Asaf Kochan, President & Cofounder of Sentra, offering data security for the AI era.
getty
Generative AI has moved quickly from experimentation to everyday use. For many organizations, the first exposure to AI does not come from building models in a lab, but from copilots embedded in productivity platforms such as Microsoft 365, Gemini, collaboration tools and internal SaaS applications.
These assistants operate inside the systems employees already use. They search across shared drives, mailboxes, documents and collaboration spaces. They summarize content, retrieve information and generate responses using whatever data those systems make available.
AI systems inherit the same access paths that already exist inside an organization, which often extend further than expected and reach a spectrum of data from current, to archived, to long forgotten.
Microsoft describes this design clearly: Copilot operates inside the Microsoft 365 service boundary, but it does not get tenant-wide visibility by being enabled. Its access is scoped to the signed-in user and the content that person is authorized to access. That’s the point. Copilots don’t “break” permissions, they operationalize them at machine speed. If access is broader than intended, a single prompt can surface what would otherwise take dozens of manual searches to discover.
Years of data and accumulated permissions often remain in collaboration environments. Shared folders stay open long after projects end. Files copied for analysis remain in personal drives. Mailboxes and document repositories accumulate records that were never meant to be widely searchable. AI assistants change how easily this information can be surfaced.
A user who previously needed to search across multiple systems can now retrieve sensitive information through a single prompt. Documents stored in separate locations can appear together in a generated response. Internal data that once required effort to locate becomes immediately accessible.
In most cases, the risk originates in the underlying data environment rather than in the AI system itself. Security teams often lack a clear inventory of the AI tools operating inside their organization and the data those tools can access.
Without that visibility, governance becomes guesswork.
Most enterprise security programs already rely on tools designed to protect data and identities. Data loss prevention monitors files and communications, typically at egress points. Identity systems manage user roles and permissions. Logs record activity after it occurs.
These tools work well within the boundaries for which they were designed. AI assistants operate across those boundaries.
A single AI query can pull data from documents, mailboxes, databases and knowledge repositories in the same interaction. The request may involve multiple services, APIs and identities acting together. The resulting output may combine fragments of information that were never previously connected.
Traditional controls were not built to map these relationships. Organizations therefore struggle to answer basic questions:
• Which AI assistants and agents currently exist inside the environment
• What data those systems can access
• Whether that access exposes regulated or confidential information
Answers to these questions form the foundation of AI governance.
AI data readiness means treating AI assistants and agents as part of the security model, with the same visibility and controls applied to them as any other identity. Gartner estimates that 63% of organizations lack the data management practices required to support AI, which helps explain why many deployments stall before reaching production.
That starts with understanding what data exists across the organization and how AI systems interact with it. Security teams need to maintain an accurate inventory of sensitive data across structured systems, unstructured repositories and collaboration platforms, and map which identities can access it.
AI systems operate using inherited permissions through service accounts, APIs and user context. If those permissions are broader than intended, the AI assistant will access data accordingly. Readiness therefore requires mapping AI identities to the data they can reach and enforcing least-privilege access based on current use.
It also depends on reliable classification. When labels are missing or inconsistent, enforcement systems cannot distinguish between sensitive and non-sensitive data. Maintaining accurate data context ensures that controls such as DLP, access policies and AI guardrails operate as intended.
In practice, this means maintaining a current inventory of AI systems and their data access, classifying sensitive data consistently across environments, treating AI agents as governed identities and feeding data context into existing controls so enforcement reflects actual exposure.
Security teams preparing for AI adoption often follow four operational steps:
1. Create a live inventory of AI systems, including copilots, custom models and automated agents and map them to the data sources they access.
2. Classify sensitive data across cloud storage, SaaS applications and legacy repositories with consistent context.
3. Treat AI agents as first-class identities; their permissions should be evaluated the same way security teams review high-risk service accounts.
4. Feed data intelligence into existing controls such as DLP, IAM and access gateways so those systems enforce policy based on real data exposure rather than pattern matching.
These steps shift security programs from reactive monitoring to governed AI access.
Enterprises are moving quickly to adopt AI assistants because the productivity benefits are clear. What many organizations discover during early deployments is that AI forces them to confront problems that already existed inside their data environment.
Oversharing, inconsistent labeling and fragmented visibility become difficult to ignore once AI systems begin interacting with enterprise data. AI data readiness addresses those conditions before they become incidents.
Organizations that treat AI assistants as extensions of their data infrastructure gain a clearer view of how information moves through their systems and who or what can access it. That visibility allows them to support AI adoption while maintaining control over sensitive data.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。