Ask how to automate something this year and you'll get one of two answers. Build a workflow: fixed steps that run the same way every time. Or hand it to an AI agent: give a model the goal and the tools, and let it work out the rest. Both camps have a point, and both get it wrong when they treat the choice as all or nothing.
The useful question isn't which one is better. It's which part of the job each one should do. Most business processes have parts of both, and the parts that need AI are usually fewer than they first look. Here's a simple way to tell them apart, one step at a time.
Four ways to get a step done
From the most predictable to the least, these are the options for any single step in a process:
- A fixed step. A rule you could write down: add the payment to the sheet, post in #sales if the amount is over $1,000. Same input, same result, every run, and no model involved.
- A judgment call. One yes-or-no question that a person answers in a second but a formula can't: is this email a sales lead? In Lodol, that's a condition set to Use judgment. It answers yes or no with how sure it is, and anything in between can take a path of its own, to a person or a review label.
- An agent. A task where the steps can't be known in advance, because what to do next depends on what it finds. In Lodol, that's the Run an AI agent step we shipped last week: you give a model the task, the tools it may use and the fields you want back, and it decides which tools to use and when.
- A person. Some decisions should stay with someone who will put their name on them. A Request approval step pauses the run until the right person approves, rejects or replies.
A test for each step
Walk through the job the way you'd explain it to a new colleague, and ask four questions of each step:
| Ask | If yes, use |
|---|---|
| Could I write this down as a rule, with no "it depends"? | A fixed step |
| Is it one question I'd answer by reading, not by calculating? | A judgment call |
| Does what happens next depend on what turns up along the way? | An agent |
| Would a mistake here reach a customer, the books or the public? | A person, before it happens |
Take a shared inbox. A new email arrives: that's the trigger. Is the sender asking to buy something? A judgment call. If yes, look up their company in your CRM and post the email in #sales with a link: fixed steps. Draft a reply: an AI step that writes, with a person reading it before anything goes out. Six steps, and two of them involve a model.
Why most processes are a workflow with one or two AI decisions
It's tempting to hand the whole inbox to an agent and let it work out the six steps on every email. It may even work in the demo. Four things go wrong once it's real:
- Small errors add up. A model works the job out afresh each time, so even a step that's right almost every time is wrong some of the time, and a process with many such steps is wrong more often still. Fixed steps don't drift.
- It costs on every run. An agent spends model usage each time, working out the same steps again. A fixed step costs nothing extra. In Lodol, a judgment call doesn't either, because it runs on our account, while an agent runs on your own Anthropic, OpenAI, Gemini or other model account.
- It's hard to read. Anyone on your team can open a workflow and follow it like a checklist. An agent's plan exists only in the run it happened in.
- Fixes don't stay fixed. When a fixed step goes wrong, you change it and it stays changed. When an agent goes wrong, a better instruction makes the mistake less likely, not impossible.
So the sensible default is a workflow, with AI at the few points where reading or judgment is really needed. Our CEO made the longer case in Agents Are Having a Moment. Reliability Is Having a Reckoning.
When an agent is the right call
Some work genuinely has no fixed recipe, and that's where an agent earns its keep:
- Research. Working out who the business behind a bank line like "SQ *BLUE BOTTLE OAKLAND CA" is means searching, reading what comes back, and deciding what to look at next. Our Verify the vendor behind a bank or card line template gives that job to an agent with three tools: a web search, a profile search and a page reader.
- Reading documents. Pulling the vendor, total and due date out of an invoice, a CV or a screenshot. Give the agent the file and the fields you want, and no tools at all.
- Conversations. Answering messages in a chat, where each reply depends on what was said before. An agent can remember earlier runs of the same conversation.
Keep the agent inside a workflow
Even when a step needs an agent, the workflow around it still decides when it runs, what it's given, what it may use and what happens to its answer. That's how the vendor template stays honest. Every fact the agent reports has to come from a page it read or a result it saw, and the workflow checks the agent's status itself: a match with no source behind it is reported as not identified. A second model then reads the summary against the evidence and says whether it claims more than the evidence supports.
In Lodol, the agent step gives you the same levers for your own work:
- A short list of tools, from your apps or your own workflows, with the inputs you don't want it to choose, like which spreadsheet it writes to, fixed in advance.
- Fields to return, so the next step gets a category and a reason it can test with an ordinary If, not a paragraph to interpret.
- A turn limit and a fallback model, so a long task ends, and an outage doesn't end it badly.
- Every move on the record: the model used, the tokens spent and every tool call, so you can see how it reached its answer.
When an agent is the wrong call
- You already know the steps. Search, then read, then write is a workflow. Build it as steps and it runs the same way every time at no extra cost.
- A yes or no is enough. A judgment call is narrower and cheaper, and "not sure" can be one of its answers.
- Nobody looks before it acts. If its output goes straight to a customer, your books or your website, put a person in front of it first.
Start narrow, and widen only when you must
Build the fixed steps first. Add a judgment call where the rules run out. Reach for an agent only for the part with no recipe, and put a person in front of anything you can't take back. Most of the time you'll stop at the second step, and the workflow you end up with will be cheaper to run, easier to read and the same on its thousandth run as on its first.
Put AI where the judgment is.
Read how judgment calls and the agent step work, see the agent guide in the docs, or start free and try both on your busiest inbox.