Use a simple automation when the steps are the same every time. Use an AI agent when each case needs reading and the next step depends on what is in it. Either way, add a log, a named person for unusual cases and an off switch. Start with the simplest option that does the job, then add autonomy only where it earns its keep.
That is how we scope work in our automation and AI agents service. We often find the right answer is smaller than the first idea.
What is the difference?
An automation is software that does a repeat task the same way every time. A new row in a form goes to a sheet, an email goes out, a task is created. Nothing is decided.
An AI agent is software that is given a goal, reads each case, picks the next step, uses your tools, and hands the unusual cases to a person. It is not a chatbot: nobody types a question to it. An agentic workflow is a process built from agents. In short, an AI agent is automation that can cope with cases nobody planned for.
Anthropic, in its guide Building effective agents, draws the same line. It calls systems where code decides the path “workflows”, and systems where the model directs its own steps and tools “agents”. It recommends finding the simplest solution possible and adding complexity only when needed, and notes that agentic systems often trade speed and cost for better results on the task.
How do you decide?
Ask four questions about the process:
- Are the steps the same each time? If yes, build an automation. It is cheaper, faster and easier to test.
- Does each case need reading? Emails, documents and support requests all differ. If the next step depends on what is written, an agent may fit.
- What does a mistake cost? A wrong tag on an internal ticket is cheap. A wrong reply to a customer or a wrong figure in a report is not. The higher the cost, the more review you need, and the narrower the agent’s job should be.
- Can you describe a right answer? If you can write down what a good outcome looks like and check it against past cases, you can measure an agent. If you cannot, you are not ready to build one.
What does that look like in practice?
These are illustrations, not client stories.
- Sending a signup to your CRM and a welcome email. Fixed steps. Automation.
- Sorting inbound emails into sales, support and spam, and drafting replies. The content varies. An agent can read each message, with a person approving the replies.
- Pulling the same figures into a weekly report. Fixed. Automation.
- Checking a document pack for missing items. The documents differ and the checks depend on content. An agent, with unclear cases sent to a named person.
For regulated work, the bar is higher. Our agentic workflows service page shows how we treat review and escalation.
What should surround any agent?
Every agent we build has the same safeguards:
- A log. Every step it took, so you can see why.
- An escalation route. A named person a case goes to when the agent is unsure.
- An off switch. One action that stops the workflow.
- A scored test set. A set of your own past cases we run the system against before launch, so its accuracy is a measured number, not a promise.
- Shadow mode. We run the agent beside your team without letting it act, then compare its decisions with your team’s on the same cases.
- Drift checks. Results can get worse when the world or the AI model changes. We re-run the tests and raise an alert before your team notices.
What do we do first?
We map how the work really runs, which is step one of our method. Some of the engagements we scope end with a process change and no build, because the process was the problem. A tidy process often needs only an automation.
Questions people ask
Is an agent always better? No. It costs more to run and test than a fixed automation, and it is harder to predict. Use one where the variety in the work is real.
Can an agent act without a person? Only in steps where the risk is low and the test results support it. Anything that publishes or sends to a customer is reviewed by a person first.
Where do we start? With one process that repeats often, has clear past examples and has a named owner. Send the brief and we will tell you whether it needs an agent, a simple automation, or neither.
