Why Enterprise AI Agents Require More Than A Technology Strategy

AI agents in software no longer answer questions. Microsoft, Google, Salesforce, and ServiceNow have announced AI agents that move between company systems and complete tasks for employees. The harder question belongs to leadership: how much work should employees hand over?

What The Vendors Have Announced

Salesforce made its redesigned Slackbot generally available in January, making Slack the place where employees access company information, workflows, and agents. Microsoft introduced Copilot Cowork in March to complete routine and multistep work across the Microsoft 365 ecosystem. Google followed in October, announcing its Gemini agent for work, along with coworker agents with their own identities that can join team projects and perform tasks just like an employee. ServiceNow’s personal agent with NVIDIA’s OpenShell is meant to let agents do the work while company policy limits what they can see and do.

Some of those safeguards still depend on system integrations. Oracle announced agentic applications for finance and supply chain in April, and SAP is adding agents across finance, HR, and supply chain work.

The products differ in what they can do and how mature they are. Each puts an agent between employees and the systems they use, creating a dependency risk. Whoever sets up the AI agent tool an employee opens first can shape how work flows, who can see company data, and what the company buys next.

Executives must carefully evaluate which AI agents they adopt and ensure their business strategy drives how their organization operates vs. a vendor’s roadmap.

Fix The Work Before You Automate It

Finance has its systems, sales has its own, and hospitals run on the electronic health record. Employees move between them all day. An agent could do that legwork and hand back finished work to review.

For instance, in a hospital that is facing a high volume of denied claims, an agent identifies missing documentation and coding mismatches and prepares the necessary corrections. This process lets staff focus on exceptions and judgment calls. But a more important question is: why are there so many errors in the first place?

If the process stays as is, you’ve automated a bad process and kept the structure that produced it. CEOs should push their business leaders to identify which processes to simplify and which jobs need to change, then set targets for what agents are supposed to improve. In the healthcare payer example above, good places to start are claim resolution time and the share of denials overturned. This is a decision about how the company runs, not an IT rollout.

Saving Time Is Not An ROI

Many AI business cases use time savings to justify return on investment (ROI). But time savings don’t necessarily translate into better business performance.

If an employee saves five hours a week using AI, the company only benefits if that time produces measurable results. Can the team handle more work without adding staff? Can the organization reduce overtime or improve service delivery? If staffing, processes, and output remain the same, the company may have improved individual productivity without realizing a meaningful financial return.

Calculating AI ROI requires consideration of licenses, infrastructure, integration, monitoring, security, and support. For thousands of employees, these costs accumulate quickly. This accumulation happens even faster when departments purchase AI agent solutions from various vendors and allow individual licenses.

CFOs must evaluate it by operating costs, capacity, service levels, and what remains after foundational or administrative costs. The number of agents deployed tells you nothing.

People Stay Accountable

AI agents should not automatically inherit all the access the users have. For example, an employee may have access to financial records or customer data, but an agent acting on that employee’s behalf shouldn’t have the same level of access by default. Preparing a financial report is one task, while authorizing a payment is another. The same principle applies in healthcare. An administrative worker cannot change a patient’s record without proper authorization. AI agents must follow the same philosophy for non-clinical automations.

Vendors have indicated that they are establishing controls. ServiceNow and NVIDIA are working on a system for centralized governance and real-time policy enforcement. Google has highlighted the importance of individual agent identities, access permissions, and audit trails. While these tools help, software alone cannot determine an agent’s level of authority.

Business leaders need to decide when AI can act independently and when it requires human intevention. Routine tasks can often be handled with little oversight, but financial transactions, sensitive data changes, and major decisions should always involve human review.

The CIO Job Gets Bigger

The CIO role involves selecting applications, integrating systems across the portfolio, managing vendors, protecting data, and controlling costs. AI agents add a new layer of complexity to how employees interact with business systems. The CIO now faces the challenge of deciding how these agents operate across platforms, setting user permissions, tracking agent activity, and preventing vendor lock-in.

The other role is to sort out which work suits personal agents and which is better run centrally. Standard, high-volume processes may be more reliable and cheaper when handled centrally. Personal agents fit work that needs flexibility, individual judgment, and coordination across several applications. Not every employee needs an agent, and agents don’t have to replace existing workflows.

Organizational leaders must own the process redesign and its outcomes. Real ROI will come not from having the most agents, but from making strategic decisions about which tasks to automate and ensuring the underlying economics are sound.

Leaders must go beyond deciding how many agents to deploy; they must focus on how much of the company’s work will change, who will be accountable for those changes, and whether the organization is prepared to adapt to a new way of operating.