When not to build an AI agent (and what to do instead)

Who it is for

Owners and managers who are evaluating "adding AI" to their operation and have been offered an agent that does everything.

The problem

Everyone sells agents. An agent decides on its own which steps to take, and that sounds wonderful until you have to explain why it did something odd with a customer's invoice. Anthropic, which builds these systems, is explicit: most successful cases do not use agents, they use predefined workflows.

Ingredients

• The list of processes you want to improve, written as steps
• One hour without interruptions
• The prompt below, which asks the right questions

Step by step

1. 1

   Write the process as you do it today, step by step

   If you can write the steps in order and they are always the same, you already have the answer: it is a workflow, not an agent. Example: receive email → classify → extract data → save to spreadsheet → notify.

2. 2

   Mark where there are decisions that a person makes today using judgment

   A workflow can include a step where the model decides something bounded (classify, summarize, extract). That is still a workflow. Only when the number of steps and their order depend on each case, and you cannot write them in advance, does an agent start to make sense.

3. 3

   Ask how much a mistake costs

   A fixed workflow fails predictably: you know at which step and why. An agent can fail in ways you did not anticipate. If a mistake touches money, customer data or legal commitments, start with a workflow even if it is less elegant.

4. 4

   Start with the simplest version that solves the problem

   Anthropic sums it up like this: look for the simplest possible solution and add complexity only when it proves necessary. A well-written prompt with examples solves more than it seems.

Replace what is in brackets with your company details, or use the guided mode above: it asks you and fills it in for you.

I am going to describe a process in my company that I want to automate with AI. Help me decide whether I need a workflow with fixed steps or an autonomous agent.

Ask me these questions one at a time and wait for my answer before continuing:

1. Can I write the steps of the process in order, and is that order always the same? Ask me to write them out.
2. In which steps is there a decision that a person makes today? Is that decision bounded (choosing among known options) or open (inventing the next step)?
3. What happens if the system gets each of those steps wrong? Does it touch money, personal data or commitments to customers?
4. How many times a day does this process happen and how much human time does it consume today?

When you finish, give me:
- Your recommendation: FIXED WORKFLOW or AGENT, with the main reason in one sentence.
- The specific steps where an AI model adds value and the ones that should remain fixed rules or humans.
- The simplest possible version to start this week.
- What to measure in the first month to know whether it works.

Process to evaluate:
[DESCRIBE YOUR PROCESS]

What usually goes wrong

• ✕Building the agent because it is what the vendor sells, not because the process needs it.
• ✕Confusing "the model makes a bounded decision" with "I need an agent". Classifying or extracting inside a workflow is not an agent.
• ✕Starting with the most complex process. Start with the most frequent and boring one: that is where the savings are and where you learn cheaply.