Shadow AI in the Enterprise

What does shadow AI look like?

Someone uses a public AI assistant to summarize meeting notes.
Someone pastes customer emails into an AI tool to draft responses.
Someone uses an AI coding assistant outside the approved developer environment.
Someone enables an AI feature inside an existing SaaS platform.
Someone asks an AI tool to analyze employee feedback.
Someone uses a browser extension to rewrite sales messages.
Someone signs up for a free AI product with a corporate email address.
Someone exports data from a business system and uploads it to an AI analytics tool.
Someone builds a small AI workflow because the official process feels too slow.

Some of this use may be low risk.

Some of it may be useful.

Some of it may be dangerous.

The problem is not that employees are interested in AI.

The problem is that unapproved AI can create hidden data, privacy, cyber, vendor, legal, compliance, and operational risk.

Shadow AI creates questions the organization cannot answer:

• What AI tools are being used?
• Who is using them?
• What data are they processing?
• Are prompts or outputs retained?
• Can the vendor use data for training?
• Is sensitive data involved?
• Is customer or employee data involved?
• Is the AI output influencing decisions?
• Are vendors or model providers involved?
• Are contracts in place?
• Are cyber controls operating?
• Are incidents being detected?
• Is monitoring happening?
• Should the use be approved, conditioned, remediated, suspended, or blocked?

What is shadow AI?

Shadow AI is the use of AI tools, AI features, AI models, AI assistants, AI-enabled SaaS functionality, or AI workflows without formal visibility, approval, inventory, risk tiering, controls, monitoring, or governance.

Shadow AI may include:

• public generative AI tools
• free AI accounts created with corporate email
• AI browser extensions
• AI writing assistants
• AI meeting note tools
• AI coding assistants
• AI analytics tools
• AI features inside existing SaaS platforms
• AI chatbots used by teams without approval
• AI automation workflows built outside IT
• AI vendor features enabled without reassessment
• AI tools used with customer, employee, or confidential data
• AI tools used for decision support without review

Shadow AI can be intentional or accidental.

Sometimes users knowingly bypass policy.

More often, they simply do not know what counts as AI governance risk.

They may think:

• “It is just a productivity tool.”
• “The vendor was already approved.”
• “It is only a pilot.”
• “I did not upload sensitive data.”
• “The tool is free.”
• “Everyone is using it.”
• “It is not making decisions.”
• “It is only helping me draft.”

Some of those statements may be true.

But the organization still needs visibility.

Why does shadow AI happen?

Shadow AI happens because demand moves faster than governance.

Employees and teams want to:

• save time
• draft documents faster
• summarize information
• write code
• analyze data
• respond to customers
• automate workflows
• compare vendors
• generate marketing content
• improve productivity
• experiment with new tools
• solve problems without waiting for a formal project

Governance teams may unintentionally encourage shadow AI when:

• intake is too slow
• policies are unclear
• approved tools are not available
• low-risk use is over-controlled
• review paths are confusing
• employees do not know where to ask
• business teams fear punishment for disclosure
• every AI request feels like a legal review
• AI governance is framed only as restriction

The solution is not to eliminate experimentation.

The solution is to create a practical path from experimentation to governance.

Low-risk AI should be easy to disclose and approve.

Higher-risk AI should be routed quickly to the right reviewers.

Prohibited AI should be blocked.

Unapproved AI should become visible.

That is the balance.

Why is shadow AI a GRC issue?

Shadow AI is a GRC issue because it cuts across multiple risk domains.

NIST’s AI RMF encourages organizations to manage AI risks across governance, mapping, measurement, and management activities, and ISO/IEC 42001 frames AI governance as a management system that should be maintained and continually improved. Shadow AI belongs in that operating model because it is unmanaged AI use that needs inventory, risk review, controls, evidence, and monitoring.

The Shadow AI Governance Model

A practical shadow AI governance model has 12 stages:

1. Define what counts as shadow AI.
2. Create a safe disclosure path.
3. Discover unapproved AI use.
4. Triage discovered use cases.
5. Classify by risk tier.
6. Link to data, systems, vendors, and owners.
7. Decide whether to approve, condition, remediate, suspend, or block.
8. Collect evidence.
9. Create issues for gaps.
10. Monitor approved or conditionally approved use.
11. Educate users and improve approved options.
12. Dashboard shadow AI posture and decisions.

The goal is not to punish first.

The goal is to govern.

1. What counts as shadow AI?

Start by defining shadow AI clearly.

Employees cannot follow rules they do not understand.

Shadow AI should include any AI use that is not:

• inventoried
• approved
• risk-tiered
• governed by policy
• reviewed where required
• monitored where required
• supported by evidence
• tied to a business owner

A policy definition might say:

Shadow AI is any AI tool, AI-enabled feature, model, assistant, automation, or AI-supported workflow used for company work without registration, review, approval, or monitoring under the organization’s AI governance process.

Be specific.

Include examples:

• public AI chatbots
• AI writing tools
• AI coding assistants
• AI meeting assistants
• AI analytics tools
• AI browser extensions
• AI features inside SaaS platforms
• AI vendor tools
• internal models or prototypes
• AI agents or automations

Also define what does not require full review.

For example:

• approved enterprise AI tools used within policy
• low-risk internal AI use registered through lightweight intake
• AI functionality already covered by an approved use case and within scope

A clear definition helps employees disclose use without fear or confusion.

Shadow AI definition checklist

2. Why does disclosure need to be safe?

If employees think disclosure leads to punishment, they will hide AI use.

The organization needs a safe path to disclose AI use and bring it into governance.

A safe disclosure process should be:

• simple
• fast
• non-punitive for good-faith disclosure
• clear about prohibited uses
• clear about sensitive data restrictions
• connected to AI intake
• able to route low-risk use quickly
• able to escalate risky use immediately
• able to create issues where remediation is needed

The message should be:

“Tell us what AI tools are being used so we can help approve, secure, or replace them. We are trying to govern AI responsibly, not punish responsible disclosure.”

A safe disclosure intake should ask:

• What tool are you using?
• What are you using it for?
• What data are you entering?
• Is customer, employee, sensitive, confidential, or regulated data involved?
• Is the output used internally or externally?
• Does it affect decisions?
• Is a vendor account involved?
• Is the tool connected to any systems?
• Is the use ongoing, pilot, or one-time?
• Who owns the use case?

This is not a full review.

It is a triage intake.

Safe disclosure checklist

3. How do you discover unapproved AI use?

Disclosure alone is not enough.

Organizations should also discover shadow AI through multiple signals.

Possible discovery sources include:

• expense reports
• procurement requests
• vendor renewals
• software asset management
• SSO logs
• DNS or network logs
• browser extension inventories
• CASB or SaaS discovery tools
• DLP alerts
• endpoint telemetry
• helpdesk tickets
• employee surveys
• training attestations
• business unit interviews
• data loss investigations
• cyber alerts
• vendor questionnaires
• contract reviews
• privacy assessments
• AI policy exception requests
• internal audit reviews

Discovery should be coordinated with privacy, cyber, procurement, legal, and IT.

Be careful with how discovery is framed.

The goal is not surveillance for its own sake.

The goal is risk visibility.

When discovery identifies unapproved AI, the next step should be triage, not automatic punishment.

Discovery checklist

Discovery should feed the AI inventory and issue workflow.

4. What should triage answer?

Not all shadow AI should be treated the same.

Once unapproved AI use is found, triage it quickly.

Triage should answer:

• Is this actually AI?
• Is it still in use?
• Who owns it?
• What business purpose does it support?
• What data is involved?
• Is sensitive data involved?
• Is customer or employee data involved?
• Is confidential data involved?
• Is a vendor or model provider involved?
• Is it connected to company systems?
• Is output used externally?
• Does it influence decisions?
• Is there immediate risk?
• Should it be suspended while reviewed?

A triage record should produce one of several outcomes:

The triage workflow should be fast.

If triage takes weeks, shadow AI will stay hidden.

Shadow AI triage checklist

5. Classify Shadow AI by Risk Tier

Discovered AI should be classified by risk tier.

Use the same tier model as approved AI use cases.

A practical model:

The EU AI Act’s risk-based approach is useful here because it reinforces that different AI systems create different levels of risk, from unacceptable to high, limited, and minimal or no risk. Internal risk tiering should apply the same principle: not all shadow AI deserves the same response.

Risk tiering should consider:

• data sensitivity
• decision impact
• affected stakeholders
• vendor involvement
• system integration
• external exposure
• autonomy
• human oversight
• monitoring availability
• legal or regulatory relevance
• potential harm

A low-risk unapproved AI tool may be brought into governance with lightweight review.

A high-risk unapproved AI tool may need to stop immediately.

Shadow AI risk tiering checklist

6. Link Shadow AI to Data, Systems, Vendors, and Owners

Shadow AI should not remain a standalone note.

It should be linked to source records.

Connect it to:

• AI inventory
• business owner
• business process
• data inventory
• data owner
• system record
• system owner
• vendor record
• model provider
• contract
• privacy review
• cyber review
• controls
• evidence
• issues
• risk acceptance
• monitoring
• dashboard

This is what turns discovery into governance.

For example:

“Marketing team uses AI writing tool” is not enough.

A connected record should show:

• marketing content generation use case
• marketing owner
• tool vendor
• data used
• whether customer data is entered
• whether outputs are external
• contract status
• risk tier
• approved scope
• conditions
• monitoring
• issues

SmartSuite’s AI Governance page describes centralized inventories with owners, lifecycle stages, linked business context, risk and performance assessments, monitoring cycles, issues, remediation, evidence, and dashboards. That connected record architecture is exactly what shadow AI needs.

Linked-record checklist

7. What happens after triage?

After triage and risk tiering, decide what happens next.

Possible decisions:

This decision should be documented.

It should include:

• decision
• rationale
• approver
• approved scope
• restrictions
• conditions
• evidence required
• monitoring
• expiration or reassessment date
• issue linkage
• risk acceptance, if needed

Do not quietly “allow” shadow AI without converting it into an approved or conditionally approved record.

Decision checklist

8. Collect Evidence

Shadow AI remediation requires evidence.

Evidence may include:

• intake disclosure
• discovery source
• business purpose
• owner assignment
• data-use assessment
• risk tier rationale
• privacy review
• cyber review
• vendor review
• contract review
• approved-use guidance
• employee attestation
• access removal evidence
• tool deactivation evidence
• monitoring evidence
• issue remediation evidence
• risk acceptance record
• approval record
• reassessment record

Evidence should answer:

• What was found?
• Who used it?
• What was it used for?
• What data was involved?
• What risk tier applies?
• What decision was made?
• What changed?
• What proof supports closure?

Do not close shadow AI issues based only on a verbal statement that the tool is no longer used.

If use must stop, retain evidence that access was removed, subscription canceled, browser extension removed, vendor disabled, data deleted, or approved alternative implemented.

Shadow AI evidence checklist

9. Create Issues for Gaps

Shadow AI often reveals control gaps.

Common issues include:

• unapproved AI tool used with customer data
• sensitive data entered into public AI tool
• AI vendor not reviewed
• contract terms missing
• prompt/output retention unknown
• model provider unknown
• browser extension installed without approval
• AI coding tool used without enterprise terms
• AI feature enabled in SaaS platform without review
• no human oversight for AI-generated output
• output used externally without approval
• monitoring missing
• policy unclear
• employees unaware of approved tools
• intake process too slow

Each issue should include:

• issue source
• affected AI use case
• affected data
• affected owner
• affected vendor
• affected system
• severity
• root cause
• remediation plan
• due date
• evidence
• validation
• risk acceptance, if needed
• dashboard status

Root cause is important.

A shadow AI issue may not be a user problem.

It may reveal that:

• approved tools are unavailable
• policy is unclear
• intake is too slow
• training is ineffective
• procurement does not route AI tools
• SaaS AI features are not monitored
• browser extension governance is weak
• vendor renewals do not include AI review

Fix the root cause.

Do not only fix the individual instance.

Issue checklist

10. Monitor Approved or Conditionally Approved Use

Once shadow AI is brought into governance, it needs monitoring if risk warrants it.

Monitoring may include:

• continued use within approved scope
• data restrictions
• prompt and output review
• vendor term changes
• model provider changes
• user access
• browser extension usage
• sensitive data detection
• human oversight evidence
• output quality
• incidents
• approval conditions
• risk acceptance expiration
• reassessment triggers

Low-risk use may require only periodic attestation.

Moderate-risk use may require usage review and owner certification.

High-risk use may require formal monitoring metrics, thresholds, and evidence.

Do not approve previously shadow AI and then forget it.

That simply moves hidden risk into approved risk.

Monitoring checklist

11. Why do approved options matter?

Shadow AI is partly a governance problem and partly an enablement problem.

If employees do not have approved tools, they will find their own.

If policies are hard to understand, they will guess.

If intake is slow, they will bypass it.

If training only says “do not use AI,” it will not work.

A better enablement model includes:

• approved AI tool list
• prohibited AI use examples
• sensitive data guidance
• prompt and output rules
• when to submit intake
• when to ask privacy, legal, or cyber
• examples by role
• quick approval path for low-risk use
• office hours
• reusable templates
• approved vendor catalog
• sanctioned AI environments
• feedback loop for common use cases

Training should be practical.

Example guidance:

• Do not enter customer, employee, confidential, regulated, or security-sensitive data into unapproved AI tools.
• Do use approved enterprise AI tools for internal drafting within policy.
• Do submit intake before using AI with customer data, employee data, vendor tools, production systems, or decision support.
• Do not enable AI features in SaaS platforms without review.
• Do report AI tools already being used so they can be reviewed.

Governance works better when approved paths are easy.

Enablement checklist

12. Dashboard Shadow AI Posture and Decisions

Executives and governance teams need visibility into shadow AI.

Dashboard views should show:

• discovered shadow AI use cases
• shadow AI by business unit
• shadow AI by source
• shadow AI by risk tier
• shadow AI involving sensitive data
• shadow AI involving vendors
• shadow AI involving production systems
• shadow AI pending triage
• shadow AI routed to intake
• shadow AI approved
• shadow AI approved with conditions
• shadow AI suspended or blocked
• shadow AI issues open
• shadow AI remediation overdue
• repeat shadow AI sources
• policy gaps identified
• decisions needed

Dashboard example:

SmartSuite’s AI Governance page describes dashboards that show coverage gaps, emerging risks, exception trends, governance readiness, issue trends, and linked governance records. That is the type of reporting shadow AI requires.

The dashboard should not shame teams.

It should show risk, action, and improvement.

Common Shadow AI Scenarios

Scenario 1: Employee uses public AI to summarize customer emails

Risk concerns:

• customer data
• prompt retention
• vendor training
• privacy review
• contract absence
• possible confidentiality issue

Response:

• triage immediately
• stop sensitive data use pending review
• assess data entered
• determine whether incident review is needed
• educate user
• identify approved alternative
• create issue if customer data was exposed
• consider blocking or monitoring tool access

Possible outcome:

Suspended pending review. Issue opened for customer data exposure assessment. Approved enterprise AI tool recommended for future use.

Scenario 2: Team uses AI meeting assistant without approval

Risk concerns:

• employee discussions
• customer conversations
• confidential business data
• recording consent
• vendor retention
• transcript storage
• access control

Response:

• identify meeting types recorded
• assess data categories
• review vendor terms
• route privacy and legal review
• determine approved use limits
• require notice or consent where needed
• define retention

Possible outcome:

Conditionally approved for internal meetings only. Not approved for customer or employee-sensitive meetings until privacy and legal review complete.

Scenario 3: Existing SaaS vendor enables AI feature

Risk concerns:

• embedded AI
• vendor already approved for non-AI use
• new data processing purpose
• model provider unknown
• prompt/output retention
• training terms
• monitoring

Response:

• create AI use case record
• link vendor and contract
• route vendor, legal, privacy, and cyber review
• disable feature until approval if risk is material
• update renewal checklist

Possible outcome:

AI feature disabled pending review. Vendor issue created for model-provider and training-term disclosure.

Scenario 4: Developer uses free AI coding assistant

Risk concerns:

• source code exposure
• secrets in prompts
• IP and output ownership
• vendor training
• insecure code suggestions
• lack of enterprise controls

Response:

• stop unapproved tool use
• assess code or secrets entered
• route cyber and legal review
• provide approved enterprise coding assistant
• train developers
• monitor for recurrence

Possible outcome:

Free tool blocked. Enterprise AI coding assistant approved with restrictions, training disabled, secret-scanning control, and developer guidance.

Scenario 5: Analyst uploads employee data to AI analytics tool

Risk concerns:

• employee data
• potential people-impacting output
• privacy and HR review
• vendor review
• retention
• model behavior
• legal exposure

Response:

• suspend use immediately
• identify data uploaded
• assess incident or privacy issue
• route privacy, HR, legal, cyber, and AI governance review
• document remediation
• consider risk acceptance only if appropriate

Possible outcome:

Use suspended. Privacy issue opened. DPIA required before any future analytics use. Executive review required for employee-impacting AI.

Shadow AI Response Matrix

Use this response matrix for consistent decisions.

This matrix helps teams act quickly without reinventing the response each time.

Shadow AI Metrics

Useful metrics include:

Do not use metrics only to penalize teams.

Use them to improve governance, training, tooling, and approved AI options.

Common Mistakes to Avoid

Mistake 1: Responding only with prohibition

If employees need AI capabilities, prohibition alone drives more shadow AI.

Create approved paths.

Mistake 2: Treating all shadow AI as equally risky

Some use is low risk.

Some use is unacceptable.

Use risk tiering.

Mistake 3: Punishing disclosure

If disclosure feels punitive, shadow AI will go deeper underground.

Encourage good-faith disclosure.

Mistake 4: Ignoring embedded AI

Shadow AI is not only public AI tools.

It can appear inside existing SaaS platforms.

Mistake 5: Not linking shadow AI to data

Data determines risk.

Every discovered use should identify data categories.

Mistake 6: Closing issues without evidence

If a tool is suspended, blocked, remediated, or approved, evidence should prove it.

Mistake 7: Not addressing root cause

If shadow AI keeps appearing, the issue may be unclear policy, lack of approved tools, slow intake, or poor training.

Mistake 8: Not monitoring recurrence

Shadow AI governance needs ongoing monitoring.

Not a one-time cleanup.

30-Day Shadow AI Governance Plan

Days 1–5: Define shadow AI and disclosure path

Create:

• shadow AI definition
• approved-use examples
• prohibited-use examples
• disclosure intake
• triage owner
• employee communication

Days 6–10: Identify discovery sources

Connect:

• expense reports
• procurement
• vendor renewals
• SSO logs
• SaaS discovery
• browser extensions
• DLP alerts
• privacy reviews
• cyber alerts
• employee survey

Days 11–15: Build triage and risk tiering

Define:

• triage questions
• risk tiers
• prohibited-use triggers
• review routing
• immediate suspension criteria
• issue triggers

Days 16–20: Create remediation and approval workflow

Create workflows for:

• approve
• approve with conditions
• route to intake
• suspend
• block
• remediate
• risk acceptance
• replace with approved tool

Days 21–25: Pilot with real findings

Pick:

• one low-risk tool
• one public AI tool
• one embedded SaaS AI feature
• one AI coding tool
• one sensitive-data use case

Run the workflow.

Days 26–30: Launch dashboard and training

Create dashboards for:

• discovered shadow AI
• risk tiers
• open issues
• blocked tools
• approved alternatives
• repeat patterns
• decisions needed

Then update training and approved AI guidance based on what you found.

Shadow AI Governance Checklist

Use this checklist to assess your program.

If several answers are no, shadow AI is probably being handled reactively.

A Practical Test for One Shadow AI Finding

Pick one unapproved AI tool or use case found in the last quarter.

Ask whether your GRC model can show:

• discovery source
• tool name
• business owner
• use case
• business purpose
• data involved
• sensitive data status
• vendor or model provider
• system integration
• output use
• decision impact
• risk tier
• privacy review status
• cyber review status
• vendor review status
• decision outcome
• approval conditions
• issue records
• remediation evidence
• validation status
• monitoring plan
• dashboard status

If answering those questions requires emails, spreadsheets, security logs, vendor notes, privacy files, and meetings, shadow AI is not connected enough.

That is common.

It is also the opportunity.

Final Thought

Shadow AI is not a sign that employees are irresponsible.

It is a sign that AI demand has moved faster than governance.

The organization can respond with fear and prohibition.

Or it can respond with visibility, triage, risk tiering, practical review, approved paths, evidence, monitoring, and better enablement.

The second approach works better.

It brings unapproved AI into the open.
It separates low-risk use from high-risk use.
It routes privacy, cyber, legal, and vendor reviews when needed.
It creates issues for real gaps.
It suspends or blocks unacceptable use.
It gives employees approved alternatives.
It gives executives a dashboard.
It turns hidden AI into governed AI.

That is the purpose of Connected GRC.

Not to stop AI.

To make AI visible, accountable, evidenced, monitored, and safe enough to use responsibly.