AI agents are moving beyond answering questions and writing drafts into software development, sales, marketing, and customer service — but the bigger enterprise challenge may be deciding how much autonomy to give them.
The enterprise AI conversation is changing.
Companies are largely treating generative AI as an assistant in the last few years, just to speed up tasks they were already performing, like asking questions, getting AI copies, and summarized documents, etc.
But right now this model is beginning to look outdated.
New enterprise usage data shows a shift toward agentic AI, where systems can research information, use tools, modify files, execute multi-step workflows, and continue working with less human intervention. As of June 2026, agentic AI accounted for 64% of combined Codex and ChatGPT enterprise output tokens, according to OpenAI.
The important shift isn’t simply that businesses are using more AI.
They’re starting to delegate work to it.
AI Is Moving From Assistance to Execution
If we talk about traditional generative AI, it’s mostly reactive: a person provides an instruction, the model produces an answer, and the person decides what the next step is.
But AI agents add another layer: execution.
An AI agent can now search for information, interact with tools, work on multiple steps, and produce an outcome without requiring a human to direct every individual action.
OpenAI’s latest enterprise data offers a useful signal of how quickly that model is spreading. It shows Codex has generated about 64% of combined Codex and ChatGPT output tokens among enterprise customers as of June 2026. Because agentic workflows typically involve longer, multi-step tasks, OpenAI says the figure reflects not only usage frequency but also the amount of work being carried out through those workflows.
The shift is also spreading beyond engineering.
According to the reports, Codex has increased in weekly active users across several business functions and departments since February 2026. And here is the data provided for this analysis: 108× in legal, 41× in sales and account functions, 41× in recruiting, 26× in marketing and communications, 24× in healthcare and clinical work, and 20× in finance and accounting.
The use of agentic AI workflows in companies has now increased engineering capabilities by 5x, making it more efficient for enterprise businesses.
The question is now entirely opposite from in 2023: “Can AI help this employee complete the task?” to “Can an AI system complete this workflow with the employee supervising the result?” in 2027.
That is a much bigger operational change now.
Where Businesses Are Actually Using AI Agents
The clearest evidence of this shift comes from the types of work businesses are already delegating.
In software development, AI agents like Claude Code and Codex can handle multi-step application building, from front-end to more technical operations in back-end development, like refactoring, testing, and terminal-based tasks inside sandboxed environments. Still, humans have full control to review the process and make requests according to the results they are looking for.
You will notice a similar approach done by GitHub Copilot Claude agents for repository work, like researching codebases, preparing implementation plans, and making changes through background development environments.
The model is spreading beyond software.
Companies like Anthropic are using managed AI agents for marketing and sales enrichment, including b2b prospect research and audience profiling. Human analytics teams can review the aggregated results, while AI workflows can run through orchestrated multi-agent processes in parallel.
Microsoft is pushing AI agents deeper into customer operations by automating parts of the customer service lifecycle, including ticketing, routing, CRM updates, and case resolution. Apart from business orientation, Microsoft customer knowledge agents are using a more supervised approach to solve customer queries by analyzing case histories, identifying knowledge gaps, and drafting content according to the data, again reviewed and published by a human support member.
Many businesses are opting for AI agents in different functions of their business and look for the pattern now.
Businesses aren’t necessarily choosing between 100% human work and 100% autonomous AI. Instead, they’re building different levels of autonomy into different workflows.
A software agent can work perfectly fine independently inside a controlled environment to get some good results. Also, a customer service agent can speed up the process of resolving existing issues and escalate the unusual ones with automation. And a marketing agent can gather, organize, and analyze information while leaving the final decision to a human.
That spectrum is likely to become one of the defining characteristics of enterprise AI adoption.
The Governance Problem Is Getting Harder
A chatbot can give you a bad answer… An agent can potentially take a bad action.
Giving an AI system the ability to act also creates a problem that can directly impact running a business.
What happens when the agent has permission to do something it shouldn’t?
A recent Boston Consulting Group analysis argues that AI agents are scaling faster than enterprise governance. The problem becomes particularly difficult when companies deploy agents across multiple platforms, business units, and vendors, each with its own governance approach.
The risks are surprisingly practical.
If the AI agent is running in a business, it can gain access to systems or models it was never intended to reach, which may lead to sensitive data leaks, and in some cases it might break further functionality of the business.
BCG identifies four recurring problems: identity gaps, lack of central agent inventory, duplicated governance work, and rising costs. The simple answer to it is a governance layer that can provide common identity, visibility and policy enforcement across different agent platforms that can still add a firewall layer. The simple answer to these problems is that they proposed a governance layer that can provide common identity, visibility and policy enforcement across different agent platforms.
The important part is that governance cannot simply become another approval queue.
It’s completely relevant to the “golden path” concept by BCG, where teams receive production-ready templates with identity, registration, monitoring, and policy controls already built in for specific work. Instead of making developers figure out governance themselves, the compliant path becomes the easiest path.
That’s increasingly important as agent deployment scales, and there are multiple companies opting for this solution to run their AI agents precisely.
At IMA Financial Group, more than 95% of its 3,000-plus associates now use AI daily across thousands of agentic workflows. The company has built an internal AI Studio to turn employee ideas and pilots into structured enterprise tools while maintaining responsible governance.
That’s probably closer to how enterprise agentic AI will actually develop: more delegation, not necessarily less human responsibility.
What This Means for Marketers and Founders
For marketers and startup founders, the most important change is that AI agents make workflow design more valuable.
Marketers are using AI agents for scheduling recurring workflows, from scheduling email marketing by experimenting with five different versions of emails to enriching CRM records and preparing outreach workflows before a human reviews them.
A founder can look at the same problem differently.
Instead of just questioning and getting text responses, AI agents can now identify repetitive workflows that consume hours every week and automate those workflows in their business.
That could include research, lead qualification, reporting, internal documentation, customer-support triage, or other information-heavy operations.
But autonomy should match the risk of the task.
An agent collecting public information doesn’t need the same permissions as an agent modifying customer records. For example, an agent that drafts marketing material doesn’t need the same authority as one capable of changing finance or operational systems.
This is where small companies may have an advantage of not needing thousands of ai agents to run their business. Small businesses can start with a handful of tightly defined workflows, and based on the results, they can eventually increase their number.
The competitive advantage won’t simply come from having access to the newest model.
It will come from knowing where to deploy autonomy safely.
The Next AI Advantage May Be Control
The companies that benefit most from AI agents may not be those that deploy the most agents.
They may be the companies that figure out where agents can act independently, where humans must remain involved, and how to monitor both.
As AI moves from assistance to execution, control may become just as important as capability.