Agent, chatbot or script: choosing the right tool for the job

"Agent" is now the most overloaded word in business software. Vendors apply it to scheduled scripts, to chat windows with a few integrations, and to systems that plan and act for hours. For a buyer the distinction is not academic: the three approaches differ in cost, predictability and risk. This post defines them in plain terms and gives a rule for choosing.

Three things, three definitions

A script is conventional automation: if X, do Y. It is deterministic, cheap and auditable. It breaks when inputs look different from what its author expected.

A chatbot is a language model wrapped in a conversation. It reads a question, produces an answer, and waits. On its own it takes no actions in your systems. It may be given retrieval over your documents so its answers are grounded in your material, but the human still drives each step.

An agent is a model that directs its own work. Anthropic, which has published one of the clearer industry definitions, separates two kinds of "agentic" system ^[1]:

"Workflows are systems where LLMs and tools are orchestrated through predefined code paths."

Agents, by contrast, are systems where the model dynamically directs its own process and tool use, keeping control over how it accomplishes the task ^[1]. The practical difference: in a workflow, your engineers decide the sequence and the model fills in language-heavy steps. In an agent, the model decides the sequence.

Where the middle ground sits

The binary of "script versus agent" hides the option most businesses actually need. Anthropic describes a set of workflow patterns that use a model inside a fixed structure ^[1]:

• Prompt chaining: break a task into sequential steps, with optional checks between them.

• Routing: classify an input and send it to a specialized follow-up.

• Parallelization: run subtasks at once, or run the same task several times and compare.

• Orchestrator-workers: a central model delegates subtasks to others.

• Evaluator-optimizer: one model drafts, another critiques, in a loop.

These give you language understanding without giving up a predictable path. Routing an inbound email into one of six queues, then drafting a reply for human approval, is a workflow. It is not an agent, and for most teams it is the better first project.

A side-by-side comparison

The final row follows Anthropic's own framing: workflows suit well-defined tasks that need predictability, while agents suit open-ended problems where the number of steps is hard to predict, and they require trust in the model's decisions ^[1].

An illustrative example

Consider a company that receives supplier invoices by email. This is a hypothetical scenario, not a description of a specific engagement.

• A script can post invoices that arrive as a clean, fixed-format file.

• A chatbot can let an accounts-payable clerk ask, "What did this vendor charge us for last quarter?"

• An AI workflow can read invoices in any layout, extract the fields, match them to purchase orders, and queue exceptions for a person. The steps are fixed; the model handles the variability of the documents.

• An agent becomes relevant only if the task is genuinely open-ended, for example investigating why a vendor's charges drifted over a year, which means deciding what to look up next based on what the last lookup revealed.

Most of the value in this example arrives at the workflow tier.

Why the market is blurring the lines

Tooling is making agent-style systems easier to assemble. OpenAI's AgentKit, announced at its developer conference on October 6, includes a visual canvas for building and versioning multi-agent workflows, an embeddable chat interface, evaluation tools and a connector registry for administrators ^[2]. Easier assembly lowers the cost of building an agent; it does not remove the trade-offs. Anthropic is explicit that agentic systems "often trade latency and cost for better task performance," and that autonomous agents bring higher costs and the potential for compounding errors, which is why it recommends extensive testing in sandboxed environments with guardrails ^[1].

A selection rule

1. Can you write the steps down? If yes, build a script or a fixed workflow.

2. Is the input unstructured but the sequence known? Use an AI workflow with a human checkpoint on consequential actions.

3. Does the right next step depend on what was just discovered? Only then consider an agent, and bound it: limited tools, spending limits, logged actions, and a person approving anything irreversible.

4. Does a single well-prompted model call with retrieval do the job? Anthropic notes that for many applications, optimizing a single call with retrieval and in-context examples is enough ^[1].

For most organizations, the right first step is the lowest rung that solves the problem. Match the tool to the predictability of the task, and treat a free-running agent as something earned by a process that has already outgrown a fixed workflow.