From cloud-ready to agent-ready: Azure and agentic AI: PwC

Greg Inks

Principal, PwC US

Email

Key takeaways:

Agentic AI raises the cost of "pets wearing cattle tags": undocumented exceptions, manual configs, and tribal knowledge that were tolerable with humans now cause agents to fail or act unpredictably.
Agent-ready infrastructure means environments provisioned, governed, and retired through code, with identity, observability, and cost controls built in from the start.
A tiered autonomy model maps agent actions to risk and should be encoded in policy before an agent ever runs, not decided in the moment.
The "beyond VMs" cloud-native argument still holds, but agentic AI adds a new requirement: the environment must be legible to a machine.
Getting to agent-ready is concrete work defined in advance rather than discovered mid-incident.

Years ago, I wrote a piece suggesting that enterprises should stop treating infrastructure like pets and start treating it like cattle (it was not my metaphor!), stop naming servers, stop nursing them through outages, stop mourning them when they pass away, and start building systems designed to lose members of the herd without blinking.

It was a good metaphor — and, for the most part, an ignored one. For years, that gap was costly. Agentic AI raises the stakes. That foundation increasingly shapes how much operational work an enterprise can delegate to AI within defined controls.

Agentic systems depend on environments consistent enough to operate without undocumented human intervention.

Here's what "not agent-ready" looks like on the ground today. Look across many enterprise Azure environments, and you'll still find pets wearing cattle tags. Production VMs with hostnames like ‘SQLPROD03’ that teams hesitate to touch on a Friday. One-off configurations that live mostly in institutional memory because “we know how it works.” Infrastructure-as-Code repos that generate infrastructure once, then get abandoned the moment a human logs in through the portal and "just fixes something real quick." And it isn't only compute, the Entra role assignment that outlived the need for it, the Azure Policy exemption with a hand-typed justification, the network dependency that lives in one engineer’s head, the approval that happens in a Teams chat instead of an Azure DevOps pipeline gate. We adopted the vocabulary of cattle and kept the emotional attachment — and the undocumented judgment calls, of pets.

Call it agent-ready infrastructure: environments provisioned, governed, and retired through code, with identity, observability, and cost controls treated as part of the infrastructure – not afterthoughts. The cloud value conversation is moving beyond cost optimization toward a governance foundation for new operating capabilities. That’s why the old metaphor suddenly has teeth.

Agentic AI raises the cost of standing still. Agentic systems, Copilot agents, Microsoft Foundry workflows, and autonomous pipelines landing across the Azure stack, don't have the context to babysit a pet. An agent provisioning compute doesn't know that ‘SQLPROD03’ has a personality, or that the network team "always does it this way." It reads policy, reads available capacity, and acts. If your environment depends on undocumented exceptions, an agent can fail against them, or act in ways the organization didn’t intend.

From infrastructure problem to platform governance model

Here’s the shift the original piece couldn’t have anticipated: infrastructure-as-cattle used to be an operations conversation, automation, orchestration, elastic capacity. Agentic AI turns it into a governance conversation. That requires an operating model, not just a principle. Infrastructure, policy, governance, and automation are increasingly converging into a single operating model, and agent-ready infrastructure is one way that model takes shape.

When a human provisions a VM, the barn and the fence are processed: a ticket, an approval, a naming convention somebody half-follows. When an AI agent provisions, scales, or retires resources on your behalf, inside Azure Landing Zones, through Infrastructure-as-Code pipelines, via Azure Policy and Azure AI Foundry guardrails, the barn and fence should be encoded in policy and controls, not left to memory and process alone. The agent acts through an approval gate in Azure DevOps or a secure GitHub Actions environment, with audit logging and time-bound access controls applied as appropriate.

In practice, that looks like three tiers, mapped to risk intensity and controls rather than task type:

Pre-authorized autonomy. Scaling within pre-approved capacity bounds, restarting a failed service, remediating a known-good drift from an Azure Policy baseline. The action, its boundaries, and its failure conditions were authorized in advance — the guardrail is the approval, already granted.
Supervised. Provisioning new resource types, modifying network topology, touching anything in scope for compliance. The agent acts, but through an approval gate in Azure DevOps or GitHub Advanced Security, with Microsoft Purview logging the decision trail and Microsoft Entra Privileged Identity Management time-bounding the access used to do it.
Approval-gated. Anything touching identity boundaries, production data planes, or regulated workloads. A human approves before the action executes, not after, when the incident report gets written.

Skip the tiering, and you'll have a stampede on your hands: ungoverned agents spinning up resources at machine speed, with nobody watching the gate.

The autonomy tier should be encoded in policy before the agent runs, not decided in the moment. And that tiering requires context. The same action may be routine on an ordinary day and materially different mid-incident, during a change freeze or in a peak-traffic window. Mature policy accounts for the operating context an action lands in, not just the action itself. The principles are reusable; mapping them to a specific client's risk appetite, architecture, and regulatory requirements is the design work that matters. The governance model can be encoded in Azure Policy and Entra Conditional Access, and extended as the agent population grows. In practice, agent-ready infrastructure isn’t “no humans.” It’s an operating model that places human judgment where the organization’s risk posture calls for it.

This isn't hypothetical. This direction is already visible in Microsoft’s Azure SRE Agent². Now generally available, it can investigate production issues, propose mitigations, and automate responses within guardrails and approval workflows defined by the organization. It’s a practical example of the broader shift from automation toward configurable autonomy. It’s also a useful signal of the broader direction of Azure operations: agents working within defined controls and configurable levels of autonomy.

Extending the "beyond VMs" argument, not contradicting it

Colleagues across our cloud practice have made the case, rightly, that lift-and-shift VM migrations cap the value of a cloud transformation, and that cloud-native, managed-services patterns can unlock more automation and operational flexibility than a VM-by-default approach.

VMs aren’t disappearing from enterprise estates. They remain appropriate for OS-level dependencies, certain compliance-bound workloads and systems that aren’t cloud-native-ready yet. The argument isn't VM vs. no-VM. It's pet vs. cattle, and it applies across the estate, not just compute: the VM, the container, the serverless function, the identity, the network path, and the policy that governs all of them together. A cloud-native workload that was migrated by hand, has undocumented exceptions, and gets modified by whoever's on call is still a pet, it just happens to be a fashionable one.

What agentic AI adds to the “beyond VMs” argument is a new operating requirement. Cloud-native and consumption-based architecture have long enabled greater automation and elasticity. Agentic operations add another consideration: whether the environment is structured so AI can act within defined policy, identity and approval boundaries. A policy-governed environment is more legible to a machine than one built around undocumented exceptions and hand-maintained dependencies.

From cloud-ready to agent-ready: what the operating model requires

If you lead technology for an enterprise right now, the question isn’t simply, “Are we in the cloud?” It’s this: could an agent operate in your Azure environment within clear policy, identity and approval boundaries, without undocumented human cleanup?

For many enterprises, the harder constraint may not be agent capability. It’s the environment around the agent. Getting to agent-ready is concrete work:

Infrastructure-as-Code coverage across the relevant estate, including exceptions that are still handled manually.
Agent identity and authorization, scoped and time-bound through Microsoft Entra, with agent actions attributable to specific, revocable identities rather than broadly shared credentials.
Policy-as-code spanning compute, identity, network, and security together in Azure Policy and Microsoft Entra, not just VM guardrails while identity exceptions and config drift stay informal.
Observability and auditability designed to help reconstruct what an agent did, what it accessed and why it acted.
Defined rollback and reversibility paths for autonomous actions, appropriate to their risk and impact.
FinOps and cost controls that cap what autonomous scaling can spend, not just what it's technically permitted to touch.
A tiered autonomy model, defined in advance and enforced through Entra PIM, Azure DevOps approval gates, and Purview audit trails, not discovered during an incident.

The original cattle-versus-pets argument was about resilience. Build systems that can withstand failure and recover automatically.

Agent-ready infrastructure raises the stakes. The question is whether autonomous systems can operate within defined controls without requiring constant human oversight.

That's the real shift AI introduces for technology teams. Infrastructure is becoming part of how the business puts AI to work.

Enterprises that make that shift can move beyond automation toward more governed autonomy. And the limiting factor may not be the capability of the agent. It may be the environment the agent is asked to operate in and the business will be the first to notice that something is amiss.

The value isn’t another assessment. It’s creating the conditions for more routine operational work to be delegated to agents within defined guardrails, freeing experienced engineers to focus on higher-value work. Getting there takes architecture, engineering and governance tailored to the enterprise’s existing estate and risk posture.

Where you start, a scoped readiness assessment, a landing-zone modernization or an extension of an autonomy model already in place, depends on your environment. The objective is an Azure estate designed for AI agents to operate within clear policy, access and human-oversight boundaries.

Here's the stake that makes this more than an IT conversation. PwC research puts the broader reinvention challenge in stark terms: 45% of CEOs surveyed said they did not believe their companies would remain viable in ten years if they continued on their current path¹. The infrastructure conversation sits inside that bigger challenge. AI isn’t an isolated technology shift. It is changing how enterprises operate, where people spend their time and how quickly businesses can turn insight into action.

The opportunity isn’t autonomy for its own sake. It’s an operating foundation that allows AI to take on more repeatable work within defined controls, so you can focus experienced people on the business problems that matter most.

That takes architecture built for agentic operations, governance encoded into the environment and human judgment deliberately placed where it matters. That’s the move from cloud-ready to agent-ready.

Stop naming your servers. Your agents were not going to remember the names anyway.

^{[1]: Matthew Duffey, Venkatesh Jayaraman, and Veronique Roos-Emonds, "From stagnation to innovation: Make business model reinvention real," PwC, May 21, 2024, https://www.pwc.com/gx/en/issues/c-suite-insights/make-business-model-reinvention-real.html. Stats cited from the article and PwC's Global CEO Survey.}

^{[2]: "Azure SRE Agent," Microsoft Azure, accessed August 2026, https://azure.microsoft.com/en-us/products/sre-agent.}