LLMs vs AI agents: Managing different AI risks

According to the IDC’s latest Market Glance on agentic AI ecosystems, 50% of organisations are already deploying AI agents across multiple business areas. 

Large Language Models (LLMs) and AI agents share many of the same underlying risks, including inaccurate outputs, inappropriate data use, bias and security vulnerabilities. The difference is that AI agents are usually given access to organisational data and systems, and the ability to take actions with varying levels of human oversight. This changes both the level of risk and how those risks need to be managed.  

This blog explores the key differences between LLMs and AI agents, what that means for your organisation’s risk profile, and how to adapt your governance as your use of AI grows. 

Key takeaway: LLMs and AI agents share many underlying risks, but those risks can change and escalate rapidly as systems gain greater autonomy, access, and authority. Apply governance proportionately, based on what an AI system can do and the potential impact if something goes wrong. 

• What is the difference between an LLM and an AI agent?
• How AI agents can amplify existing risk
• How to determine the required level of AI governance
• How to avoid an AI governance gap

What is the difference between an LLM and an AI agent?

LLMs respond to prompts by producing or working with content, for example by drafting text, summarising information, answering questions, or analysing documents. 

AI agents go further. They may use LLMs alongside tools, data sources, and memory to pursue a goal or complete a task. Rather than simply generating an output for a person to act on, they can plan multiple steps, interact with other systems, and take actions with varying levels of human involvement. 

From a governance perspective, the key difference is the level of autonomy, access and authority involved.  

• LLMs: Can you trust the output and use it appropriately? 
• AI Agents: Can you trust what the agent can access, decide and do, including the actions it takes, with the right level of human oversight?  

How AI Agents can amplify existing risk

An inaccurate AI output can have very different consequences when using an AI agent compared to an LLM, depending on what the system can do. 

When an AI output becomes an action: 

• Using an LLM, an incorrect output is corrected and reviewed before it is used 
• Using an AI Agent, the same error could be acted on automatically, for example updating a record, sending a communication, or triggering the next stage of a workflow 

Similar risks also apply to personal dataInformation which relates to an identified or identifiable natural person.. With an LLM, an employee might enter unnecessary personal data into a prompt. And with an AI agent, it may be able to retrieve that information directly from connected systems if its permissions are too broad. 

However, this does not mean every AI agent is inherently higher risk than every LLM. The level of risk depends on the system’s capabilities, the context in which it is used, and the potential consequences if something goes wrong. 

How to determine the required level of AI governance

These core elements all form part of effective AI governanceThe framework of policies, processes, and roles that ensure Artificial Intelligence (AI) is developed and used responsibly, ethically, and in compliance with applicable laws, societal expectations, and corporate values.  for both LLMs and AI Agents: 

• Data Protection Impact Assessments (DPIAs) 
• AI Impact Assessments (AIIAs) 
• Policies, access controlsA series of measures (either technical or physical) which allow personal data to be accessed on a need-to-know basis., and staff training 
• Human oversight and clear lines of accountabilityPerhaps the most important GDPR principle, which requires controllers to take responsibility for complying with the GDPR and, document their compliance. 
• Testing and monitoring

The level of control needed depends on the system and how it’s being used. Four factors are particularly important: 

1. Autonomy: How independently can the tool operate?

You should define when human approval is required and implement stronger mechanisms for intervention and escalation as autonomy increases.  

2. Access: What can the tool see?

Access should be limited to the information necessary for the system’s approved purpose. You should understand what data the system can retrieve, including personal or special category dataTypes of personal data listed in Article 9(1) GDPR that are considered sensitive and thus require extra protection. Article 9(1) lists data relating to: • racial or ethnic origin • political opinions • religious or philosophical beliefs • trade union membership • genetic data • biometric data • health • sex life • sexual orientation Where these types of personal..., and whether connections to other applications create additional data flows that need to be assessed. 

3. Authority: What can the AI tool do?

Consider what the LLM or AI Agent can change or cause to happen: 

• Can the system only retrieve a customer record, or can it edit or delete that record? 
• Can it draft an email, or can it send one? 
• Can it recommend an action, or can it make or implement a decision? 

Permissions should be limited to what the AI needs to perform its approved task, with appropriate safeguardsWhen transferring personal data to a third country, organisations must put in place appropriate safeguards to ensure the protection of personal data. Organisations should ensure that data subjects' rights will be respected and that the data subject has access to redress if they don't, and that the GDPR principles will be adhered to whilst the personal data is in the... in place to prevent or limit unintended actions. 

4. Impact: What happens if the AI tool gets something wrong?

Consider both the potential effect on individuals and the organisation, as well as whether an incorrect action can easily be identified and reversed. The greater the potential impact, the stronger the safeguards that may be required. 

Download our White Paper: Deploying LLMs with confidence for further guidance on generative AI governance.

In case you missed it…

• Governing AI agents: What organisations need to consider 
• AI Officer vs DPO: Defining roles in AI governance 
• AI Impact Assessments: What are they and why do you need one?