The short answer.
A small company should build workflows first and agents later. A workflow that a person starts and checks teaches the company what a piece of work needs: its inputs, its exceptions and what good looks like. That written, tested method is the material anything more automatic is made from, and a company that has not written down how its work is done has nothing to hand to an agent.
Here “agent” means software that carries out steps by itself toward a goal, with little or no person in the loop. A workflow is your best way of doing one piece of work, written down precisely enough that your AI assistant prepares it the same way every time, with a person checking and deciding. A workflow, an automation tool or an agent sets the three side by side on one piece of work.
What a workflow teaches you about the work.
A workflow in daily use teaches a company three things about the work that no one can learn from a whiteboard.
- The inputs.Running a workflow every week shows exactly which records the work starts from, which fields matter and which ones arrive blank. Most teams discover the work depends on a fact nobody wrote down.
- The exceptions.The person checking each output sees the cases that break the pattern: the order with two delivery addresses, the client on old terms, the request that belongs to another team. Each one goes in the exception log and, once handled, in the workflow’s instructions.
- What good looks like.Every time the person approves or corrects an output, the company gets a clearer standard. After a few months the standard is written down, with test cases that prove the workflow meets it.
An agent needs all three before it can act on its own. Inputs, exceptions and a standard are what you give it, and they are what you check it against. A workflow builds them while your people stay in charge, so the step after a workflow is short. The step from nothing straight to an agent is long, and the agent learns its exceptions on your customers.
A workflow runs in your own ChatGPT, Claude, Copilot or Gemini, on your account, run by your people. The workflow prepares; a person decides. We have a plan. How do we build the first AI workflow? walks through the first one.
Everyone is talking about agents. Are we behind?
No. A company with ten workflows in daily use is ahead of a company with an agent nobody trusts. The ten workflows prepare real work every week, each one has an owner, and each one is a written method ready to be run with less attention when it earns it. The agent nobody trusts gets checked so closely that it saves nothing, or gets switched off.
The talk about agents is mostly talk about where the work is going, and the direction is right. Software will carry out more steps by itself. The companies that get there first are the ones that wrote their methods down, tested them and kept them current. That is disciplined, focused work, and it starts with a plan. A plan before agents sets out what that plan holds, and The stages of AI at work shows where workflows and agents sit on the path.
Five signs a workflow is ready to run with less attention.
A workflow is ready to run with less attention when it shows all five signs below. Copy the list and check one workflow against it at the monthly review.
- It has run on real work for a sustained period.Weeks of live cases, run by the team that owns the work, not a demonstration on a handful of examples.
- Its exception log has gone quiet.The reviewer has stopped finding new kinds of case. Every exception the log recorded is now handled in the instructions or sent to a person by rule.
- It passes a written test set.The test cases cover the normal case and every known exception, and the workflow passes all of them after every change.
- Its inputs arrive the same way every time.The records it starts from come from one system, in one format, without someone tidying them first.
- A miss is caught before it leaves the building.Anything that sends, pays or changes a record still waits for a person to approve it, and the owner can say who that person is.
A workflow that shows all five signs is a candidate for the next step: a scheduled run, a connection that pulls from your systems, or an agent that carries more steps by itself. How to schedule repeatable business routines covers the first of those. A workflow that misses any sign stays where it is, with a person starting it and checking it, until it shows that sign too.
One company deciding what to automate.
Illustrative example.
Take a 45-person property management company. The owner hears a pitch for an agent that answers tenant maintenance requests and books contractors by itself. Instead of buying it, the operations lead puts maintenance on the plan as one team with three pieces of repeating work: logging each request, sorting it by urgency and preparing the contractor brief.
The team builds a workflow for sorting requests by urgency first. The maintenance coordinator runs it on every request and checks each result. In the first weeks the exception log fills: requests that mention a smell of gas, buildings with their own contractor rules, tenants who report two problems in one message. Each one goes into the instructions and the test set.
At the monthly review, the operations lead checks the sorting workflow against the five signs. Four are met. The fifth is not: requests arrive by email, phone and a web form, so the inputs are not the same each time. The team decides to bring every request through the web form before anything runs with less attention. The contractor brief stays a workflow a person starts and checks. When the agent pitch comes back, the company has a written method, a test set and an exception list to hold it against.
What is different in a company of under 200 people.
Most advice about agents is written for large enterprises with a platform team, a security review board and engineers to build and watch the software. A company of 20 or 150 people decides differently.
- The person who checks the work is close to the work.In a small company the reviewer often does the job as well. A workflow puts that person’s judgment into writing, which is the method an agent would need.
- One bad automated action is visible.A wrong email to a client or a wrong booking reaches the owner by the afternoon. A person approving each action is cheap insurance at your size.
- You already pay for the assistant.Workflows run in the ChatGPT, Claude, Copilot or Gemini seats you have, which is how 10X the ROI of your existing AI begins. No new platform comes first.
- You can cover every team.A large company pilots one agent in one division. You can give every team its first workflow and learn where more automation fits across the whole company.
Where to start.
Start with the plan: your teams, the work each one repeats and the first workflow for each. How organizations plan their use of AI sets out the three steps: plan it, develop the workflows, deploy them. FullyPowered Planner is where the plan lives, with your teams, their roles and the work between them, and the order you build in. Free workflow templates in the Workflows Directory are a starting point, and we develop workflows with you in a 3-Day Workflow Sprint or a 2-Week Workflow Sprint. Once workflows are in use, How to deploy AI workflows to a whole team, and keep them working covers the rollout.