Agentic AI in the Workplace: Who Owns the Decision?

A dashboard can show an overdue customer request. An AI agent can read the request, update the record, and send the reply. The dashboard informs a person. The agent acts on that person’s behalf. 

For operations and IT leaders at mid-sized organizations, that shift changes the first question to ask about agentic AI in the workplace. What can this system change, and who owns the result? A useful agent needs access to workflows and data. Once it has that access, a plausible but wrong output becomes a real business action. 

Agentic AI in the Workplace Is Outpacing Accountability 

McKinsey’s 2025 survey found that 62% of respondents said their organizations were at least experimenting with AI agents, while 23% were scaling them in at least one business function. Its August 2026 survey shows the gap widening by size. Among organizations with more than $1 billion in revenue, 40% of respondents report scaling agents, up from 27% a year earlier. At smaller organizations, the share held flat at 22%. 

Governance has not kept pace. In Gartner’s 2025 survey of 360 IT application leaders, only 13% strongly agreed they had the right governance structures to manage agents. Gartner also predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing costs, unclear business value, and inadequate risk controls. 

Gartner also warns that some products sold as agents are existing assistants or automations with a new label. Before buying one, ask what decisions it makes, which systems it can change, and why a simpler workflow would not do the job. 

Classify Each Action Before Granting Access 

I sort proposed agent actions by the cost of a mistake. 

Easy to reverse

• Examples: Drafting an internal summary, reordering a task queue, or flagging duplicate CRM records
• Starting control: Allow execution within a defined scope and review samples

Recoverable at a cost

• Examples: Updating a customer or donor record, or sending an external message
• Starting control: Set limits and require approval at the consequential step

Difficult to reverse

• Examples: Releasing a payment, deleting data, signing a contract, or making an employment decision
• Starting control: Require a named person to authorize the action

Reversibility is the first test. Volume, data sensitivity, and external reach matter too, because a small change repeated thousands of times can cause substantial damage. Tighten the control whenever an action touches sensitive data or reaches someone outside the organization. 

Employment decisions need particular care. Under the EU AI Act’s amended timetable, obligations for high-risk systems in areas including employment apply from December 2, 2027. The Act also reaches organizations outside the EU when their systems’ outputs are used there. In the United States, Illinois has required employers since January 1, 2026, to provide notice to employees when AI is used for specified employment purposes. An approval step alone does not establish compliance, so assess the system and how it is actually used. 

Build the Pilot Around One Measurable Workflow 

Start with a narrow task, such as finding mismatched records across two systems and preparing them for review. Record the current time, error rate, and rework before the agent arrives. Then grant access only to the data and actions that task requires. 

For every agent, document four things: its owner, its permitted actions, its approval points, and a record of its actions that a non-engineer can follow. 

Then test the failures. See what happens when data is missing, an instruction conflicts with policy, or a connected system goes down. Review the scope whenever the model, tools, or workflow changes. 

Make Accountability Part of the Design 

An agent has to operate inside a business process that people can inspect and manage. That is the integration work Pumex does across AI, custom software, and data systems, following our staged delivery approach. 

Agentic AI in the workplace earns its place when it removes measurable work without quietly expanding its authority. Choose one workflow, define its boundaries, and compare results with the baseline. 

If you are deciding how agents should connect to your Salesforce, Appian, or custom systems, talk with Pumex about the workflow and its controls before you scale.