What shadow AI is, and why it matters
Shadow AI is any use of AI that the organisation has not assessed, approved or recorded. It is rarely malicious. Staff use public assistants to draft documents, analyse spreadsheets or summarise client emails because it saves time. Vendors switch on AI features inside platforms you already pay for. Teams connect automation tools to corporate data using personal accounts.
The risk is not AI itself. It is the combination of sensitive data, unknown access and no accountable owner:
- Data exposure: personal, payment or confidential client data pasted into tools that retain prompts or use them for training
- Unmanaged access: AI agents and connectors holding broad permissions to email, files, CRM or code repositories
- Decision risk: AI outputs influencing credit, pricing, hiring, claims or customer communications without human review
- Regulatory exposure: obligations under the EU AI Act, GDPR, POPIA and sector regulators that apply whether or not the use was approved
Questions every board should ask
Directors do not need to understand how models work. They need answers to five governance questions:
- Where is AI used? Is there a current register of AI tools, features, integrations and use cases, including vendor-enabled features?
- What can it access? Which data types and systems can each AI tool reach, and under whose identity?
- Who owns the risk? Does every material use case have a named business owner and technical owner?
- How are decisions made? Who approves new AI use, on what criteria, and where is the decision recorded?
- What happens when it fails? How would we detect misuse, data leakage or harmful output, and is AI included in incident response plans?
How to discover shadow AI
Discovery works best when it combines technical evidence with conversation:
- Review identity-provider logs and OAuth grants for AI applications connected to corporate accounts
- Analyse web-proxy, DNS or CASB data for traffic to AI services
- Audit SaaS admin consoles for AI features that vendors have enabled by default
- Check expense and procurement records for AI subscriptions
- Run short, non-punitive surveys asking teams what tools they use and why
Make discovery an amnesty, not an investigation. People disclose more when the goal is to make useful tools safe rather than to remove them.
From discovery to control
Once you know what exists, apply proportionate controls. Approve and provide enterprise versions of the most valuable tools so staff have a safe alternative. Set clear rules for prohibited data and mandatory human review. Restrict high-risk connectors and agent permissions. Then monitor for new, unapproved use and review the register on a set cadence.
Remember that obligations already apply. Under Article 4 of the EU AI Act, providers and deployers must take measures to ensure sufficient AI literacy among their staff, and this has applied since 2 February 2025. Data-protection law applies whenever prompts or outputs contain personal information.
What board readiness looks like
A board is ready when management can report, at each meeting or quarterly, on the AI register and changes since the last report, the material use cases and their risk ratings, exceptions and incidents, and decisions needing board direction. That reporting should rest on evidence from discovery and monitoring, not on self-declaration alone.
Common mistakes to avoid
- Running discovery as a disciplinary exercise, which discourages disclosure
- Relying only on surveys and missing OAuth grants and vendor-enabled features
- Approving enterprise AI tools without configuring data retention, access and logging
- Forgetting AI agents and automation platforms that act on behalf of users
- Reporting AI risk to the board once, rather than on a regular cadence
Frequently asked questions
Is shadow AI a data-protection breach?
Not automatically, but it can become one. If staff paste personal information into a tool without a lawful basis, appropriate contract or adequate safeguards, it may breach GDPR or POPIA. Under POPIA, any resulting security compromise must be reported to the Information Regulator.
Should we block public AI tools?
Blocking alone tends to fail. A better approach is to provide approved enterprise tools with suitable data protections, set clear rules for prohibited data, and block or monitor only the highest-risk services.
How often should the board hear about AI risk?
At least quarterly while AI use is growing quickly, with immediate escalation for material incidents. Many boards fold AI into their regular technology and cyber risk reporting.
What evidence shows that shadow AI is under control?
A current AI register, discovery results over time, records of approval decisions, monitoring of exceptions, and training completion for relevant staff.
Board readiness means being able to explain where AI is used, what it can access, who owns the risk and how the organisation will respond when it fails.
SecuriCentrix can run a shadow AI discovery exercise and deliver a board-ready AI risk briefing within weeks.
Free readiness checklists
Work through the checks with your team, then talk to us about any gaps.
ISO 27001 Readiness Checklist
25 checks · ISO/IEC 27001:2022
Get the free checklistIreland and EUNIS2 Readiness Checklist
25 checks · Directive (EU) 2022/2555
Get the free checklistRelated reading
- AI Governance QuickStart
- What boards should receive from a cyber exposure assessment
- AIBound AI governance platform
Sources
This article is general guidance, not legal advice.
