In July, a group of OpenAI agents broke through testing guardrails and hacked the open-source AI company Hugging Face. The incident exposed two distinct problems: the guardrails failed to contain the agents, and the agents pursued harmful actions that should never have occurred ('misalignment'). Recent research has also shown the tenacity and ingenuity with which the agents pulled off the hack.
Incidents like this are often framed as evidence that AI models are growing too capable to trust. But capability alone doesn’t explain what went wrong. The outcome depended on both the model’s behavior and the controls around it: what it could access, where it could act, and whether a human had to approve consequential steps. That same distinction matters in IT support, where agentic AI can create real value only when its authority is deliberately constrained.
For Markus Gremmelmaier, one of our Senior Product Managers working on AI, the Hugging Face hack “shows how capable the systems are. But the real question is whether the system and the boundaries around it were designed properly.”
In IT support, agentic AI becomes more useful when it can interact with systems and produce real effects, but the consequences of weak boundaries or misaligned behavior also increase. As a result, rather than restricting AI, the goal is to build a control framework that contains what the agent can do, aligns its actions with the user’s intent, and still leaves enough flexibility to deliver meaningful value.
That control framework has several layers. It starts with administrator control over whether the capability is enabled, who can use it, and where it can act. It then separates advice and diagnosis from system-changing resolution, requires meaningful human approval for consequential actions, and keeps decisions and outcomes visible and traceable. These AI-specific safeguards sit within the access, security, compliance, and governance controls IT teams already trust. Finally, IT professionals need the skills and confidence to supervise AI effectively as their role evolves.
Start with administrator control
Sustainable trust in AI starts with control, and that control begins long before anyone uses an AI tool or agent.
First things first: Should this capability be enabled at all? Consider the short and long-term consequences of integrating it into your workflows, and whether the benefits outweigh the possible downsides.
Answer this question in the affirmative, and you can move onto more granular questions. Who can use the capability? Which devices or environments can it be used on? Which level of capability does each person receive? Is advanced functionality enabled by default?
The administrator must first define the parameters of an agent’s remit. The approach will vary by business: a smaller IT team may have one administrator making these decisions, while an enterprise may need role-based access and different policies across teams or device groups.
But the underlying principle remains the same: authority and access should be assigned intentionally by human admins. Or, as Markus put it: “Admins decide whether they want to use these capabilities, who can use them, and on which devices. Admins are in control of deploying the functionality.”
Separate advice, diagnosis, and resolution
In IT support, different levels of AI capability call for different levels of access. As Markus put it: “Sometimes experts just want advice. Sometimes they only want diagnosis. And sometimes they want end-to-end troubleshooting.”
Secure AI use requires differentiated access. Not everyone has the same expertise or authority, so each person should receive only the capabilities needed for their role. This applies the principle of least privilege (PoLP): access is limited to what is necessary for the task.
For junior employees, it might be enough for them to use AI agents for advice. This will allow them to access possible causes and troubleshooting steps, tapping into the experience of more senior employees. A low risk starting point, this mode keeps control firmly in human hands.
More experienced employees can access more complex information to diagnose a problem; for example, endpoint context and diagnostic capabilities to support root-cause investigation and validate findings. After assessing this information, they decide what will happen next.
The highest level of functionality – that is, permitting AI to resolve issues independently and make system-altering changes – should be reserved for the most experienced employees. This demands stronger control and appropriate human approval. The user needs to be experienced enough to understand the change that is about to happen.
Instead of treating AI adoption as an all-or-nothing decision, think of authority as a spectrum that can be adjusted to the user’s role and expertise. Assigning different levels of access creates a differentiated and more secure model without preventing experienced users from taking advantage of advanced capabilities.
Serious about AI security
Learn how TeamViewer builds security and control into every AI workflow, so you can adopt AI capabilities with confidence
Ask for human approval when it matters
Our recent global research study, Path to the Autonomous Workplace, found that three quarters of respondents use AI daily, but 61% still prefer human oversight before AI takes action. Only one in 20 (5%) have no concerns about AI operating without a so-called ‘human in the loop’.
To build trust and foster adoption, we need to ensure that human oversight plays a key role in AI-powered workflows. This should not mean interrupting users after routine steps. Rather, actions should be classified according to their potential effect. The greater the possible effect, the greater the need for human judgement.
AI should pause for human approval before taking consequential action; but, by the same token, the person giving approval should be given all the information they need to make an informed decision. The choice to reject a proposed action or remediation should be a real one.
Putting a human in the loop also creates a sense of accountability. It means that the decision to automate always lies with an experienced professional, who can reliably discern the level of oversight needed.
Human oversight is not measured by how often the AI asks for permission. It depends on whether the right people have real control when the decision carries consequences. Or, as Andrew Hewitt (Vice President, Strategic Technology) put it in a recent article: “They want confidence that they can intervene when it matters.”
Make actions visible and traceable
Black boxes do not inspire trust because they require blind faith. But we all already know that AI can make mistakes and requires oversight. With agentic AI, that concern becomes more immediate because the system can take action in real environments. As a result, users need visibility into what it's doing, why it's doing it, and what effect its actions have.
For this reason, any actions taken by AI need to be fully visible and traceable: before, during, and indeed after the action has taken place.
Before an action, the user should understand what problem the AI identified, what it proposes to do, whether the action will change the system, and the projected outcome. During execution, the user should be able to see what is happening and confirm that it matches the approved plan.
Whatever the AI said it was going to do, should correspond exactly with what happens during the action. Then, after the action you should be able to review what was approved, who approved it; what action occurred, whether it succeeded, and what happened if it failed.
When it comes to agentic AI, there shouldn’t be any surprises. For control to be really meaningful, you need to have all necessary information at hand. This helps investigate failures, demonstrate accountability, and improve policies. Instead of relying on blind faith, you can feel reassured that AI is doing what it should – no more and no less.
Fit AI into the controls IT already trusts
Integrating agentic AI into your support operations shouldn’t demand a complete rethinking of your security and privacy stance. It should instead plug into and build on your preexisting controls.
Agentic AI should fit within the access, security, compliance, and governance controls IT teams already trust. Existing policies, user settings, and granular access controls still apply, while AI-specific modes, approval rules, and transparency requirements extend those foundations.
The takeaway? Most well-known security best practices will still apply in your use of agentic AI. You don’t need a whole new security stance. As Markus put it: “We use the whole control, compliance, and security toolset, including policies, user settings, and conditional access, and then add the best practices for AI.”
Prepare IT professionals to supervise AI
Responsible adoption depends not only on technical controls, but also on the people supervising the AI. IT professionals are understandably anxious about what AI means for their job security. With the arrival of technologies such as Claude Code, some worry about being replaced or about technological development outpacing their skills.
It’s impossible to dismiss these concerns fully, just as it’s impossible to really say what the future holds. But one thing is for sure: AI is going to play an increasingly large role in IT workflows. And IT support is no exception.
For support teams, AI may indeed assume more of the troubleshooting work. In turn, the supporter’s expertise will become crucial, as their role shifts toward supervision, judgment, and handling tricky edge cases.
It’s important to emphasize that human expertise does not disappear. Instead, it moves to the decisions demanding context and accountability. Human expertise and so-called ‘process knowledge’ are needed to make sense of AI recommendations and know exactly when to intervene.
IT works should increase automation only when the available evidence, controls, and confidence support it. Over time, more routine support cases may indeed be opportunities for increased automation, but as it stands, sensitive cases will always require human oversight.
Turn AI capability into real-world results
Tia helps IT teams troubleshoot and remediate issues faster, with the governance, transparency, and controls enterprises expect.
A practical checklist for responsible adoption
The principles above translate into a practical sequence for adopting AI without losing control. Use this checklist to define the use case, set boundaries, test the system, and expand its authority only when the evidence supports it.
- Define the use case. Choose a specific troubleshooting problem or workflow.
-
Assign access. Decide who can use the capability and where.
-
Choose the capabilities AI can perform. Choose advice, diagnosis, or resolution.
-
Begin with a small group. Let experienced users evaluate the system.
-
Classify consequential actions. Ensure that human approval is required.
-
Provide decision context. Show AI’s logic and reasoning for proposed actions.
-
Pause for approval. Do not proceed with a system-changing action until authorization is received.
-
Retain a record. Capture decisions, approvals, actions, and outcomes.
- Review the evidence. Use successes, failures, and rejected suggestions to improve the governance model.
Responsible adoption is not one decision to permit or prohibit AI. It's a controlled process of testing, learning, and gradually assigning authority.
How Tia applies these principles
At TeamViewer, we aim to build AI products that inspire trust, and these principles shaped how we developed Tia, the TeamViewer Intelligent Agent. Tia is designed to simplify IT support through automated troubleshooting and remediation.
Built into our platform with enterprise-grade security functionality, Tia is made for companies who need to know exactly what’s happening with their AI. Actions, approvals, and outcomes remain traceable at all times, supporting governance, compliance, and security.
Tia also offers differentiated functionality – Advise, Diagnose, and Resolve modes – meaning only more experienced workers can activate its agentic capabilities. All the while, administrators control who can do what, and system-changing actions always require user approval. Users always get to review proposed actions, so humans remain fully in control.
The result is an approach to agentic AI that delivers the benefits of the technology while preserving the appropriate level of human control. Human feedback helps Tia improve over time, allowing oversight requirements to be refined as the evidence, controls, and confidence develop.
Conclusion: Control makes capable AI usable
Balancing AI’s capability with human oversight is the precondition not only for getting the most from AI, but also for building sustainable trust in it.
And trust, Andrew Hewitt says, “is not something organizations can demand from employees when AI is introduced. It is built through clear boundaries, meaningful human decisions, visible actions, and experience.”
The more capable AI becomes, the more we need guardrails and a differentiated approach, based on authority. This in turn creates accountability, which is crucial for sustainable governance.
The right AI solution won’t surprise you. It will explain why it recommends an action, show the projected outcome, and carry it out only when you authorize it.
Crucially, the organizations that benefit most will not simply give AI the greatest authority. They will match its capability with clear boundaries and meaningful human decisions.
