When Should a Company Skip an AI Agent?

Payroll, bank transactions and airline seat assignments depend on predictable outcomes. A SiliconANGLE column by Jason Bloomberg argues that teams should weigh an agent’s variable, goal-directed behavior against deterministic software before deploying one. Agents can write code, communicate with customers and identify or resolve IT issues, but their flexibility also creates risk ¹. Start with the work that needs doing, not with the fact that an agent can do it.

Why “can use an agent” is the wrong first question

An LLM-based agent may produce different outputs for repeated inputs. That nondeterminism supports its ability to make decisions and take different actions toward a goal, but it can be a liability when a workflow needs predictable results ¹.

Deterministic software is useful when the same inputs should produce the same outputs. The column names payroll, bank transactions and airline seat assignments as examples where predictable behavior matters, and argues that conventional software is preferable when there is a right answer to find ¹.

Analysis: Treat agency as a design choice, not a default. If the workflow has clear rules and a known-correct output, first ask whether ordinary software can handle it. Consider an agent when the work involves uncertainty or coordination that simpler code cannot address, not merely because the process feels complicated.

AI agent business use cases: what actually benefits from an agent

Agents may fit situations with changing inputs and no single correct answer. The column gives weather-dependent planning, incident mitigation in complex IT environments, and business forecasts based on multiple inputs as examples ¹. These examples illustrate potential fit; they do not make every complex task a good candidate.

The contrast is a workflow whose target result can be specified in advance. If the work requires exact correctness or predictable behavior, an agent’s possibility of giving a poor answer or taking an undesirable action is a risk to weigh ¹.

Analysis: Ask directly: If we wrote down the rules and expected output, could ordinary software do this? If yes, consider building that first. If not, identify what varies, such as the available information, competing goals or need for multi-step coordination. “It’s complicated” is not a justification by itself.

AI agents vs automation: who decides the next step

Approach	Who chooses the next step	Good fit when
Deterministic automation	Software follows explicit rules	Inputs and outputs are known; repeatability matters
Human-directed AI assistance	A person directs the work and decides what happens next	AI drafts, analyzes or prepares; a person retains judgment
Agent	The system picks steps and may call tools to pursue a goal	Conditions vary enough to justify goal-directed action

This is a screening framework, not a claim that every product fits neatly into one row. The column warns that an agent may give a poor answer or take an undesirable action ¹. Human-directed assistance is another option: AI can help with work while a person retains the decision and authority to act.

Analysis: Moving from automation to assistance to agency can increase flexibility, while also requiring clearer limits on which decisions the system may make and which stay with a person. Choose the least autonomous approach that does the job. For a deeper look at what evidence should exist before an agent acts, see AI agent governance: what evidence is enough before an agent can take action.

Setting an error budget before you deploy

Before deploying an agent, consider the cost of misbehavior. The column uses 5% as an illustration, not a general rule, and says that an error budget of zero means the agent should not be deployed ¹. The appropriate tolerance depends on the consequences of mistakes in the specific workflow.

Analysis: Express the budget in terms the business tracks, such as incorrect cases, financial loss, delayed orders or affected customers. Consider both frequency and severity: a workflow might tolerate frequent, inexpensive-to-fix errors while rejecting a single harmful one.

Write down, before launch:

• Acceptable error frequency
• Maximum damage from one error
• Who can stop or reverse an action in flight

The workflow owner should define these limits. An agent is a poor fit where exact correctness is required or misbehavior could cause serious harm ¹.

AI agent ROI: when the economics kill the idea

Agents may make several model calls, retry actions and invoke tools. The column argues that if the cost of that behavior outweighs the expected business benefit, an agent does not make economic sense ¹.

Analysis: Price the whole workflow, not just the model call. Include review time, retries, tool use, corrections, escalations and slower execution. Compare those costs with what the agent is expected to provide, such as faster resolution or broader coverage. Retries in particular deserve testing before launch; see testing for retries that cost money and outputs that vary.

A rough worksheet can make assumptions explicit:

expected value of outcome
– baseline process cost
– agent and tool cost
– review, correction and escalation cost
= estimated net benefit

This is a decision aid, not a forecast. If review and correction absorb the expected benefit, redesign the workflow or drop the agent.

Where human judgment is the product, not overhead

The column says that LLMs and agents can simulate qualities such as empathy, creativity, insight and understanding, but are mimicking human behavior. It cautions against using agents when those human qualities are essential to the goal ¹.

Analysis: Ask what a person contributes to the outcome. If the work depends on judgment tied to context or relationships, AI might draft or summarize while a person retains the decision and action.

Review is not the same as ownership. If nobody can meaningfully interrupt or override an agent, sign-off may not provide meaningful judgment. Build the human role around the decision that needs a person rather than adding review after the fact.

A screen to run before enterprise AI agent adoption

Run this before choosing a model from a model catalog or writing agent code:

• Do inputs change in ways that change the right next step?
• Is there one correct answer, or several reasonable ones?
• Do results need to be repeatable?
• Are the actions reversible?
• What is the worst credible consequence of an error?
• Is human judgment central to the outcome?
• Would deterministic software already cover this?

These questions apply the column’s distinction between predictable tasks and situations with changing inputs, along with its warnings about poor answers and undesirable actions ¹.

If an agent still looks necessary, write a one-paragraph why agency justification: name the specific decisions and tool calls that require goal-directed behavior. “It’s complex” does not clear that bar.

What to watch after you ship

Analysis: Measure the approach against the reason you chose it. For automation, track repeatability, exceptions, processing time and correction effort. For human-directed assistance, track adoption, time saved, review burden and who owns the decision. For agents, consider tool calls, retries, escalations, undesirable actions and cost per completed outcome. Usage analytics, audit logs and spend controls in an enterprise AI gateway are one place such data can come from.

Tie each metric back to the original case for agency. If variability is not improving outcomes, consider reducing it in favor of automation. If human review is absorbing the expected benefit, redesign the workflow rather than granting the agent more autonomy by default.

A practical synthesis is: use deterministic automation for exact, repeatable outcomes; consider human-directed AI when a person should remain in charge but AI can help; and use an agent only when changing conditions and goal-directed action justify the risk and cost ¹. Run the workflow through the screen before writing agent code.