keepcalculator

06 Numbers and Powers

AI Agents and Strategy: Who Chooses the Course?

Card No. 06·78019 · Updated · Standard formula, shown below

AI agents can carry out a strategy because they can follow goals, rules, and feedback through a sequence of tasks; humans must choose the course because deciding what matters, which risks are acceptable, and what outcomes are worth pursuing involves values and responsibility. An agent can help compare options and execute a chosen plan, but its ability to act does not make it the legitimate owner of the decision.

What an AI agent does

An AI agent is a software system that takes in information, selects actions, and may use tools or other software to pursue an assigned objective. In practice, the label covers a wide range of systems. Some agents follow a fixed workflow; others use a language model to interpret instructions, plan intermediate steps, and adjust their actions when new information arrives.

Execution can include tasks such as sorting incoming requests, drafting a response for review, searching a permitted database, or moving information between business systems. A more autonomous agent may complete several of these steps without asking at every turn. It still operates within conditions established by people: its permissions, available tools, instructions, and criteria for success.

That distinction matters. Producing actions that advance a specified objective is not the same as deciding that the objective is the right one. A navigation system can find a route to a destination; it does not decide whether the destination is worth reaching.

Execution is not the same as choosing a strategy

A strategy connects an intended outcome to priorities, trade-offs, and a plan for responding to changing circumstances. It answers questions such as: Which customers should receive attention first? How much cost or delay is acceptable? What should happen if two goals conflict? These questions often have no single answer that follows from the available data alone.

Once people define a bounded objective, an agent may help pursue it. For example, a support team might decide that urgent service outages should be handled before routine requests, while also protecting customer privacy and escalating uncertain cases. An agent can sort requests by specified signals, gather relevant information, and prepare a suggested response. The team—not the sorting mechanism—sets the priority and decides what counts as an appropriate escalation.

In a business setting, an agent might also compare delivery options against a chosen balance of cost and speed. It cannot determine from arithmetic alone whether a company should prioritize lower prices, faster delivery, reduced environmental impact, or resilience during disruption. People must establish how those aims rank and what constraints cannot be crossed.

Why human judgment remains necessary

Strategy involves judgments that cannot reliably be reduced to one measurable target. Even when a goal sounds clear—such as “improve service”—people must decide what improvement means. A shorter wait, a more accurate answer, and a fairer distribution of attention may compete with one another.

Human involvement is also important because decisions affect people who may not be represented in the system’s data or instructions. A pattern that looks efficient in historical records may reproduce a past imbalance. Data can inform a decision, but it does not by itself establish that the result is fair, acceptable, or consistent with an organization’s commitments.

Responsibility is another dividing line. When an agent acts, people and organizations remain accountable for choosing to deploy it, defining its permissions, and responding to its consequences. The system cannot meaningfully accept moral or institutional responsibility in the way a person or organization can. This is why consequential decisions commonly need a named human owner, a route for appeal, or a review process appropriate to the situation.

How people and agents can work together

A useful arrangement assigns each side the work it can do well. People set direction and boundaries; agents handle repeatable steps and surface information; people review exceptions and revise the plan when circumstances change. The level of autonomy should depend on the consequences of a mistake, how reversible the action is, and how confidently the system can determine that it is operating within its instructions.

  • Define the outcome: State the goal in concrete terms and identify priorities that may compete.
  • Set boundaries: Specify permitted data, tools, actions, spending or access limits where relevant, and cases requiring escalation.
  • Start with review: Have the agent recommend or prepare actions before allowing it to commit consequential changes.
  • Monitor results: Check for errors, unexpected patterns, and changes in the task or operating environment.
  • Provide a stop or correction path: Make it possible for an authorized person to pause the workflow, undo a reversible action, or handle an exception.

These are not guarantees of good outcomes. They are practical ways to make authority explicit and to keep a capable system from being treated as if it had chosen the organization’s priorities.

Common problems and limits

A frequent misconception is that a sufficiently advanced agent will naturally develop sound judgment about what should be done. More capable planning can help a system handle complex tasks, but it does not resolve disagreements about values or give the system a mandate to set goals. A fluent explanation of a choice is not proof that the choice is wise or justified.

Agents can also misunderstand instructions, rely on incomplete or inaccurate information, or take an action that is technically permitted but contrary to what the operator intended. A sequence of individually reasonable steps may lead to a poor result if the original objective was poorly framed. Tool access adds another concern: a mistaken suggestion is different from a mistaken change to a record, message, or transaction.

For these reasons, oversight should be more than a person occasionally glancing at a dashboard. Review needs to be meaningful: the reviewer should understand the purpose of the workflow, have enough information to spot important errors, and have authority to intervene. High-impact, hard-to-reverse, or contested decisions call for stronger human control than routine, reversible work.

Frequently Asked Questions

Can an AI agent make strategic decisions?
An agent can analyze options, forecast consequences from available information, and recommend a course. Whether it should make the final decision depends on the stakes and the authority people have deliberately delegated. It cannot settle value judgments simply by producing an answer.

What is the difference between an AI agent and an AI assistant?
The terms are used inconsistently. Usually, an assistant responds to requests or helps with individual tasks, while an agent is designed to pursue a goal through multiple steps and may use tools. Neither term guarantees a particular level of autonomy or reliability.

Why can’t people just give an agent a goal and let it decide?
A broad goal leaves important choices unstated: how to resolve conflicts, which risks to accept, and whose interests count. An agent may optimize a measurable proxy while missing the intent behind it. Clear boundaries and human accountability help address that gap.

When should a person approve an agent’s actions?
Approval is especially important when an action could significantly affect someone, expose sensitive information, be difficult to reverse, or create substantial operational consequences. Routine, low-impact tasks may need less direct review if they are bounded and monitored.

Does human oversight mean checking every step?
No. Oversight can be designed around risk: review samples and exceptions for routine work, and require explicit approval for consequential actions. The essential point is that people can understand the system’s role and intervene when needed.

Related

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →