AgenticChasm

Agent use cases are defined by workflow

Companies are finding their next agent use case in work they already understand. Nearly half start with high-volume, repetitive processes. About one in three look first at whether an agent can connect with the systems already in place, while another third need data that is structured and accessible.

That logic is visible across the market. Software development and customer support remain the two most common agent use cases, but agents are also being applied across research, process automation, sales, and marketing. The point is less about finding a single killer use case than finding work with enough volume, structure, and measurable value to justify an agent.

Review data from the Agent Builder category on G2 shows how one part of that market is changing. As general-purpose platforms have expanded beyond their earlier customer-service roots, software engineering use cases have risen 77% year over year, and process automation has grown 42%. Sales development use cases have more than halved over the same period.

The vendors that cross the chasm will be the ones that can show buyers how work changes after deployment, then use that proof to identify the next workflow worth automating.

Trust has to hold up in the workflow

An agent that performs well in a clean, bounded workflow may not be enough from a buyer’s perspective. Once it encounters incomplete data, conflicting policies, or an exception that does not fit the original design, the quality of the surrounding operating model becomes just as important as the agent itself.

Early adopters can accept a promising product with a few rough edges when it helps solve an urgent problem. Pragmatic buyers need dependable improvement inside the systems they already use, supported by trusted data, clear guardrails, reliable recovery paths, and a practical way to measure whether the gains hold up.

As agents scale across more workflows, those requirements become harder to manage through manual intervention alone. Teams need a clearer view of which agents are operating on their behalf, who owns them, what they can access, and where human involvement is still required.

Agent governance is returning to IT

Functional leaders will continue to identify the use cases where agents can create value. However, as agents connect to shared systems and take on more consequential work, accountability is moving to IT and the teams that manage the underlying technology.

Central IT, along with AI, data, and machine learning (ML) teams, now manage about half (49%) of agent operating costs, and they collectively own 60% of governance accountability. The center of gravity is shifting from simply deciding which software enters the business to governing the systems that increasingly perform work within it, including AI agents.

This changes the buying center. For vendors accustomed to selling directly to customer success, sales, or other functional leaders, IT is no longer only an approver. It is increasingly the team that must be convinced on risk, orchestration, and cost control alongside business value.

That creates a new division of responsibility. Business teams need room to design and improve the workflows they know best, but IT teams need a reliable way to set policy, assign ownership, manage permissions and spend, monitor activity, and coordinate a response when an agent produces an unexpected result. 

Vendors must address this lack of visibility. Visionaries may tolerate scattered experiments and incomplete visibility to gain an early advantage, but for pragmatists, governance is part of the product. They need to know who owns each agent, how it connects to systems, what it costs to run, and how the business will respond when an action goes wrong.

Without that control, they will keep agents on a short leash, limiting the autonomy and scale required to reach the mainstream.

  Download   Download

Agents are reshaping the software stack

The next shift in software will happen one workflow at a time. Agents will coordinate work across applications and increase the value of products whose capabilities, data, and controls help complete that work.

By 2030, 56% of buyers expect agents to orchestrate work across their SaaS stack and moderately reduce new SaaS purchases. Another 20% expect agents to replace a significant share of application functionality, while 5% expect an agent-dominant environment. Together, those expectations describe a meaningful reset in how software is consumed, and the reset is already underway.

  Download

Among organizations with deployed agents, 66% say those agents exchange work across vendors or both within and across platforms. Only 8% say their agents operate independently. The software stack is becoming less a collection of separate destinations and more a set of capabilities that agents can call on to complete work.

Trust does not equal total autonomy

Trust arrives on a mule, but it leaves in a Maserati, and that is proving true for AI agents. G2 review data shows that 67% of B2B software buyers trust their organization’s AI system quite a bit or very much, yet 48% prefer agents with guardrails. Another 29% want them limited to specific tasks, and only 7% would hand over total autonomy.

A low-risk task can run with a light touch, but consequential work requires more evidence, closer monitoring, and a clear path to intervene. Evidence is what lets teams expand control and narrow the trust gap without turning every autonomous action into another approval queue. 

Today, that type of evidence is largely absent from the agent market. Without a standardized way to measure agent performance in critical areas like accuracy, compliance, relevance, and completeness, buyers can’t make an informed decision. It’s one of the biggest reasons G2 recently beta-launched AI Agent Evaluations — because the market has a strong need for a neutral, trusted evaluation layer to help pragmatists compare enterprise agent performance before they buy.

Vendors that make their evidence legible, repeatable, and relevant to a buyer’s own work will have a much clearer path to the pragmatists who decide how quickly agents cross the chasm.