Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

3 questions for your AI agent strategy

Understand the risks and rewards of AI agents.

[Photo: Getty Images]

Profile photo of Rebecca Finlay

CEO · Partnership on AI

In 2023, we saw an explosion in AI following the public release of ChatGPT. While industry leaders declared 2025 the year of agents, it’s clear that hasn’t been the case.

Despite marketing campaigns featuring the term “agentic AI,” there is widespread confusion. What’s really new about today’s agents? Why should businesses, already focused on deploying generative AI systems, consider agentic AI?

What is agentic AI?

Generative AI applications produce content, whereas agents are built to take direct action with varying degrees of autonomy.

For example, a generative AI app can generate a list of top restaurants and available reservation times. A virtual assistant as agentic AI can independently book a reservation and send a calendar invitation. When working well, agents offer a win for productivity.

With increased capability comes risks because agents raise the stakes.

If a generative AI application fails, the customer receives a list of incorrect restaurants and dates. But if an AI agent fails, the impact is much greater.

In our example, the risks might seem low at first—perhaps the app books the wrong date or restaurant. However, if failures compound, agents could cause havoc with a restaurant’s booking system, affecting profits or leading to the unlawful release of personal information.

When applied to other sectors, such as health care or banking, risks increase. As systems scale, human and automated monitoring systems are challenged to keep up.

3 stress-test questions

Before deploying agents widely, leaders must evaluate the opportunities and risks of any application and assess the potential effects of agents deployed by others that could affect their business. To get started, my organization recommends three questions to stress test your AI agent strategy.

1. What are the stakes?


What are the potential hazards of the actions the agents can take? You’re facing higher stakes if agents are deployed in complex, consequential settings or allowed to take actions such as accessing sensitive personal and financial information.

Take, for example, an internal-facing agent that generates meeting notes and sends follow-up emails. If this agent failed in its assigned actions, it would be an inconvenience. Notes that didn’t accurately reflect the conversation, or were shared with the wrong colleague.

On the other hand, if a medical office deploys an agent to share post-visit summaries with patients, an agent sending information to the wrong email address compromises patient privacy and safety.

Conduct a full risk analysis of the range for any agent application. Start with low-risk use cases and a well-constrained set of tasks.

2. How easy is it to reverse an action?

Can the agent’s actions be undone? An agent’s actions are safer in a testing environment than in a real-world application, where an action may not be undone. Irreversible actions, such as deleting files or sending sensitive information, carry significant legal, financial, or reputational risks.

Last summer, there was the cautionary tale of an AI agent making an irreversible error, when an app-building platform’s AI deleted an entire production database. The CEO of the company behind the agent later said they would work on a “planning/chat-only mode so [users] can strategize without risking [their] codebase.” Without that mode available, the customer using the agent lost data on more than 1,200 individuals and 1,000 companies.

Think carefully about the ease or difficulty of reversing an agent’s possible actions and potential effects.

3. How much autonomy can we afford?

What level of independence is the agent allowed? An agent’s architecture with predefined tools and workflows is more predictable than one with fewer constraints and more generative capabilities.

An AI agent used for tax preparation, for example, could be given a great deal of autonomy to use multiple tools, acquire up-to-date information on codes and regulations, request additional documentation, and execute complex workflows to submit documents without human intervention.

Yet at every step, there is room for error, which could lead to costly mistakes. Constraining the agent to well-defined workflows, such as reviewing a specific set of documents and using a single tool to complete tax forms, lowers the risk of errors and, if any occur, makes it easier to identify what went wrong.

Starting with clear rules and controls in place ensures more opportunity to learn the warranted level of autonomy, over time.

Prepare for AI with frameworks and systems

By assessing the stakes, reversibility, and autonomy of an agent application, the appropriate real-time failure detection levels become clearer. This fortifies your AI agent strategy with reliable monitoring systems to prevent harmful or costly failures. Sectors implementing real-time failure monitoring, such as the automotive industry, can provide valuable insights for AI agents.

As with any emerging technology in business, collective work on technical architectures, evaluation practices, and policy frameworks establishes trust between actors.

Gather input from employees and other stakeholders, who are best positioned to identify both automation’s benefits and the times when human control is essential. Together, assess how your current frameworks apply and whether they need updating. They likely will.

Join other leaders through organizations, such as the Partnership on AI, to define the terms of smart and secure AI innovation. Together, we will make 2026 a year for real progress on AI agents.

Rebecca Finlay is the CEO of Partnership on AI.