Data access governance for AI agents needs data-level controls

At a glance

What this is: This is Collibra’s view that data access governance now needs data-level controls because traditional IAM and ticket-based workflows are too slow for analytics and AI use cases.

Why it matters: IAM, IGA and data governance teams need to reconcile fast self-service access with masking, entitlement control and auditability across both human users and AI-driven consumers.

By the numbers:

• 92% of data consumers at banks said the data they need is often unavailable or takes too long to be made available, according to a Deloitte survey cited by Collibra.

Context

Data access governance is the discipline of deciding who can reach which data, under what conditions, and with what controls applied to that data in use. Collibra’s article argues that the bottleneck is no longer just identity authentication or application access, but the last mile between approved users and governed data delivery.

That matters because analytics, AI initiatives and operational decisions increasingly depend on data that must be both fast to reach and tightly controlled. The article frames the issue as a mismatch between traditional IAM workflows and the need for granular data-level controls such as masking, policy enforcement and source-aware access provisioning.

The article also situates the problem in a broader compliance context, pointing to GDPR, NIS2 and the EU Data Act as examples of the regulatory pressure around secure and governed data sharing.

Key questions

Q: What breaks when access governance is still built around tickets and long-lived credentials?

A: Ticket-based and credential-heavy models break when work moves faster than human approval cycles. Engineers wait, security loses visibility, and permissions accumulate instead of expiring. In cloud and AI environments, that creates more standing access, weaker accountability, and a larger blast radius. Governance has to shift from slow approval to time-bound control and automatic revocation.

Q: Why does governance that stays at the documentation layer create risk for AI and analytics programs?

A: Documentation alone leaves a gap between policy intent and actual data use. When controls are not enforced in the workflow, teams either wait on slow manual reviews or work around governance altogether. That creates inconsistent masking, weak consent handling, and limited auditability, which increases compliance risk while making it harder to scale data-driven work safely.

Q: How do security teams know whether data masking is actually working?

A: Data masking is working when users can complete their task without seeing sensitive values they are not authorised to view. Teams should test whether the same policy behaves consistently across data sources, roles and consumption paths. If masking depends on manual exceptions or duplicate views, the control is not reliably enforced.

Q: Which teams are accountable for governed self-service data access?

A: Accountability usually spans data governance, IAM, platform engineering and the business data owner. The key is that access decisions must be governed centrally while being enforced where the data lives. If ownership is split but policy is not, the result is slower approvals, weaker audit trails and inconsistent access decisions across platforms.

How it works in practice

Why application-centric IAM struggles with data entitlements

Traditional IAM is optimised for granting access to applications, not for controlling specific datasets, rows, columns or masked views. That distinction matters because a valid identity session does not automatically mean the user should see the same data everywhere. In modern data estates, the control plane must decide not only whether a user is authenticated, but what subset of data they may see, how that access is enforced across multiple platforms, and whether the policy is applied consistently at the point of consumption.

Practical implication: move access decisions closer to the data model instead of relying on coarse application permissions.

How masking and central policy enforcement change the access model

Data masking is a governance control that lets an authorised user query a dataset without exposing every sensitive value. When paired with central policy enforcement, it reduces the need to create duplicate views for different audiences and helps avoid manual rework across platforms. The technical shift is from repeated, local exceptions to a single policy source that can apply database, schema or table-level rules and then render sensitive fields conditionally based on role or entitlement.

Practical implication: define masking and access rules once, then enforce them consistently wherever the data is consumed.

What changes when AI agents become data consumers

AI agents add pressure because they can act at machine speed and often need broad, repeatable access patterns to complete tasks. If governance still depends on ticket queues, spreadsheet tracking or manual approval loops, access delivery becomes the bottleneck rather than the control. The article’s central architectural point is that governed self-service has to work for both humans and AI agents, which means access provisioning, auditability and policy evaluation need to keep pace with automated consumption patterns.

Practical implication: design governed data access for both people and machine consumers, not just traditional end users.

NHI Mgmt Group analysis

Data access governance is becoming a data-plane problem, not just an identity problem. When the business value sits in the dataset itself, coarse application access is too blunt to express who should see which fields, rows or masked values. That shifts the governance burden from login-time authorisation to ongoing data entitlement enforcement. Practitioners should treat the data layer as part of the identity control surface.

The real control gap is not access approval, but access precision. Manual workflows can say yes or no, but they struggle to express conditional visibility, masking and platform-specific policy consistency at scale. Once access decisions must be applied across Snowflake, Databricks and BigQuery, the programme needs a way to keep one governance model intact while adapting enforcement to each source. The practitioner takeaway is to reduce exception handling, not just queue length.

AI governance inherits the weaknesses of data governance. If AI initiatives consume data through the same slow, fragmented access model used for people, then the governance bottleneck simply moves upstream into the AI pipeline. That is why this problem is not only about speed. It is about whether the organisation can prove that sensitive data stays constrained while still being usable for analytics and machine-assisted workflows.

Governed self-service will become the default expectation for both humans and AI agents. The market is moving toward access models that preserve oversight without forcing every request through a manual ticket chain. That will push IAM, IGA and data governance teams to align on a shared operating model for entitlements, masking and audit trails. Organisations that keep treating data access as an exception process will struggle to scale analytics and AI safely.

Data-level controls create the next governance boundary for sensitive data. The article points toward a named concept we can call the data access last mile: the point where approved identity meets governed data exposure. That is where policy enforcement, masking and source integration determine whether access is actually usable. Practitioners should evaluate this boundary as a first-class control layer rather than an operational afterthought.

What this signals

Data access governance is moving into the same operating layer as IAM and IGA. Teams can no longer treat data access as a separate helpdesk function if analytics and AI are going to scale safely. The practical question is how to preserve auditability while shifting entitlement decisions closer to the data itself.

The data access last mile is where modern governance programmes will be judged. If masking, entitlements and source-specific policy enforcement do not line up, the organisation will keep compensating with manual workarounds. That weakens both delivery speed and control consistency, which is exactly the trade-off practitioners need to remove.

AI adoption raises the bar for data governance rather than replacing it. Machine consumers do not reduce the need for policy. They make weak policy models fail faster, because access patterns become more frequent, more repeatable and harder to correct after the fact.

For practitioners

• Define the data access last mile Map where approval ends and data exposure begins across your main warehouses, schemas and sensitive tables. Identify the point at which identity decisions need to turn into row, column or masking controls, not just application permissions.
• Replace ticket-based access with governed self-service Measure how much access friction still depends on spreadsheets, manual approvals and repeated coordination across platforms. Move recurring, low-risk requests into policy-driven flows that preserve auditability without forcing every request through a human queue.
• Standardise masking policies across data platforms Use one policy model for sensitive fields so users see only what their role permits, even when the data is replicated across Snowflake, Databricks or BigQuery. The goal is consistent enforcement, not separate one-off views for each audience.
• Align AI consumption with data governance rules Review how analytics pipelines and AI workloads inherit access to sensitive datasets, including any service identities or delegated access paths they use. Make sure the same masking and entitlement rules apply before automated consumers start scaling usage.

Key takeaways

• Traditional IAM alone does not solve the problem of governed access to sensitive data across modern analytics and AI environments.
• The article’s core evidence is that data consumers still experience major availability delays, which shows that access friction remains a business issue as well as a governance issue.
• Practitioners need data-level controls such as masking, central policy enforcement and auditable self-service if they want fast access without losing oversight.

Key terms

• Data Access Governance: Data access governance is the practice of deciding who or what should reach specific data based on sensitivity, business purpose, and observed access paths. It combines classification, entitlement analysis, and review workflows so access decisions reflect exposure, not just permission status.
• Data Masking: Data masking obscures sensitive values so people or systems can work with the information without seeing the original content. In modern environments it is used as a runtime protection, not only as a database preparation technique, and it can be static, dynamic, or context-aware depending on the workflow.
• Controlled Self-Service: Controlled self-service is a delivery model where users can access approved tools quickly without waiting on manual IT support, while the organisation still controls data, permissions, and oversight. It reduces the incentive for shadow IT by making sanctioned options easier to use than unofficial alternatives.
• Data Access Last Mile: The data access last mile is the final step between an approved identity and actual data exposure. It is where policy enforcement, masking and source-specific permissions determine what the user or machine can truly see, making it a critical governance boundary for analytics and AI.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are building or maturing an IAM programme, it is worth exploring.