AI agent or automation? A five-minute test for your workflow

An automation follows a path you wrote down: when a form arrives, add a row, send an email. An agent chooses the path: it reads the request, picks which tool to use, checks the result and tries again if it is wrong. Both are useful. Paying for an agent where a workflow would do costs more to build, more to run, and is harder to test.

The five-question test

• If the same input always needs the same action, it is an automation.
• If the input is messy (free text, emails, PDFs, voice notes) but the action is fixed, it is an automation with one AI step to read the input.
• If the next step depends on what the last step found, it leans agent.
• If the person doing it today handles exceptions by reasoning rather than a lookup table, it leans agent.
• If a wrong action costs money or trust, it needs a person to approve it, whichever one you build.

Three examples

• An invoice arrives by email and its numbers go into the accounts sheet. Automation, with AI reading the invoice.
• A customer asks on WhatsApp where their order is. Automation with an AI front end: confirm it is the buyer, look up the order, reply. The path never changes.
• A prospect asks a vague question about a group booking, then changes the dates twice. Agent. It has to ask follow-ups, check availability and decide when to bring in a person.

Why the answer is usually both

The pattern that holds up in production is an agent for the judgement and plain code for the execution. The agent decides what should happen; ordinary, tested code does it, inside permissions you set. That keeps the cost per run down, makes every failure traceable, and means a model update cannot quietly change how refunds work.

Count how many times last week someone did the same task by hand, and how many of those needed a judgement call. The first number tells you what to automate. The second tells you whether you need an agent.

What we do

We build both, and we tell you which one your workflow needs before we quote. Most first projects are a workflow with one or two AI steps. A full agent comes in when the workflow keeps meeting cases the rules cannot handle.