AI Agents at Work: What Businesses Should Automate, and What Managers Should Still Control
A customer-service agent that can issue refunds, a purchasing assistant that can order inventory, and an accounting system that can investigate overdue payments all promise to save employees time. Increasingly, these tasks can be handled by artificial intelligence systems that do more than generate information. They can interact with business software, complete sequences of activities, and act with varying degrees of independence.
For managers, this creates a difficult question. Giving an AI system access to company records is one thing; allowing it to spend money, alter customer accounts, or make decisions affecting employees is another. The more authority an AI agent receives, the greater the potential consequences when its information is wrong or its actions exceed what the business intended.
Businesses therefore need to distinguish between work that can be automated and decisions that require human authority. The objective should be to reduce unnecessary manual effort while preserving accountability, sound judgment, and control over consequential outcomes.
How AI Agents Change Business Automation
Traditional automation operates largely through predefined instructions. A system may send an invoice reminder after 30 days, generate a report every Monday, or notify employees when inventory reaches a specified level. These processes can be highly effective, but their actions normally depend on rules established in advance.
AI agents introduce greater flexibility. Depending on their design and permissions, they can interpret an objective, gather information, decide which tools to use, and perform several connected steps. For example, an agent handling a delivery complaint might retrieve an order record, examine tracking information, check the company's refund policy, and prepare a response. If authorized, it could go further and issue compensation without waiting for an employee.
This ability to act makes AI agents commercially interesting, but it also changes the nature of oversight. An inaccurate report can be corrected before anyone relies on it. An unauthorized payment or disclosure of confidential information may be much harder to reverse.
Business adoption is nevertheless increasing. In its August 2026 State of AI survey, McKinsey reported that 40% of respondents from organizations with annual revenue exceeding $1 billion said their organizations were scaling AI agents, compared with 27% in the previous survey. Among smaller organizations, the proportion remained at 22%. These findings indicate growing deployment, especially among larger enterprises, although adoption alone does not establish that the investments are profitable.
Percentage of respondents reporting that their organizations are scaling AI agents
Source: McKinsey, The State of AI in 2026: On the Road to ROI, August 25, 2026. Survey-reported adoption does not establish profitability. Bars use a common scale from 0 to 50%.
Which Business Tasks Are Suitable for AI Agents?
The best opportunities generally involve activities with clear objectives, accessible information, recognizable exceptions, and manageable consequences when errors occur. These conditions matter more than whether a task appears repetitive or technically easy.
In administration, agents can retrieve information from approved systems, organize documents, classify requests, and prepare reports. A project manager, for instance, might use an agent to collect progress updates and identify tasks that appear delayed. The software could reduce the time spent assembling information, leaving the manager to investigate causes, assess dependencies, and decide whether intervention is necessary.
Customer service offers similar opportunities. Agents may answer routine questions, retrieve delivery information, and handle straightforward requests within established policies. A customer asking where a package is located presents a different level of risk from someone disputing a substantial charge or alleging unfair treatment. In the latter situation, the business may need a person who can investigate circumstances, interpret policy, and exercise discretion.
Finance and procurement require more careful boundaries. An agent could compare supplier quotations, identify invoice discrepancies, or recommend replenishment based on inventory records. Yet the authority to pay an invoice, change a supplier, or commit the business to a major purchase should depend on established approval controls. Automating the preparation of a decision can be worthwhile even when the final decision remains human.
The financial case also needs examination. An agent that processes 500 invoices quickly may appear productive, but that advantage becomes less convincing if its errors require extensive investigation. Managers should compare total processing costs, correction time, accuracy, and business outcomes against the existing process, not simply measure the number of tasks completed.
What Decisions Should Remain Under Management Control?
The distinction between automated execution and managerial authority is particularly important in strategy, financial commitments, employment, and business ethics. These decisions involve consequences that cannot always be evaluated through predefined rules or historical data.
Strategic decisions require management to determine which objectives deserve priority and what level of uncertainty the organization is prepared to accept. An AI agent might compare expansion opportunities, analyze competitor information, and develop financial scenarios. It cannot assume responsibility for the decision to enter a market, close a division, or redirect investment. Those choices involve commercial judgment, organizational capabilities, and obligations to stakeholders.
Financial control presents a more immediate problem. A purchasing agent may identify a genuine need for additional inventory, but that does not establish whether the company has sufficient cash or whether the proposed order is commercially sensible. Managers should determine appropriate spending limits, approval thresholds, and exceptions. Low-value, routine transactions may be suitable for controlled automation, while larger or unusual commitments require additional scrutiny.
Employment decisions deserve particular attention because errors can directly affect people's careers and livelihoods. AI may help arrange interviews, organize training records, or prepare administrative information, but decisions involving hiring, dismissal, promotion, and disciplinary action require appropriate human assessment and compliance with applicable law. An automated recommendation may reflect incomplete evidence or reproduce biases present in the information used to generate it.
The European Union's AI Act, Article 14, provides an example of how regulation approaches this issue for systems classified as high-risk. Its human-oversight provisions address the ability of authorized people to understand system limitations, monitor operation, recognize excessive reliance on AI outputs, and intervene or override decisions where appropriate. The legal requirements depend on the system and applicable provisions, but the management principle extends beyond Europe: oversight should be meaningful and proportionate to the risks involved.
Managers must also retain responsibility for ethical judgments and sensitive stakeholder relationships. A system may calculate the financial advantages of reducing staff or refusing an exceptional customer request, but the lowest-cost option is not necessarily the most responsible decision. Commercial consequences, employee welfare, customer trust, and reputation require consideration together.
The Risks of Giving AI Agents Too Much Independence
The problem with an autonomous agent is not limited to whether it produces an inaccurate answer. Its access to tools and records can allow an error to become an action. A system that misunderstands a customer request might change an account, approve a refund, or share information with an unintended recipient before the mistake is detected.
The OWASP Top 10 for Agentic Applications, published in December 2025, identifies security risks including agent goal hijacking, misuse of connected tools, excessive privileges, manipulation of an agent's memory or context, and cascading failures. One concern is that malicious instructions embedded in an email, webpage, or document could influence an agent into performing actions unrelated to its legitimate assignment. This is particularly dangerous when the agent has broad access to sensitive systems.
A second risk comes from employees trusting automated recommendations too readily. Human approval provides limited protection when a reviewer simply accepts whatever the system proposes. Effective oversight requires access to the relevant evidence, enough understanding to challenge the recommendation, and genuine authority to reject or stop an action.
Organizations should therefore limit each agent's permissions to what its task requires. A reporting assistant does not need unrestricted access to payroll records, and a customer-service agent should not automatically possess authority to approve substantial financial adjustments. Clear access restrictions, transaction limits, monitoring, and escalation procedures can reduce the consequences of mistakes without preventing useful automation.
A Practical Way to Divide Work Between AI and Managers
Consider a hypothetical medium-sized distributor introducing an AI agent to manage inventory replenishment. Employees currently examine stock levels, compare recent sales with forecasts, and prepare purchase recommendations. The company wants to reduce the time spent gathering information and identifying potential shortages.
The agent could inspect approved inventory records, calculate expected requirements, compare suppliers, and prepare proposed orders. However, it might misinterpret an unusual sales increase as permanent demand or fail to recognize that a supplier's quoted terms have changed. Allowing the agent to place every recommended order automatically could create unnecessary inventory costs or cash-flow pressure.
Management could instead give the agent authority to complete routine analysis and prepare recommendations, while requiring purchasing approval for significant orders, unfamiliar suppliers, and unusual quantities. Smaller transactions might eventually be automated within defined limits if testing demonstrates reliable performance. The company would measure forecast accuracy, inventory availability, purchasing errors, correction costs, and actual staff time saved before expanding the system's authority.
This illustrates why automation should be designed around the consequences of an action rather than the apparent sophistication of the technology. The same agent may operate independently for one activity, require approval for another, and be prohibited from making certain decisions altogether.
The U.S. National Institute of Standards and Technology offers a useful basis for this approach through its AI Risk Management Framework, which organizes risk management around governing, mapping, measuring, and managing AI risks. Its Generative AI Profile provides additional guidance relevant to systems using generative AI. Businesses can apply these principles by defining an agent's purpose, assessing its operating risks, evaluating performance, and maintaining controls throughout its use.
The emphasis on measuring actual outcomes matters. Business Studies Q&A's earlier case study on worker productivity examines how productivity measures are influenced by both company-specific and industry-specific factors. The same caution applies when assessing AI automation: faster task completion does not automatically mean greater organizational productivity or profitability. A credible evaluation must consider the entire process, including the costs of supervision, correction, and system maintenance.
Conclusion: Automate Work Without Surrendering Accountability
AI agents can reduce the administrative effort involved in gathering information, processing routine requests, and coordinating business activities. Their usefulness, however, depends on more than technical capability. Managers need evidence that an automated process produces reliable results, costs less than the alternatives, and operates within acceptable limits.
The strongest approach is selective delegation. Businesses can give agents greater independence in low-risk, well-defined activities while retaining managerial authority over significant financial commitments, strategic choices, consequential employment decisions, and sensitive exceptions. Human oversight must also involve genuine judgment rather than automatic approval of machine-generated recommendations.
As AI agents become more capable, the responsibility of management will increasingly involve deciding where automation is appropriate and how its results should be evaluated. The business that automates the most tasks will not necessarily perform the best. More important is whether it can use automation to improve decisions and operations while keeping responsibility firmly where it belongs.
Sources and References
- McKinsey & Company. (August 25, 2026). The State of AI in 2026: On the Road to ROI.
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
- National Institute of Standards and Technology. (July 2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile.
- OWASP GenAI Security Project. (December 9, 2025). OWASP Top 10 for Agentic Applications for 2026.
- European Union. (2024, consolidated text subsequently updated). Regulation (EU) 2024/1689, Article 14: Human Oversight.