How can we help you?

Governance Shouldn’t Slow AI Down. It Should Be What Lets It Scale.

by Angus Palmer, ServiceNow Solution Architect at AC3

One of the biggest misconceptions in enterprise AI is that governance and innovation sit on opposite sides of the table.

In reality, the organisations that scale AI successfully tend to do the opposite. They treat governance not as a brake on innovation, but as the mechanism that allows AI to move beyond experimentation and into real business operations.

Five years from now, I suspect many organisations won’t be regretting the AI pilots that failed. They may regret the processes they never had the confidence to redesign.

A failed pilot can actually be useful. It might expose poor data quality. It might show that nobody really owns a process. It might highlight gaps in security controls. Or it might simply prove that the original use case was not particularly valuable.

That is still learning. Standing still teaches you very little.

Right now, a lot of enterprise AI adoption is still sitting in what I think of as the safe productivity layer:

  • “Summarise this meeting”.
  • “Draft this email”.
  • “Search this document”.
  • “Extract these notes”.

Those prompts are useful. I use them myself. But they are not the same as redesigning an end-to-end business process.

The bigger productivity shift happens when AI can help move work through an entire process, from the initial request through investigation, decision support, approval, action and finally evidence that the work was completed.

And that’s where platforms start to matter. AI needs context. It needs access to workflows. It needs identity and access controls. It needs to understand what it’s allowed to do right now, not what somebody decided it could do six months ago. And whatever it does needs to leave an audit trail.

That is why I think the conversation is gradually shifting away from simply asking, “Which model should we use?” and towards a much broader question:

How do we orchestrate AI safely across the enterprise?

This is where security joins the conversation.

I don’t think we should design AI systems around the assumption that we can create a model that can never be fooled. Prompt injection is OWASP’s highest-ranked security risk for large language models and is inherent in how these models operate.

A better approach is to design the environment around the model so that, when something eventually does go wrong, the blast radius is limited. This layer of software around the model that permits it to take action is the Agentic AI Harness. It is at the harness that we can control what the agent can access. Control what it can change. Control which actions need approval.

In other words, don’t rely on the AI always behaving perfectly. Build the controls around it.

This is where ServiceNow becomes interesting. ServiceNow's AI Platform brings together data, AI, workflows, automations and security so that AI can move beyond recommendations and start taking action across enterprise workflows.

AI Agents can perform work within the guardrails of your existing workflows, leverage deterministic automations for exact outcomes and access data under the same security controls applied to human users. Veza can map and manage permissions across human and AI identities and AI Control Tower provides a way to manage, monitor and optimise AI. Together, these capabilities govern AI not only on the ServiceNow platform, but across the enterprise.

For me, that lines up closely with the problem Australian organisations need to solve. We don’t necessarily need more AI experimentation. We need better ways to turn experimentation into governed execution.

An organisation needs to be able to answer some fairly basic questions.

  • What AI capabilities are operating in the organisation?
  • Who owns them?
  • What data are they using?
  • What can they access?
  • What actions can they take?
  • How are they being monitored?

And if something goes wrong, how quickly can we stop them?

One idea I keep coming back to is this:

The first AI governance control is not a policy. It’s an inventory.

If you can’t see the AI agents, models, prompts, datasets, integrations and non-human identities operating in your environment, it’s very difficult to govern them properly. You can’t govern what you don’t know exists.

Innovation and governance are not opposites

I think we’ve created a false choice between innovation and governance. In practice, governance is what allows innovation to scale. Without governance, AI tends to stay trapped in pilots, proofs of concept and low-risk productivity use cases. With the right controls in place, organisations can gradually increase what they’re prepared to let AI do.

I particularly like the idea that autonomy should be earned.

A read-only assistant answering questions from approved knowledge articles presents a very different risk from an agent that can update records, trigger fulfilment, approve an exception or communicate with a customer. As autonomy and potential impact increase, the controls around identity, access, monitoring, approval and shutdown should increase with them. That’s the practical architecture conversation we should be having.

How do we let organisations start small without allowing AI to start invisibly?

How do we give teams enough freedom to experiment while still connecting AI to ownership, policy, risk and operational control?

And how do we move away from governance that happens once a year towards governance that can operate at the speed of AI?

The real issue is organisational confidence

AI is not creating all of these governance problems. In many cases, it is exposing problems that were already there.

  • Poor data quality was already there.
  • Unclear process ownership was already there.
  • Excessive access was already there.
  • Fragmented systems and weak operational controls were already there.

AI simply makes those weaknesses much harder to ignore.

So when organisations hesitate to use AI more broadly, I don’t think it’s always because they are afraid of AI itself. Often, they’re not confident in the organisational machinery around it.

  • They’re not confident the data is good enough.
  • They’re not confident the process is well defined.
  • They’re not confident the access controls are tight enough.

Or they’re not confident that everybody understands who is accountable when something goes wrong.

That is why I think Australia's AI opportunity is not simply to copy the loudest parts of the US market. We can take a more trusted approach to adoption: practical, governed, workflow-led and focused on outcomes.

But we still need to move. Caution only becomes an advantage if it leads to better execution. If it simply creates slower decision-making, it is not really caution anymore. It’s inertia.

Where I would start

For organisations trying to make this practical, I would start by asking a handful of fairly simple questions:

  • What AI capabilities are already being used across the organisation, including shadow AI?
  • Who owns each use case, agent, model, dataset and integration?
  • What data can each AI capability access, and should it have that access?
  • What actions can it take?
  • Where is human approval still required?
  • How are outputs, decisions, costs, performance and incidents being monitored?
  • If something goes wrong, what is the process for quickly pausing or disabling that capability?

These are not abstract governance questions. They’re the foundations for scaling AI safely. And they become much easier to answer when AI is connected to the same platform that already understands work, risk, assets, identity, service delivery and operational evidence.

Final thought

The AI race may not be won by the organisations running the boldest pilots.

It may be won by the organisations that figure out how to balance ambition with control and then turn experimentation into something they can trust enough to use every day.