Dr Victoria Holt: life, the universe and everything

Thursday, 4 June 2026

The Reality of Compliance in the Age of AI

Compliance used to be retrospective. Policies were written, audits were conducted and evidence was gathered after the fact to demonstrate that controls had been followed. That approach is no longer sufficient. AI has introduced a level of complexity where decisions are made faster, data is reused in ways that are difficult to track, and accountability becomes harder to define. Compliance cannot keep up if it remains a reactive process. It has to become something that is designed into how organizations operate.

The problem beneath the surface

Most organizations still treat compliance as a separate function. A team that interprets regulation. A set of policies that sit alongside operations. A series of controls that are checked periodically but the real challenge is not understanding regulation. It is applying it consistently across processes, systems, and increasingly, AI-driven outcomes.

What data can be used for training
How decisions are explained
Where sensitive information is retained or deleted

These are operational questions, not just compliance ones.

Where Purview comes in

Purview Compliance capabilities focus on managing these challenges in a structured way. Data lifecycle management defines how long data should exist and when it should be removed. Records management strengthens that by applying legal and regulatory context. Compliance Manager provides a framework to track controls and measure progress against requirements. More recently, these capabilities are being used to address AI-related concerns. Understanding data usage, managing retention, and demonstrating control are all foundational to responsible AI. The technology does not replace compliance thinking. It enables it to be applied consistently.

The technical layer that matters

Retention labels and policies are often seen as administrative tools. In reality, they directly influence how data is stored, preserved, or deleted across workloads. Records management introduces immutability and defensibility. Compliance Manager maps controls to regulatory standards, providing visibility into gaps and progress. These are not isolated features. They form a system where compliance is codified into policies that operate at scale.

Why this matters now

Regulation is evolving. The EU AI Act, data protection laws, and industry-specific requirements are all pushing organizations towards greater accountability. At the same time, AI is accelerating how data is used. This creates a tension. Organizations want to move quickly, but also need to demonstrate control. Manual processes cannot bridge that gap. Compliance has to become embedded. It has to operate continuously, not periodically.

The reality

In the age of AI, compliance is no longer about proving that controls exist. It is about proving that they are applied, monitored, and effective in a constantly changing environment. Purview provides the mechanisms to do this, but like governance and security, it depends on how it is used. Policies must reflect real business requirements. Controls must be implemented consistently. Ownership must be clear. Otherwise, compliance remains a reporting exercise rather than a capability.

References and learning

https://learn.microsoft.com/en-us/purview/compliance
https://learn.microsoft.com/en-us/purview/data-lifecycle-management-overview

Microsoft Announces Scout: An Always‑On Autonomous Agent for Work

Microsoft has unveiled Scout, its first Autopilot agent, an always‑on, autonomous digital assistant designed to work across Microsoft 365, proactively coordinating tasks, managing workflows, and keeping work moving even when you’re not in the loop. What makes Scout different is its ability to operate with its own identity, act within organisational policies, and build long‑term context through WorkIQ, learning how you work and what matters most. But beneath the excitement, there’s a deeper story for those of us working in data governance and Responsible AI.

Where Scout Meets Data Governance

Microsoft has been explicit: Scout is built with enterprise‑grade security, policy enforcement, and auditability from day one. Key governance‑aligned capabilities include: Policy‑constrained identity Scout acts only within the permissions and boundaries your organisation sets. Execution containers & OS‑level sandboxing reducing risk when agents access files, run code, or interact with networks. Continuous policy conformance checks every action is validated against organisational guidelines, producing an audit trail. This is a significant shift: AI agents are no longer “black boxes” running in user sessions, they’re governed, monitored, and contained as first‑class enterprise actors.

Responsible AI: Built Into the Foundation

Microsoft has also published Responsible AI documentation for Scout, reinforcing that it is part of a broader commitment to safe, transparent, and accountable AI systems.

Highlights include:

Responsible AI FAQs explaining how Scout works, what data it accesses, and how system owners can shape behaviour.
Tiered permission systems for file access, shell commands, and browser automation.
Human‑in‑the‑loop expectations and environmental considerations for deployment.

This aligns Scout with Microsoft’s AI principles of fairness, reliability, safety, privacy, security, inclusiveness, transparency, and accountability.

Why This Matters

For organisations already investing in data governance, AI assurance, and operational Responsible AI, Scout represents a new category of enterprise agent:

Autonomous enough to reduce coordination overhead
Governed enough to meet compliance and risk expectations
Context‑aware enough to become a durable part of the digital workforce

This is the moment where AI agents stop being assistants and start becoming accountable digital colleagues operating within the same governance frameworks as humans and systems.

Wednesday, 3 June 2026

Microsoft Build 2026: The Moment Governance Became the Bottleneck, Not Innovation

If last year’s narrative was about what AI can do, Microsoft Build 2026 marked a noticeable shift: the conversation has moved firmly to what organizations must control.

Across two days of announcements, Microsoft made one thing clear. The next phase of enterprise AI will not be defined by better models or more copilots. It will be defined by whether organizations can operationalise data readiness, governance, and trust at scale.

And that is where the most important announcements sit.

From “AI Features” to “AI Systems That Act”

The headline innovation at Build wasn’t just new models, it was the emergence of autonomous AI agents as first-class enterprise actors.

Microsoft introduced Scout, an always-on AI agent capable of continuously operating across enterprise systems, taking actions rather than waiting for prompts.
This marks a fundamental shift from assistive AI to operational AI, software that executes tasks, interacts with systems, and makes decisions within workflows.

But this also introduces a new governance reality.

When AI moves from generating content to acting on behalf of a business, the questions change:

Who is accountable for the action?
What data did the agent access?
What policies constrained its behaviour?

Microsoft’s answer is not a single tool but an emerging governance architecture for agents.

Governance Is Now Part of the Platform (Not an Add-On)

Across the announcements, governance was not positioned as a compliance afterthought. It was embedded into the core platform.

Three developments stand out.

Agent identity, control, and auditability

Agents are now designed with their own identities, permissions, and audit trails that essentially are becoming governed entities within enterprise systems.
This is a critical shift: governance is no longer about users accessing data, but about non-human actors operating within policy boundaries.

The rise of the agent control plane

With capabilities such as Agent 365 and broader governance frameworks, Microsoft is building what can only be described as a control layer for AI agents covering access control, visibility, monitoring, and compliance.

This moves governance from static policies to continuous oversight of autonomous systems.

Built-in safety, evaluation, and testing

The introduction of evaluation frameworks like ASSERT (for testing AI behaviour against policy expectations) signals a shift toward engineering governance into the development lifecycle itself.

This aligns closely with emerging standards (ISO/IEC 42001, EU AI Act), where governance is expected to be designed, evidenced, tested and not assumed.

Data Governance Quietly Took Centre Stage

While the headlines focused on models and agents, the more important story sits underneath: data is now the limiting factor for AI.

Microsoft’s investment in Fabric including a GPU accelerated data warehouse positioned as an execution layer for AI workloads reflects a deeper truth: organisations don’t lack AI capability, they lack AI-ready data environments.

This reinforces a theme many of us have been seeing on the ground:

The challenge is no longer can we use AI?
It is can we trust the data, control its usage, and scale it responsibly?

Even outside the keynote announcements, updates across Microsoft Purview continue to evolve around:

data quality management,
data loss prevention for AI interactions,
and governance across expanding AI estates.

Taken together, this signals a more mature positioning that data governance is not supporting AI, it is enabling it.

A New Stack: AI, Data, and Governance as One System

Perhaps the most important architectural shift is how Microsoft is framing the AI stack.

At Build 2026, governance was explicitly treated as a foundational layer alongside compute, models, and tools.

This is subtle but significant.

Previously, governance sat outside the stack:

something imposed after deployment,
owned by risk or compliance functions,
often disconnected from engineering.

Now, governance is:

integrated into runtime environments,
embedded in agent frameworks,
and enforced through platform capabilities.

This is a move toward operational governance, not theoretical governance.

What This Means for Businesses

For organizations, these announcements are less about new features and more about a change in expectations.

AI adoption will be constrained by governance maturity

The organizations that succeed will not necessarily be those with the most advanced models but those with:

clear data ownership,
defined policies for AI usage,
and the ability to monitor and control AI behaviour continuously.

Governance becomes a cross-functional discipline

AI governance can no longer sit solely with data teams or compliance functions. It now spans:

data governance,
security,
enterprise architecture,
and operational risk.

Tools alone will not solve the problem

While Microsoft is building an increasingly comprehensive governance ecosystem, the platform assumes something critical:

Organisations already understand their data, risks, and policies.

In reality, many do not.

This is where the gap and the opportunity sits.

The Real Announcement wasn’t a Product

If you step back, the most important announcement at Build 2026 wasn’t a model, a Copilot update, or even an agent.

It was a shift in narrative.

Microsoft is signaling that:

AI is no longer experimental.
Agents will become embedded in everyday business operations.
And governance is now the primary barrier to scale.

In other words, we’ve moved from the innovation phase of AI to the industrialisation phase.

And industrialisation always introduces the same question:

How do you scale safely, consistently, and with accountability?

That is not a tooling question. It is a Data and AI governance question.

References

forbes.com dqindia.co theneuron.ai microsoft.github.io pulse2.com

forbes.com learn.microsoft.com theneuron.ai

Tuesday, 2 June 2026

The Reality of Data Security in M365 (Purview Protection)

Data security has traditionally been viewed as a problem of access. Who can see what. Who can download it. Who can share it. In a world of structured systems and defined boundaries, that was enough. That world no longer exists. Data in M365 is fluid. It moves between emails, Teams chats, SharePoint sites, and endpoints. It is copied, embedded, summarised, and now increasingly, generated by AI. The idea that you can secure it by controlling entry points alone is no longer realistic. Data security has shifted from protecting locations to protecting the data itself.

The problem organizations are actually facing

Most organizations believe they have data security in place because they have policies but when you look closer, those policies are often disconnected from how data is used in practice. Labels are defined but not applied consistently. Data loss prevention rules exist but generate noise rather than insight. Users find workarounds because controls are either too restrictive or not aligned to real workflows. The result is a false sense of security. Controls exist, but coverage is inconsistent. Risks are identified, but not prioritised. Sensitive data continues to move, often without visibility.

What Purview is doing differently

Purview Protection capabilities bring together several controls that are often treated separately. Information Protection classifies and labels data, ideally at creation. Data Loss Prevention applies policies to control how that data moves. Insider Risk adds behavioural context, identifying patterns that indicate potential misuse or compromise. Individually, these are familiar concepts. Together, they form a model where protection is persistent and data-centric. Classification stays with the data. Policies follow it across services. Signals from usage and behaviour start to inform risk in real time. It is not just about blocking actions. It is about understanding how data is used and where risk actually exists.

The technical reality that matters

There is a level of technical depth that is often overlooked. Sensitivity labels are not just tags. They drive encryption, access control, and downstream policy enforcement. DLP is not just rule matching. It combines classification, conditions, and contextual signals. Insider Risk does not operate in isolation. It correlates activity across multiple workloads. These capabilities rely on integration. They rely on consistency. They rely on governance decisions being made upfront. Without that, the tooling becomes fragmented. With it, you start to see a unified security posture that is driven by data, not systems.