Is an AI agent better than a simple automation?

Not by default. They suit different jobs. An automation repeats fixed steps the same way every time, which is a strength for predictable tasks. An agent chooses its own steps, which helps when the path cannot be written down in advance. Anthropic notes that agentic systems often trade latency and cost for better task performance.

What is the simplest way to use AI in a business process?

A single call to a language model, often with relevant information supplied to it and a few examples in the prompt. Anthropic says that for many applications, optimizing single language model calls with retrieval and in-context examples is usually enough. Retrieval means finding relevant information for the model to read.

When is an AI agent worth the extra cost?

OpenAI points to three signs: decisions that need judgement, rules that have become too tangled to maintain, and heavy use of unstructured data such as free-text messages or documents. Anthropic adds that agents suit open-ended problems where you cannot hardcode a fixed path. In our reading, the cost of a mistake also matters.

Can I start with a simple setup and move to an agent later?

Yes, and Anthropic suggests it. Anthropic's summary advises starting with simple prompts, optimizing them with evaluation, and adding multi-step agentic systems only when simpler solutions fall short. It also says to consider adding complexity only when it demonstrably improves outcomes. Measure results first, then decide whether more autonomy is justified.

Do You Need an AI Agent, or Is a Simple Automation Enough?

Imagine you run the Lagos fabric shop from the first post, and after a week of reading about AI agents you want one for everything. Then you write down three jobs: send a confirmation email when an order arrives, sort the support inbox, and sort out an order like 1042 that never arrived. Should all three get an agent? Probably not, and this post explains how to tell.

(This is an illustrative scene, not a real business.)

Do you need an AI agent, or is a simple automation enough?

Often a simple automation is enough. In its December 2024 post “Building effective agents”, Anthropic writes: “we recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.” Anthropic has since added a note to that page saying “Much of the tooling landscape described in this post has changed since December 2024.” We therefore use only its general guidance on when to use agents, not its advice on tools.

OpenAI’s guide to building agents says the same from the other side. After listing the cases where an agent helps, it adds: “Otherwise, a deterministic solution may suffice.” A deterministic solution is one that always does the same thing with the same input, like a fixed rule. If you have not read the first two posts, What Is an AI Agent, and How Is It Different from a Chatbot? defines the terms and How Does an AI Agent Actually Work, Step by Step? shows the loop an agent follows.

What options sit between a plain automation and an agent?

There is a ladder with four steps, and you should climb only as far as the task needs. This is our own way of laying out the choices, using Anthropic’s terms.

A fixed automation. A workflow is a set of steps that people write in advance, and the software follows them the same way every time. No language model is needed. Our guide to workflow automation with AI is a companion read on automating business processes.
A single call to a language model. The software sends one request to the model and uses the answer. Anthropic says “For many applications, however, optimizing single LLM calls with retrieval and in-context examples is usually enough.” LLM is short for large language model, the kind of AI model that reads and writes text. Retrieval means finding relevant information for the model to read. In-context examples are sample questions and answers placed in the prompt to show the model what you want.
An AI workflow. Code fixes the path, and a model does the thinking at set points. Anthropic lists five patterns. Two are easy to picture. In prompt chaining, a task is split into a sequence of steps and “each LLM call processes the output of the previous one”. In routing, the software “classifies an input and directs it to a specialized followup task”. The classifier can be a language model or a more traditional program.
An agent. The model chooses the steps itself, using tools, and loops until it is done.

Anthropic’s rule of thumb is: “workflows offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale.”

When does an AI agent make sense?

An agent makes sense when you cannot write the steps down in advance and the task needs judgement. Two vendors give signs. Anthropic says agents suit “open-ended problems where it’s difficult or impossible to predict the required number of steps, and where you can’t hardcode a fixed path”, and adds that “you must have some level of trust in its decision-making.”

OpenAI’s guide tells teams to prioritize workflows that have previously resisted automation, and names three signs:

Complex decision-making. Nuanced judgement or exceptions, such as refund approval in customer service.
Difficult-to-maintain rules. Rule sets so extensive and intricate that updating them is costly or error-prone, such as vendor security reviews.
Heavy reliance on unstructured data. Interpreting natural language or documents, such as processing a home insurance claim.

Unstructured data means information that does not sit in neat fields, such as an email, a letter, or a scanned form.

What does an agent cost you?

In our reading, an agent usually costs more time, more money, and more checking than a fixed automation. Anthropic says “Agentic systems often trade latency and cost for better task performance”, where latency is the delay before you get an answer. It also warns that “The autonomous nature of agents means higher costs, and the potential for compounding errors”, meaning that a mistake early in the loop can carry into later steps.

The previous post in this series showed why this matters, since agents can be inconsistent when the same task is repeated. OpenAI adds that actions that are “sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent’s reliability grows”. So part of the price is a person who reviews the risky steps.

What four questions should you ask before choosing?

These four questions are our own checklist, not a vendor standard. They turn the sources above into something you can answer in a few minutes.

Is the path predictable? If you can write the steps as a list that holds every time, use an automation. If the steps depend on what you find along the way, consider an agent.
Can you check the result? If a person or a test can quickly confirm the outcome, an agent is safer to try. If nobody can tell a good result from a bad one, start simpler.
What does a mistake cost? A wrong tag on an email is cheap to fix. A wrong refund or a message sent to the wrong customer is not. The costlier the mistake, the more you want fixed rules or a person approving the step.
How often does the task come up? A job that repeats hundreds of times a week rewards the effort of setting up a fixed process. A rare, messy job may not be worth building anything for yet.

What does this look like for the fabric shop?

Illustrative example. Imagine the three jobs from the start. Everything here is invented to show the reasoning.

Job 1: the confirmation email. The path is fixed, the result is easy to check, and a mistake is cheap. A plain automation is enough: when an order arrives, send the email and add a row to a spreadsheet. A language model adds nothing.
Job 2: sorting the support inbox. Messages arrive in plain language and fall into a few groups, such as order status, returns, and wholesale questions. The path is predictable once a message is labelled, so this is a single model call or a routing workflow, and a person handles each group. Anthropic’s own routing example is directing “different types of customer service queries (general questions, refund requests, technical support) into different downstream processes, prompts, and tools”. If you want a chatbot to answer the common ones, our chatbot implementation guide covers building AI-powered customer support.
Job 3: order 1042 never arrived. Imagine it is Monday morning, the order shipped last week, and the customer is waiting. You cannot write the steps in advance. The agent might look up the order, check the courier, find the parcel held at a depot because the address is incomplete, and draft a message to the customer. A refund or a message to the customer can wait for your approval. This task has no fixed path, it mixes a customer’s free-text message with records in your systems, and a person can approve the risky step. It fits an agent.

In our reading, the result is one agent for one job out of three. The shop gets the benefit where flexibility matters and keeps the cheap, predictable parts cheap.

What to remember

Start with the simplest thing that works. A fixed automation or a single model call is enough for many jobs.
An agent fits when the steps cannot be fixed in advance, the task needs judgement, and a person can check the risky parts.
Ask four questions: is the path predictable, can you check the result, what does a mistake cost, and how often does it happen.

Next in the series: how AI agents use tools, and what it means when a model “calls a function”.