Why AI agent adoption in businesses still faces governance and evaluation challenges

The implementation of artificial intelligence agents has advanced rapidly, but the numbers vary depending on the measurement method. A PwC survey shows that 79% of executives say their organizations already use AI agents. However, studies analyzing systems actually in operation reveal much lower percentages.

A survey compiled by BvLogic indicates that only 31% of companies have at least one AI agent in production. This discrepancy explains why many companies experiment with the technology without fully integrating it into their daily operations.

From testing to operation: a leap that requires more than a demonstration

Developing an agent capable of performing a task in a controlled environment can be relatively simple. The situation changes when the agent needs to access internal systems, manipulate real data, follow corporate rules, and continue functioning in situations not anticipated during testing.

An agent can deliver impressive results in a demonstration, but this does not mean it is ready to perform activities that impact customers, payments, documents, or internal processes without adequate supervision.

Governance defines the limits of autonomy

The greater the autonomy granted to technology, the greater the need to establish clear limits. Who can authorize an action? What information can the agent access? When is a human review mandatory? What happens when the tool makes a mistake?

A study by Kore.ai reveals that 62% of surveyed companies have already postponed deployments due to concerns about governance and observability, while 53% admit to having put an agent into production without fully understanding or trusting its behavior.

Rigorous evaluation is essential to move pilot projects into production

Measuring whether an agent actually works for the proposed task is another critical point. Testing the technology solely based on the quality of a few answers can hide flaws that arise when it needs to handle hundreds or thousands of different situations.

BvLogic points out that evaluation gaps are among the main obstacles preventing agent projects from advancing to production. According to data compiled by the organization, 88% of pilot projects do not reach this stage.

Therefore, before expanding usage, companies need to define objective criteria: which tasks will be automated, what results are expected, what errors are acceptable, and when human intervention is mandatory.

Professionals need to take on a new role in the face of automation

The expansion of AI agents changes the type of knowledge required within organizations. Professionals involved in these projects need to understand processes, identify where automation makes sense, monitor results, and recognize situations that require human intervention.

This applies to technology areas, but also to marketing, customer service, human resources, finance, and operations. The more an agent performs steps in a process, the more important it is for someone to be able to supervise its work and connect the technology to the area’s real objectives.