AI data readiness is the real gap behind agent visibility

At a glance

What this is: This analysis argues that AI agent visibility is secondary to a deeper problem: most organisations do not continuously know what sensitive data they have, where it lives, or who and what can reach it.

Why it matters: IAM, NHI, and AI governance teams need data readiness because access controls cannot be judged accurately when the underlying data estate is unclassified and access paths change faster than periodic reviews.

By the numbers:

• 68% of organisations cannot reliably distinguish human activity from AI agent activity.

Context

AI agent visibility is not the same thing as governance, because counting agents does not reveal whether they can reach sensitive data that was never classified. The core control gap is data readiness: organisations often know more about tools and identities than they know about the data those systems can touch, which leaves AI security decisions built on partial inventory and stale assumptions.

In practical terms, this is a governance problem across IAM, NHI, and AI workflows. A service account, application, or AI agent with access to unclassified customer, financial, or HR data creates the real exposure, while the access path itself is only the route to that exposure. The article treats that condition as common rather than exceptional.

Continuous discovery and classification are the missing layer. When data classifications are stale or absent, periodic access reviews simply certify uncertainty rather than reduce it.

Key questions

Q: What breaks when AI agents are added before data is classified?

A: Security teams lose the ability to judge whether an AI workflow is actually risky because access reviews and agent counts do not reveal what sensitive data is reachable. The result is a governance process that certifies unknown exposure. Classification must exist first, or every downstream control is built on uncertainty.

Q: Why do unclassified datasets create AI governance risk?

A: Unclassified data can be reached by humans, service accounts, or AI agents without anyone recognising that the access path leads to something sensitive. That means the control failure is not the model or the agent, but the missing sensitivity label and reachability map. Once the data is classified, the actual blast radius becomes visible.

Q: How do security teams know whether AI access is actually working safely?

A: Look for three signals: complete discovery of the AI estate, clear mapping of source data to each system, and logs that prove what was accessed and why. If any of those are missing, the control environment is incomplete. Safe AI access is evidenced, not assumed.

Q: How should organisations govern human and non-human access during mergers?

A: They should treat both as part of the same identity estate. Human onboarding, service account inheritance, API keys, and integration credentials can all introduce hidden access paths if they are not reviewed together. A merger is the wrong time to separate IAM from machine identity governance.

Technical breakdown

Why agent counts do not measure exposure

Agent inventories describe how many automated actors exist, but they say little about the sensitivity of the data those actors can reach. Risk is driven by the intersection of identity scope and data classification, not by the number of agents alone. If a system can access public documentation, its presence matters far less than a single identity that can read regulated or proprietary data. In governance terms, visibility without classification produces a false sense of control because the control plane tracks actors while the exposure lives in the data layer.

Practical implication: Treat agent inventory as supporting evidence, not as the primary control for AI security decisions.

How data readiness changes the access review problem

Traditional access reviews assume that reviewers already know which resources are sensitive and which access paths matter. That assumption breaks when data is spread across collaboration tools, file shares, data warehouses, and SaaS systems with no current classification state. Data readiness means continuously discovering data, classifying it, and mapping identities back to that data so access decisions are made against the asset at risk. This is where NHI governance intersects with AI governance, because service accounts, applications, and agents often inherit broad access long before anyone evaluates the underlying data exposure.

Practical implication: Move review logic from identity lists to classified data sets and the identities that can reach them.

Why DSPM for AI is a governance model, not a point tool

DSPM for AI is best understood as a control pattern that keeps data discovery, classification, and access analysis in continuous motion. The important architectural point is that classification must feed enforcement, not sit as a one-time discovery output. When the same dataset can be reached by humans, applications, service accounts, and AI agents, the governance problem becomes shared and dynamic. That is why periodic projects fail: they freeze a moving environment into a static report and then ask security teams to act as if the report is still current.

Practical implication: Require continuous classification and access evaluation before permitting AI workflows to consume sensitive datasets.

NHI Mgmt Group analysis

AI visibility is a downstream metric, not the control boundary. The article correctly shifts attention away from counting agents and toward what those agents can reach. That is the real governance boundary because the exposure sits in the data estate, not in the presence of automation itself. For practitioners, the question is whether classification and access governance are current enough to support AI at all.

Data readiness is the named concept that AI governance teams should adopt. It means knowing, continuously and operationally, what sensitive data exists, where it lives, and which identities can reach it. Without that state, AI governance becomes a policy layer floating above unknown exposure. The practitioner conclusion is straightforward: classification must precede meaningful enforcement.

This is also an NHI governance problem, not only an AI problem. Service accounts, applications, and AI agents often inherit access before anyone evaluates whether the underlying data should be reachable at all. That makes non-human identities part of the data governance surface, especially when access is overbroad or unreviewed. Practitioners should govern access based on data sensitivity and identity type together.

Periodic access review is structurally too slow for AI-era data movement. The article shows why point-in-time certification cannot keep pace with integrations, copilots, and automated workflows that open new access paths continuously. The governance model has to move from retrospective review to continuous evidence of data classification and reachability. Practitioners should treat stale classification as a control failure, not a documentation issue.

OWASP-style agent visibility metrics are useful only when tied to data exposure. A figure such as 68% unable to distinguish human from agent activity is meaningful, but it does not by itself define the risk. The operational question is which identities can reach regulated or proprietary data and whether that access is justified. Practitioners should use visibility as a trigger for data-centric control testing, not as proof of control effectiveness.

From our research library:

• 53% of security leaders expect AI to run major portions of their infrastructure autonomously within the next three years, according to the 2026 Infrastructure Identity Survey.
• 19% of organisations give AI systems dramatically more access than human employees, nearly one in five granting unrestricted privilege, according to the 2026 Infrastructure Identity Survey.
• Read next: Top 10 Agentic AI Identity Issues

What this signals

Data readiness is the missing control plane for AI governance. Organisations do not need more lists of agents if they still cannot say which datasets are sensitive, who can reach them, and whether that access is justified. That is where classification, reachability mapping, and enforcement have to converge if AI security is going to be measurable rather than rhetorical.

Continuous classification is the practical answer to shadow AI and overpermissioned access. When AI workflows, service accounts, and human users all touch the same datasets, the only sustainable model is to classify first and govern access against current sensitivity. This is where NHI and agentic AI governance meet data security in a single operational control. According to the 2026 Infrastructure Identity Survey, 19% of organisations give AI systems dramatically more access than human employees, nearly one in five granting unrestricted privilege.

Access reviews need evidence from classified data, not just entitlement lists. If a team cannot tell whether an accessed record set is sensitive, the review is already behind. Organisations failing to scope AI access properly are 4.5x more likely to experience a security incident, according to the 2026 Infrastructure Identity Survey.

For practitioners

• Implement continuous data classification Classify sensitive data continuously across cloud, SaaS, collaboration platforms, and data warehouses so access decisions reflect current exposure, not last quarter's inventory.
• Map identities to sensitive datasets Build and maintain a living map of which humans, service accounts, applications, and AI agents can reach each sensitive dataset, then prioritise the highest-risk paths first.
• Right-size access before AI enablement Require explicit justification for any AI workflow that can touch regulated, financial, HR, or customer data, and remove unnecessary entitlements before rollout.
• Make classification drive remediation Use classified data findings to trigger cleanup of stale content, redundant datasets, and overbroad access instead of sending unreviewed alerts to operations.

Key takeaways

• AI agent visibility alone does not tell security teams whether the underlying data is sensitive, which is why exposure remains hard to quantify.
• The control gap is continuous data readiness, meaning discovery, classification, and reachability mapping must stay current as access paths change.
• For IAM, NHI, and AI programmes, the practical priority is to govern access from the data outward instead of treating agent counts as the primary security signal.

Key terms

• AI Data Readiness: AI Data Readiness describes whether an organisation can safely expose data to AI systems without losing control over sensitivity, purpose, or access scope. It combines discovery, permission management, and continuous oversight so data use remains aligned to governance expectations.
• Data classification: Data classification is the process of labelling information according to sensitivity, regulatory impact, or business value so controls can be applied consistently. For AI governance, it allows policy to follow the data into prompts, sessions, and destinations rather than relying on brittle text matching.
• Non-Human Identity (NHI): A digital identity assigned to a non-human entity such as a software application, service account, API key, bot, machine, or AI agent that enables it to authenticate and interact with systems without direct human involvement. NHIs now outnumber human identities in most enterprises by 25 to 50 times.
• Reachability Mapping: Reachability mapping is the practice of determining which identities, applications, or automation can actually access a given dataset or resource. It matters because nominal permissions and real access are not always the same, especially across integrated cloud, SaaS, and AI environments.

Deepen your knowledge

NHI Mgmt Group covers identity security, NHI governance, and agentic AI through the NHI Foundation Level course, the industry's only accredited NHI security programme. It gives practitioners a shared control model for machine identities, access scope, and governance decisions that now shape AI risk.