending entirely on what document is involved. The situation becomes genuinely concerning when a staff member pastes customer details, internal strategy, financial information, source code, or meeting notes into a prompt without any consideration of where that data goes.

    Shadow AI also shows up at a team level: a department purchasing an AI writing, transcription, image-generation, or meeting-summary tool without procurement approval. Nobody in IT knows the subscription exists, nobody reviewed the vendor’s data retention terms, and the account is registered to a personal email address.

    That last detail matters more than it might seem. A tool registered to a personal email creates access problems, offboarding gaps, billing confusion, retention questions, and audit gaps. When the employee leaves, the organisation may have no way to access, transfer, or close the account. Responsible experimentation means using an approved tool under documented data rules with appropriate human review. Unmanaged experimentation means the organisation simply does not know what is happening, and cannot assess whether it is acceptable.

    The distinction between those two things is administrative and cultural, not technical. It comes down to whether the organisation has established an approval path and whether staff know to use it.

    Assess the Risks by Data Task and Outcome

    Confidential data and privacy risks are where shadow AI most commonly creates serious exposure. Sensitive information entered into an unapproved AI tool may be retained, processed, or disclosed in ways the organisation has never evaluated. NIST’s AI Risk Management Framework identifies data privacy, leakage, and unauthorised disclosure as generative-AI risks that organisations need to address in their governance approach.

    Intellectual property risk runs alongside privacy risk. Proprietary prompts may expose trade secrets, and AI-generated output raises copyright and ownership questions that many businesses have not thought through. The US Copyright Office stated in its January 2025 report that copyright protection for AI outputs depends on sufficient human authorship, and that prompting alone is not enough to establish it.

    Hallucination, sometimes called confabulation, is a third risk category that gets underweighted. AI output can be confidently presented but false, incomplete, or misleading. Employees who treat generated content as verified facts, legal advice, financial guidance, or a final business decision are compounding whatever data risk already exists with a decision-quality problem.

    The other risk categories are real but often more visible: compliance and contractual obligations the tool use may breach, uncontrolled subscription costs, fragmented workflows that become harder to audit, and reputational damage from an inaccurate public statement or an improper automated decision. A clear risk rule worth emphasising: higher-risk use involves more sensitive data, more external impact, less human review, or a more consequential decision. Those four factors travel together.

    An AI Acceptable Use Policy Should Make Safe Work Easier

    An AI acceptable use policy is a documented set of rules that explains which AI tools, accounts, data types, and tasks are permitted for work, and what users must do before relying on AI-generated output. It is distinct from a general technology-use policy because it addresses AI-specific inputs, outputs, human review, disclosure obligations, intellectual property considerations, and escalation paths.

    Understanding how it differs from related policies helps position it correctly inside an organisation’s governance framework. A security policy covers protection of systems, accounts, and information. A privacy policy covers handling of personal data and applicable rights or obligations. A technology-use policy covers devices, networks, software, and general acceptable behaviour. The AI acceptable use policy fills a gap those documents leave open: the specific questions employees face when they sit down to use an AI tool for real work.

    The policy should cover employees as a baseline and extend to contractors, temporary staff, and third parties acting on the organisation’s behalf where they have access to company information or systems. Someone should be clearly named as the owner who approves exceptions and schedules reviews.

    Risk-based governance is preferable to treating every AI use identically. NIST’s Generative AI Profile, published 26 July 2024, is a voluntary resource organisations can reference when adapting AI risk management to generative AI tools specifically. It does not prescribe a single approach; it offers a structure for thinking through the risks that apply to a given use case. For most SMBs, that means the policy does not need to be exhaustive. It needs to be clear enough that an employee can decide, before uploading anything, whether a task is in or out.

    Write an AI Policy for Employees Around Permission Data and Decisions

    One useful organising principle for an AI policy for employees is the Permission Data Decision Rule: before using an AI tool for a work task, an employee should confirm that the tool and account are approved, that the data involved is allowed in that environment, and that any decision or output the AI assists with has appropriate human review before it goes anywhere.

    That is not a checklist employees need to memorise. It is a habit the policy reinforces through its structure and language. In practical drafting order, the policy should cover:

    • Purpose and scope: explain why the organisation allows controlled AI use and identify the people, departments, and work activities covered.
    • Approved tools and accounts: name approved services, specify whether personal accounts are prohibited for work purposes, and explain how to request approval for a new tool, integration, or use case.
    • Permitted uses: list lower-risk examples such as brainstorming, drafting non-confidential material, improving grammar, summarising public sources, and generating agenda ideas.
    • Data boundaries: prohibit passwords, credentials, customer data, personal data, confidential business information, financial records, proprietary code, regulated information, and unreleased plans in unapproved AI tools.
    • Human accountability: require fact-checking, editing, source verification, and human ownership before output is shared externally or used in a business decision.
    • Disclosure, records, and rights: explain when AI assistance must be disclosed, when records must be retained, and when copyright, licence, or attribution review is required.
    • Incident reporting: tell staff how to report accidental disclosure, suspicious output, a harmful result, or an apparent policy breach quickly and without fear of immediate punishment.
    • Policy ownership: identify who approves exceptions, how non-compliance is handled, and when the policy is reviewed.

    Policy language should answer the question every employee eventually faces: what do I do when I’m not sure? The answer should be: stop, do not upload the information, and contact the designated manager, IT, security, privacy, legal, or compliance person before proceeding.

    Set a ChatGPT Policy for Employees by Account Type and Data Sensitivity

    A ChatGPT policy for employees must start with a clear statement about which version and account type is permitted. A public free account used for approved, non-confidential tasks is a very different situation from an organisation-managed workspace used for internal work. Those two scenarios carry different data terms, different administrative controls, and different organisational obligations.

    A paid or business-branded account does not automatically make every category of data or decision appropriate. Configuration, access control, retention, contractual terms, and the organisation’s own legal and compliance obligations all require independent review regardless of what a vendor’s marketing says.

    As of 18 November 2026, ChatGPT Business standard seats are listed at US$25 per user per month when billed monthly or US$20 per user per month when billed annually; premium seats are listed at US$125 monthly or US$100 annually, and the workspace requires at least two paid seats. OpenAI states that workspace data is not used to train models by default and that centralised administration is included. Prices, availability, and features can vary by market and change over time, so verify current terms before making purchasing decisions.

    A ChatGPT policy should require employees to use only the approved account for the approved task, verify outputs before sharing them externally, and request approval before adding integrations, entering sensitive information, or applying AI assistance to a high-impact decision.

    Account ownership and administrative control matter more than many organisations realise. We completed a Microsoft 365 migration for a client whose 365 environment came through GoDaddy. That federated tenant structure blocked customisations and third-party integrations across mailboxes, SharePoint, OneDrive, and Teams entirely. Moving the client to their own tenant gave them full administrative control and the ability to implement automation workflows. The lesson applies directly to AI governance: before connecting any new AI service to company systems, the organisation needs to understand who controls the account, what integrations are permitted, and what administrative visibility exists.

    Approved Restricted and Prohibited Examples Make Policy Usable

    A well-written AI policy for employees does not make people read terms and conditions every time they want to draft an email. It gives them examples they can match to real tasks. The table below is a decision aid, not a substitute for the organisation’s legal, privacy, security, or contractual obligations.

    Approved (subject to tool rules)RestrictedProhibited in public or unapproved AI tools
    Improving grammar or clarity in non-confidential writingSummarising internal documentsUploading customer databases, contracts, or unreleased product plans
    Producing agenda ideas or brainstorming campaign conceptsAnalysing business data or financialsEntering credentials, passwords, or authentication data
    Summarising material that is already publicly availableWorking with customer-related informationUploading medical records, confidential source code, or regulated records
    Creating a routine first draft that a staff member will verify and editUsing AI to support decisions affecting customers, employees, finances, or complianceAsking AI to make an unreviewed employment, credit, legal, health, or safety decision
    Generating template structures or agendas for internal planningPurposes requiring a specifically approved environment and documented controlsConnecting an unapproved AI app to company email, file storage, or collaboration platforms

    Restricted uses are not banned. They require a specifically approved environment, an approved purpose, appropriate permissions, and documented data controls. If those conditions are not in place, the task moves into prohibited territory until they are.

    One point that requires explicit policy language: employees should not attempt to anonymise sensitive information by removing an obvious name if the remaining details could still identify a person, customer, project, or confidential matter. Partial anonymisation creates a false sense of safety. The shadow AI risk does not disappear because a column header was deleted.

    A Generative AI Policy for Small Business Can Be Proportionate

    A generative AI policy for small business does not require a governance programme the size of an enterprise compliance function. It requires a document employees can actually read, understand, and follow in the middle of a workday when they are deciding whether to paste something into a prompt.

    A proportionate minimum policy covers approved tools, prohibited data categories, mandatory human review before output is shared or acted on, escalation contacts, incident reporting, and a simple request route for new tools or integrations. That can be a single page with a short annexe of examples. What it cannot be is a vague statement that employees should “use AI responsibly,” which answers no question anyone actually faces.

    A central register of approved AI services should exist from the start, even if it begins as a shared spreadsheet. Each entry should capture the service name, account owner, business owner, integrations in use, renewal date, data categories permitted, and approval status. That register is what makes offboarding manageable and audit responses possible.

    most OST managed IT clients run both as foundational account and endpoint controls: MFA and EDR

    Most On-Site Technology managed IT clients run MFA and EDR as standard controls. MFA is a foundational layer of account protection that applies to AI workspace accounts the same way it applies to email and cloud services. What MFA does not do is replace an AI policy, enforce data boundaries, or ensure human review of output. Those are governance and training problems, not authentication problems. Treat MFA as a prerequisite for using managed accounts, not as a substitute for a policy.

    Co-managed environments need particular attention here. Where an internal IT team and an external IT provider share responsibility, the organisation should document clearly who approves new AI tools, who reviews integrations, who monitors accounts, and who responds to incidents. OST has an established co-managed IT practice, and in those engagements, clearly assigned responsibilities are what keep security and governance from falling through the gap between internal and external teams.

    For businesses planning Microsoft 365 Copilot, licensing is a separate planning issue. As of 4 November 2026, Microsoft’s licensing documentation indicates that Microsoft 365 Copilot availability depends on the underlying Microsoft 365 plan and purchase route, with Microsoft 365 E7 or eligible subscriptions referenced for relevant Copilot offerings. Verify current eligibility, market availability, and pricing before budgeting, since these change.

    How to Monitor AI Use at Work Without Undermining Trust

    How to monitor AI use at work is not purely a technical question. Monitoring should identify unmanaged risk while remaining lawful, fair, transparent, and proportionate to the actual risk involved. A hidden AI surveillance programme that employees discover is worse for trust and policy compliance than the shadow AI it was intended to find.

    Practical monitoring starts with what the organisation already controls. Maintain an approved-tool inventory covering services, accounts, integrations, business owners, and permitted data types. Review software discovery data, single sign-on records, access logs, and procurement records for signs of unapproved subscriptions. Expense claim review catches a surprising number of personal AI subscriptions being submitted for reimbursement. Review AI integrations that can access cloud files, email, collaboration platforms, or customer systems, since those integrations carry the highest data-exposure potential.

    Alongside those technical controls, ask employees what tools they use and provide a non-punitive approval path. Staff who feel they will be disciplined for disclosing a useful tool they adopted will simply not disclose it. The goal is to surface AI use so it can be evaluated, not to create an environment where people hide tools from IT.

    Transparency is not optional. Tell employees what is monitored, why it is monitored, what information is collected, who can access it, and how long it is retained. UK ICO guidance states that worker monitoring must be lawful, fair and transparent, and that email and message monitoring needs a clear purpose and must be necessary and proportionate. Even for US-based businesses outside UK jurisdiction, that standard is a reasonable bar for internal policy design.

    Training is part of monitoring in the sense that educated employees surface problems faster. Cover data classification, prompt safety, hallucinations, fact-checking, copyright considerations, escalation, and incident reporting. Give teams four decision prompts they can keep at hand:

    • Is this AI tool and account approved for this task?
    • Is the information permitted in this specific environment?
    • Have I verified the output and accepted responsibility for it?
    • Should I ask before using AI because this task involves sensitive data, an integration, or a consequential decision?

    The EU AI Act’s AI-literacy provisions became applicable on 2 February 2025, requiring providers and deployers to take measures supporting staff AI literacy in light of their knowledge, experience, and context of use. For businesses operating in or selling into the EU, that is a compliance planning item. For everyone else, the underlying principle applies regardless of jurisdiction: employees who understand what AI can and cannot do make better decisions about when and how to use it.

    Frequently asked questions

    QHow can a business identify shadow AI before it creates a data leak?

    Start by combining approved-tool inventories, software discovery, expense reviews, access logs, integration audits, and direct employee disclosure. Ask what work problem employees are trying to solve with unapproved tools, then provide a safe alternative or a clear approval route. Discovery should be transparent and proportionate: covert surveillance typically damages the trust that makes voluntary disclosure possible, and that disclosure is often the fastest way to surface real risk.

    QWhat should an AI acceptable use policy include for employees?

    Cover scope, approved tools and accounts, permitted tasks, prohibited data, human review requirements, disclosure obligations, incident reporting, exception approvals, and policy ownership. The critical element most policies omit is practical examples. Employees need to know whether a specific task is approved, restricted, or prohibited before they act, not after. Legal language alone does not answer the question someone faces at two in the afternoon when they are deciding whether to paste a client summary into a chatbot.

    QDoes a paid ChatGPT account make it safe to enter confidential business information?

    No. A paid account may offer additional administrative controls and data-retention settings, but suitability for a specific task depends on account configuration, access controls, retention terms, OpenAI’s current data agreements, and the organisation’s own legal, privacy, and contractual obligations. Employees should use only the approved account for the approved purpose and seek guidance from IT, security, legal, or compliance contacts before entering confidential, personal, regulated, or customer information into any AI tool, regardless of the subscription tier.

    QHow often should a small business review its generative AI policy?

    Review it on a regular cycle and whenever the business adopts a new AI tool, connects an integration, changes how sensitive data is handled, experiences an incident, or faces new legal, customer, or contractual requirements. A short, practical policy should evolve with real use cases and employee feedback rather than sit unchanged for years. The AI tool landscape changes quickly enough that a policy written without a scheduled review date will be outdated before most people have read it.

    Need a Reliable IT Partner?

    OST manages and supports IT for small and mid-sized businesses, from help desk to network and security.

    Learn more about our managed IT services.

    Or call (973) 777-7227