Shadow AI Executive Briefing for Boards

Staff are already using AI tools the organisation has not approved, and company data, credentials and customer information are going with them. This briefing is for directors, chief executives, audit and risk committee members, CISOs and general counsel who must decide how the organisation allows, governs and monitors AI use. It explains what shadow AI is in business terms, why banning it rarely works, what the evidence says about its cost, how to test management's answers and what a credible first 90 days looks like. It stays at board level. The practitioner detail is in the Shadow AI Discovery Guide and the Shadow AI Policy Template.

The risk in one paragraph

Shadow AI is any use of AI tools, models, browser extensions, plug-ins or agents for work without the organisation's knowledge or approval. It ranges from an employee pasting a contract into a free chatbot, to a team connecting an AI note-taker to the corporate calendar and mailbox, to a developer giving an AI coding agent a production API key. Each one can send confidential data to a provider the organisation has no contract with, create OAuth grants and tokens that nobody tracks, and produce outputs that nobody has checked before they reach customers. The risk is not that people use AI; that is now normal and often valuable. The risk is that the organisation cannot say which tools hold its data, which ones can act on its systems, or who is accountable when something goes wrong. The board influences the outcome through three choices: whether approved tools are good enough that staff prefer them, how clearly the rules on data are set and enforced, and whether AI tools are governed as identities with access rather than as software purchases.

Why this is a board-level issue now

Use is widespread and mostly self-provided. The Microsoft and LinkedIn 2024 Work Trend Index, a survey of about 31,000 knowledge workers in 31 markets, found 75% using generative AI at work and 78% of those users bringing their own AI tools, rising to 80% in small and medium-sized companies.
Shadow AI now shows up in breach costs. IBM's Cost of a Data Breach Report 2025 found that one in five organisations studied reported a breach due to shadow AI, and that organisations with high levels of shadow AI had average breach costs $670,000 higher than those with low or none. In shadow AI incidents, personal data was compromised in 65% of cases against a 53% average. See Shadow AI Statistics.
Governance lags use. The same IBM study found 63% of breached organisations had no AI governance policy or were still developing one, and only 34% of those with a policy audited regularly for unsanctioned AI.
Regulation reaches ordinary staff use. Article 4 of the EU AI Act, on AI literacy, has applied since 2 February 2025 and requires providers and deployers of AI systems to take measures on the AI literacy of staff who use AI on their behalf. Privacy law already applies whenever personal data goes into an AI tool. See the EU AI Act AI Literacy Obligations (Article 4) Guide.
AI tools now act, not just answer. Agents, AI browser extensions and connected assistants hold OAuth tokens and API keys to mail, files and code. In the Vercel Context.ai OAuth breach, a forgotten AI app's stolen token exposed customer secrets. Shadow AI has become a shadow identity problem.

The forms shadow AI takes

Form	Example	Main business risk	Control management should evidence
Consumer chat and assistant tools	Staff paste documents, code or customer data into a free AI service	Confidential or personal data leaves under consumer terms	Approved enterprise alternative; data rules; web and data loss prevention controls
AI features inside approved SaaS	A supplier switches on AI summarisation in an existing platform	Data processed in new ways the contract and privacy notice do not cover	Supplier AI review in procurement and renewals
Connected AI apps and note-takers	An AI tool granted OAuth access to calendar, mail or files	Persistent tokens with broad access, held by an unknown third party	OAuth app consent governance and periodic grant reviews
Browser extensions	An AI writing or meeting extension that reads every page	Access to sessions and everything on screen	Extension allow-listing on managed browsers
Developer and agent tools	Coding agents, local models, unsanctioned MCP servers using company API keys	Code, secrets and production access exposed; actions taken without oversight	Inventory of agents and the credentials they use. See the Unsanctioned AI Agents and MCP Servers Guide.
Unvetted outputs	AI-written advice, code or content sent to customers unchecked	Errors, liability and reputational harm	Human review rules for customer-facing and decision-making use

Ban, allow or govern: the decision

Approach	What it means	When it fits	Trade-off
Block	Prohibit and technically block public AI services	Short term, for highly classified environments or while approved tools are procured	Pushes use to personal devices, where there is no visibility at all
Allow with rules	Permit tools, with written rules on data and use	Small organisations with limited tooling budgets	Relies on staff judgement; little evidence of compliance
Govern and provide	Offer approved enterprise tools, set data tiers, monitor and review exceptions	Most organisations, and any regulated one	Requires licence spend, monitoring and an intake process that is fast enough

The pattern that works is to make the approved route the easiest route. After staff pasted source code into a public chatbot in the Samsung ChatGPT leak, the company restricted use of external generative AI; a ban buys time but does not remove demand. See the Top 10 Shadow AI Risks.

Five questions for the board to ask

Which AI tools are in use across the organisation today, and how do we know? A credible answer comes from evidence (network and browser telemetry, expense data, OAuth grants, extension inventories), not a staff survey alone.
What approved AI tools do staff have, and are they good enough that people choose them? If the approved option is slower or weaker, shadow use will continue regardless of policy.
What data are staff allowed to put into which tools? Expect a simple data classification mapped to approved tools, with contract terms that cover data retention and use for model training.
Which AI apps and agents hold access to our systems, and who owns each one? Every OAuth grant, API key and service account used by an AI tool is a non-human identity that needs an owner, minimum permissions and a review date.
How long does it take to approve a new AI tool, and how many requests are waiting? A slow intake process is one of the strongest drivers of shadow use.

Red flags in management answers

"We have banned ChatGPT": a ban without an approved alternative and evidence of enforcement usually means use has moved out of sight.
"Everyone has signed the policy": attestation is not evidence. Ask what monitoring shows.
"Our SaaS suppliers handle AI": supplier AI features change how data is processed; ask when contracts and privacy notices were last reviewed for AI.
"AI is an IT procurement matter": AI tools that connect to mail, files and code are access decisions and belong in identity governance.
"We do not have any AI agents": developer tools and assistants increasingly act autonomously. Ask how that statement was checked.

What good looks like

A clear policy: an AI acceptable use policy that names approved tools, data tiers and prohibited uses, owned by an executive. See the Shadow AI Policy Template.
Approved tools with the right terms: enterprise agreements covering data retention, training use and location, with single sign-on so access ends when people leave.
Visibility: continuous discovery of AI services, extensions, OAuth grants and agents, recorded in an AI inventory (NIST CSF 2.0 ID.AM-02).
Data guardrails: data loss prevention rules covering uploads and pastes to AI services for the most sensitive data classes. See the Data Loss Prevention Guide.
Identity governance for AI: AI apps and agents treated as identities with owners, least privilege and access reviews. See the Identity Governance for AI Agents Guide.
Literacy: role-based training on safe AI use, recorded as evidence. See the AI Literacy Training Plan Template.

Where to invest first

Priority	Investment	Why first	Typical owner
1	Discovery of AI use (web, browser, OAuth grants, expenses)	Gives the facts for every later decision	CISO
2	An approved enterprise AI tool with SSO and suitable contract terms	Removes the main reason for shadow use	CIO with procurement and legal
3	AI acceptable use policy and fast intake process	Sets the rules and a route to say yes	CISO, legal and data protection officer
4	OAuth consent and extension controls	Closes the routes that give AI tools standing access	Identity and endpoint teams
5	Data loss prevention for AI destinations	Protects the most sensitive data classes	Data security team

Smaller organisations get most value from items 2 and 3 plus admin-only consent for third-party apps in their productivity suite. Larger organisations need items 1, 4 and 5 to see and control use across many teams. See the AI Usage Control and Shadow AI buyer's guide.

Metrics to track

Metric	What it shows	Good direction
Distinct unapproved AI services seen per month	Scale of shadow use	Falling as approved tools are adopted
Share of AI traffic going to approved tools (%)	Whether the approved route is winning	Rising
AI apps and agents with OAuth grants or keys, and share with a named owner	Unowned standing access	100% owned
Data loss prevention events to AI destinations, by data class	Sensitive data exposure attempts	Falling for restricted data
Median days to approve or reject an AI tool request	Whether intake drives shadow use	Short and stable
Staff completing AI literacy training (%)	Evidence for Article 4 style duties	Rising to full coverage

See the AI Usage Monitoring Guide.

A 90-day plan

Days 1 to 30, find out: run discovery across web traffic, browsers, OAuth grants in the productivity suite and expense claims; produce a first AI inventory with the top tools, the teams using them and the data involved.
Days 1 to 30, stop the worst: revoke OAuth grants held by unknown AI apps with broad mail or file access; set third-party app consent to admin approval for high-risk permissions.
Days 31 to 60, provide and set rules: confirm an approved enterprise AI tool for general use; publish the acceptable use policy with data tiers; launch a light intake process with a target decision time.
Days 31 to 60, govern access: give every AI app, agent and integration a named owner, minimum permissions and a review date, inside the existing access review cycle.
Days 61 to 90, monitor and train: turn on data loss prevention rules for restricted data to AI destinations; start role-based AI literacy training; publish the first metrics pack.
Day 90, review: the risk committee reviews the inventory trend, open exceptions and intake backlog, and sets risk appetite for agentic tools.

Common failure patterns

Policy without provision: rules that forbid public tools while offering nothing approved.
Counting websites only: discovery that sees chat sites but misses OAuth-connected apps, extensions and AI features in existing SaaS.
Forgotten grants: AI trials that end while their tokens stay valid, as in the Vercel Context.ai case.
No owner for agents: developer agents using personal or shared API keys with production access. See the Secrets Management Guide.
One-off survey: a single inventory exercise that is never repeated while new tools appear weekly.

Framework mapping

Framework	Reference	Board relevance
NIST AI RMF 1.0 and AI 600-1	Govern, Map, Measure, Manage; Generative AI Profile	Governance structure for AI risk, including third-party generative AI
NIST CSF 2.0	GV.PO-01, ID.AM-02, PR.DS-02	Policy, inventory of software and services, data in transit
NIST SP 800-53 Rev. 5	CM-10, CM-11, PL-4, AT-2, AC-6	Software usage restrictions, user-installed software, rules of behaviour, literacy training, least privilege
EU AI Act	Article 4 AI literacy	Staff literacy duty for providers and deployers
ISO/IEC 42001	AI management system	Certifiable governance structure for AI use
MITRE ATT&CK	T1567 Exfiltration Over Web Service	Detection logic for uploads to web services, including AI services

Practitioner checklist

Maintain an AI inventory covering services, SaaS AI features, extensions, OAuth-connected apps and agents.
Provide at least one approved enterprise AI tool behind single sign-on with reviewed contract terms.
Publish data tiers that state which data may go into which AI tools.
Require admin approval for third-party app consent with mail, file or directory scopes.
Allow-list browser extensions on managed browsers.
Assign an owner, least privilege and a review date to every AI app, agent and API key.
Apply data loss prevention rules for restricted data to AI destinations.
Run role-based AI literacy training and keep completion records.

Standards and references

EU AI Act: Article 4 AI Literacy
NIST: AI Risk Management Framework
IBM: Cost of a Data Breach Report 2025 announcement
Microsoft and LinkedIn: 2024 Work Trend Index
MITRE ATT&CK: T1567 Exfiltration Over Web Service
NIST SP 800-53 Rev. 5: CM-10, CM-11, PL-4, AT-2, AC-6; NIST CSF 2.0 GV.PO-01, ID.AM-02

How NHI Mgmt Group can help

We provide independent board briefings and risk assessments on the identities AI tools create: OAuth grants, API keys, service accounts and agent credentials that shadow AI leaves behind. Our NHI Foundation Level Training Course, including its agentic AI section, helps security and engineering teams govern these non-human identities. Contact us to arrange a briefing.

Shadow AI Executive Briefing: Board Questions, Controls and a 90-Day Plan