Skip to main content

When AI moves from answering to acting: Getting governance right

AI agents bring greater autonomy to the enterprise, making clear rules about access, actions and accountability increasingly important

16 November 2026

There’s a big difference between getting an answer wrong and carrying out an action unchecked. In the early stages of AI, a colleague of mine asked a voice bot we had developed to suggest an itinerary for a trip to France. Instead, it supplied one for India. It was an obvious hallucination and could be verified easily.

But, if AI could get something so simple wrong, imagine what could happen when an autonomous AI agent makes the wrong decision and runs with it. The consequences can be far more serious, affecting your organization, your customers and your reputation.

As organizations use AI more, they are moving away from what I call “vanilla” AI, which simply answers questions, toward AI that can make decisions, act and carry out entire business processes. That’s why governing your agentic AI has become so important.

Governance needs to catch up with what AI can do now

The governance you have in place right now will remain relevant. You will still need to manage the quality and accuracy of your AI, protect your data, and consider privacy, security and bias. Agentic AI doesn’t change any of that. It simply gives you more to govern.

Traditional AI produces an output, and a person decides what to do with it. Agentic AI can take that output, go on to use other tools and systems, communicate with people or initiate the next step in a process — all without any human intervention at all. This means we must now question the output as well as the activity throughout the workflow.

4583774246_Blog_design_AI_governance_v3

Traditional AI stops at the output, leaving a person to decide what happens from there

Although this model works well when a person makes the final call, using it to govern agentic AI leaves you exposed to risks it was not designed to manage.

This is where I believe some organizations are falling behind.

The excitement around new technology, especially agentic AI, is understandable. But letting that excitement get ahead of the controls, on the assumption that you can put them in place later, is a mistake I see far too often. Because with AI agents, by the time you discover a problem, the agent may already have acted on it.

Organizations without effective AI governance can face compliance failures and operational problems. When AI acts without guardrails, those consequences are amplified.

When an answer becomes an action

Once AI has the freedom to act, there are three things I believe every organization should think about.

1. Risk moves into your day-to-day operations

This gives you a far greater blast radius. AI risk is no longer confined to the information an agent produces. It now extends to the running of your business.

2. There are more ways to influence what AI does

AI agents don’t work in isolation. They can take information from emails, documents and websites, and interact with different tools, systems and even other agents. Any of these can influence how an agent behaves. This makes controlling what it can access — and what it can do with that access — incredibly important.

3. You need to keep watching while the agent is working

Testing an agent before it starts working in your business is essential, but you can’t stop there. The information it receives, and the environment in which it operates, will continue to change. You need to know how it is behaving, what it is doing and when something needs your attention.

The bottom line is, the more responsibility you give an agent, the more carefully you need to define the boundaries.

Your digital workers need rules, too

I like to think of AI agents as digital employees. An agent carrying out one small, contained task is a little like hiring a freelancer for a specific piece of work. But once that agent becomes part of your day-to-day operations, with ongoing responsibility and access to your systems and information, the relationship changes. You expect an employee to work within company policies and understand the limits of their authority. Your agents need rules, too.

Every agent should have a clear business purpose and a named owner. It needs its own identity and access only to the information and tools required for its job.

Then there is the question of autonomy. Some actions can happen automatically. Others require approval. An action that is difficult or impossible to reverse needs a different level of control than a simple day-to-day task. And when an agent starts working with other agents or systems, you need to know exactly where it is going and where it stops.

Everything must be visible and traceable. If something goes wrong, you should be able to see what information the agent was working with, which tools it used and what approvals it received.

4583774246_Blog_design_AI_governance_v3

Agentic governance spans identity, context, tools, actions, monitoring and human accountability

Pay attention to your data

Data has always been fundamental to AI. With agentic AI, there is more riding on getting it right. An agent can use the information you give it to decide and act. It makes the quality, security and governance of that data even more important.

The same applies to the information an agent remembers as it completes a task. You need to know what it can retain, how long it keeps that information and whether anything sensitive could be exposed.

It comes down to something quite simple: If you want your AI agents to make good decisions, you need to give them good data to work with from the outset.

Governance doesn’t stop when the agent starts

Governance needs to follow an AI agent throughout its working life. Before you build, be clear on the agent’s purpose, who owns it, the data it will use and how much freedom it should have. Before it goes live, check its permissions, testing and approvals and make sure you can reverse or stop an action if something goes wrong. Once it is working, monitor how well it performs, the decisions and actions it takes, how much it costs to run and whether it delivers the results you expect.

Asking some simple questions can tell you a lot about where your governance stands. Can you see every agent operating in your business, who owns it, and what it can access and do? Can you trace what happened if something goes wrong? And can you step in quickly?

If the answer to any of these questions is no, start with visibility. Know which agents you have, what they are doing and who is responsible for them. That will give you a practical starting point to tighten the controls.

That kind of oversight is becoming a hallmark of organizations that are further ahead with AI. Our 2026 Global AI Report: A Playbook for AI Leaders found that 55.9% of AI leaders use a centralized governance model, compared with 37.6% of other organizations and 33.3% of AI laggards.

Start with the end in mind

We’ve learned many of these lessons by being our own client zero. We experiment with AI internally, encounter security and compliance issues ourselves, and find out what can go wrong before taking that experience to our clients.

One of the most important lessons we’ve learned is to start with the end in mind. Be clear about what you want an AI agent to achieve, how you will measure its value and what you’re prepared to let it do to get there. Then build the governance around that from the start. Trying to add the right controls once an agent is already operating across your business is far more difficult.

By applying governance to your AI agents, you’re not limiting their freedom. You’re putting the right rules in place for these agents to operate as intended and for you to get the most out of them.

What to do next

Learn more about NTT DATA’s Agentic AI Studio and how we can help you build, govern and manage AI agents across their lifecycle.

Author

Professional headshot of a bearded man in a light gray blazer and white shirt, seated indoors.

Ashish Mishra

Lead Solutions Architect: Agentic AI Services at NTT DATA

Ashish Mishra is Lead Solutions Architect: Agentic AI Services at NTT DATA, helping organizations move beyond AI experimentation to build practical, production-ready GenAI and agentic AI solutions. He advises clients on AI architecture, governance, technology choices and adoption strategies across the Microsoft ecosystem. This enables them to scale AI with confidence and deliver meaningful business outcomes.