Do you need an AI agent, or a workflow with one AI step?
Most small businesses don't need an AI agent. If you can write the job down as steps in order, build a fixed workflow and give AI the one step that needs judgment, like sorting an email or drafting a reply. Pick an agent when the steps change with every case and a wrong move is cheap to catch.
You know when a demo shows an AI agent reading an email, updating the CRM and writing the reply, all on its own? It's tempting to want one for every task. Most work in a service business already follows known steps, though. A lead comes in, gets sorted, gets a reply and lands in the CRM, and that order doesn't change from one lead to the next.
The difference is who decides the next step. In an agent, the AI model picks which tool to use and what to do next, and keeps going until it decides the job is done. In a fixed workflow, the steps are set in advance, and AI does one bounded job inside it, like sorting an email into a category, pulling the details out of a message, or drafting a reply. Anthropic's guide to building agents draws the same line, and its advice is to find the simplest setup that works, which might mean not building an agent at all.
| Question | AI agent | Workflow with one AI step |
|---|---|---|
| Does it do the same thing each time? | Not always. The model picks its steps, so two similar emails can take different paths. | Yes, apart from what the AI step writes or decides. The steps around it run in the same order every time. |
| What one run costs | It varies with how many steps the model takes, because each one is another AI call. Anthropic's guide says agents often trade speed and cost for better results. | Close to the same each run, because the number of AI calls is set in advance. |
| How you test it | Run whole tasks in a safe copy of your tools and check every path it took. Anthropic recommends extensive testing in sandboxed environments, with guardrails. | Test the AI step alone, many times, on the same sample inputs, then test the fixed steps like any other workflow. |
| What happens when it fails | A wrong early choice feeds every later one. Anthropic's guide names compounding errors as a risk that comes with an agent's independence. | The failure sits at one named step. The run stops there, and an alert can say which step it was. |
| Who can check what it did | Someone technical, reading the model's choices one by one in a detailed log. | An owner can read the workflow map and follow each step. Only the AI step's output needs judging. |
| Jobs it suits | Open-ended ones, where the next step depends on what it finds, like researching a company across several sources before a call. | Jobs you could write down as steps: sort the email, pull out the details, draft a reply, log it, ask a person. |
Two systems I built, run on my own accounts, are fixed workflows with bounded AI steps. In lead to reply, a single AI call scores the lead and drafts the reply, and on one test lead on 2026-09-16 that call cost about a hundredth of a cent. On one test inquiry run live on SignalNine's own inbox on 2026-09-13, the inquiry desk made three AI calls, one to sort the email, one to pull out the details and one to draft the reply, and together they cost about a cent. In both, the order of the steps never changes, and I approve every reply before it goes out.
Where each one fails
An AI agent
An agent fails when the job has a right answer and a wrong one that costs money. Every choice the model makes is a place it can choose wrong, and the next step builds on that choice. OpenAI's guide to building agents says sensitive, irreversible or high-stakes actions should trigger human oversight, and Anthropic's recommends testing in sandboxed environments with guardrails. For an owner, that means someone technical watching it closely for a while, and a person approving anything a client will see.
It also fails on the bill when it wanders. Each extra step it takes is another AI call, so a confused run costs more than a clean one, and you only see that after it has run.
A workflow with one AI step
A fixed workflow fails when the work doesn't fit the steps. An email that is half a new lead and half a complaint gets sorted into one category, and the workflow carries on as if that category were right. Each new kind of case means someone adds a branch, and a workflow with too many branches gets hard to change. OpenAI's guide lists rules that have become hard to maintain as one sign a job may suit an agent.
It also fails at the AI step itself. A bounded step can still be wrong, so whatever comes after it needs a check. An email the step can't place goes to a person, and a drafted reply waits for approval before it reaches a lead.
What I'd pick, per situation
- If you can write the job down as steps in order, pick a fixed workflow, even when one step needs judgment. Give AI that step and nothing else.
- If the AI's only job is to sort, pull out details or draft text, pick a fixed workflow. In n8n, the tool I build on, the AI Agent decides which tool to use for the task, while the Text Classifier sorts text into categories you name and the Information Extractor pulls structured details out of it.
- If the steps depend on what the AI finds along the way and a wrong move is cheap to undo, an agent can earn its cost. Researching a new lead's company before a call, or answering a staff question from your own documents, are fair cases.
- If it would act on clients, money or records that are hard to fix, keep a person between the agent and the action, or use a fixed workflow.
- If you're not sure, start with the fixed workflow. If one step keeps meeting cases the rules can't handle, you can try an agent on that step alone.
Questions
What is the difference between an AI agent and an automation?
In an automation, or workflow, the steps are set in advance and run in the same order. An agent gets a goal and some tools, and the AI model decides which tool to use next and when it's finished. A workflow can still use AI for one step, like sorting an email, without becoming an agent.
Do I need an agent to use AI in my business?
Usually not. Most of the useful AI work in a service business is one step inside a workflow: sorting incoming email, pulling details out of a form or a document, or drafting a reply that a person approves.
Is an AI agent more expensive to run?
Usually, because each step it chooses is another AI call, and the number of steps changes from run to run. Anthropic's guide says agents often trade speed and cost for better results. A fixed workflow makes the same number of AI calls each time, so its cost per run stays close to steady.
When is an AI agent the right call?
When the steps depend on what it finds, the inputs vary a lot, and a wrong move is cheap to catch and undo. OpenAI's guide points to complex judgment, rules that have become hard to maintain, and a lot of unstructured text. Start it on internal work before anything a client sees.
How do I test an AI step before I trust it?
Run the AI step alone on a set of sample inputs, many times each, and count how often it gets one wrong. A rare mistake won't show up in a handful of runs. Then test the rest of the workflow like any other, and keep a log of every run.
- Anthropic, Building effective agents: workflows orchestrate models and tools through predefined code paths, while agents dynamically direct their own processes and tool usage; find the simplest solution possible, which might mean not building agentic systems at all; agentic systems often trade speed and cost for better task performance; the autonomy of agents means higher costs and the potential for compounding errors, so test extensively in sandboxed environments, with guardrailsREAD
- OpenAI, A practical guide to building agents: agents suit complex decision-making, rules that have become difficult to maintain, and heavy reliance on unstructured data; before committing to building an agent, validate the use case against these, otherwise a deterministic solution may suffice; sensitive, irreversible or high-stakes actions should trigger human oversightREAD
- n8n docs, AI Agent: the agent determines which tool to use depending on the task, and needs at least one tool connectedREAD
- n8n docs, Text Classifier: classifies incoming data into the categories you provide, each with a name and a descriptionREAD
- n8n docs, Information Extractor: extracts structured information from incoming dataREAD
Related
- Should a chatbot answer your leads, or should you approve each reply?A chatbot that answers leads on its own, or an AI-drafted reply a person approves first: speed, what each can promise, what happens when it's wrong.
- n8n vs Zapier vs Make: which should a service business run on?What each one bills for, who can change a workflow, what happens when a step fails, and where n8n, Zapier and Make each fall short for a small service firm.
- Human-in-the-loop approvalWhat human-in-the-loop approval means in an automation, which steps should wait for a person, and how an approval card makes sure an email sends once.
- What a reliable automation looks likeFive properties any automation should have, checkable by an owner who doesn't build: logged runs, named alerts, one send per approval, safe reruns, an owner.
- Website form leads, scored and answered with a reply I approveA system I built: each website form lead is checked, added to HubSpot once and scored. Gmail sends the drafted reply when I approve it in Telegram.
- A drafted reply to each lead, approved from my phone before it sendsAI drafts replies to lead emails that pass a junk screen, from my business facts. Risky asks get flagged, and nothing sends until I approve it on my phone.
- WebhookA webhook is a message one app sends another the moment something happens. What that means in plain English, how it fails, and how a system catches it.