AI agent or automation? When a Zap is enough and when you need an agent

Over the past year you cannot open LinkedIn without someone explaining why you need an AI agent. Salesforce has Agentforce. Make shipped a new generation of AI agents, Zapier sells enterprise ones, and the OpenAI API lets you roll your own. Every other conference talk opens with "agents are the future of work."

Maybe. But most of the mid-size company CEOs I talk to do not need an agent at all. They need automation that someone bothered to configure properly. The distinction is worth money, and it decides whether anyone can still maintain the thing six months later.

Automation and agents are not the same tool

Automation is "if X happens, do Y"

Zapier and Make rest on a single idea. You define a trigger and an action. An invoice email arrives, so the attachment lands in Google Drive and a notification goes to Slack. An order status changes in the CRM, so the accounting record updates.

Deterministic and predictable, which is exactly the point. You know in advance what will happen, and once it is set up it runs for months untouched.

An agent picks its own steps

An agent works differently. It gets a goal rather than an instruction, along with access to tools, data and history, and it works out for itself which steps get it there.

Tell it to handle a customer complaint and it reads the message, pulls the customer's history from the CRM, weighs how serious the problem is, and decides whether to answer or hand off to a person. If it answers, it picks tone and content from context.

One sentence covers the difference. Automation executes instructions and an agent makes decisions. Decisions carry risk, and somebody has to own that risk.

Choosing between them

Ten months of deploying both for clients left us with a simple decision framework. We reach for it even when a client opens the conversation asking for an agent.

Structured data and clear rules mean automation

If your inputs have a predictable format and the rules fit into if/then logic, skip the agent. Automation costs less, deploys faster, and survives without a specialist babysitting it. CRM and accounting sync, status-driven notifications, documents generated from templates, scheduled database reports. None of that needs an agent.

Unstructured data and judgment mean an agent

When inputs arrive in whatever format the sender felt like using, in natural language, and someone has to exercise judgment to process them, an agent earns its keep. Think complaint analysis, classifying documents by what is actually in them, answering awkward questions out of a knowledge base, or sizing up supplier proposals.

Hybrids win in practice

Here is the interesting part. Real solutions are rarely a pure agent or pure automation. They are automation workflows that call AI at the decision nodes. Make or Zapier drives the flow, AI decides where deciding is needed, and everything else stays deterministic. You pay for intelligence only in the handful of steps that need it.

Three practical examples

Email triage with Zapier and GPT classification

The sales inbox takes more than 200 messages a day, and the sales team burns an hour daily just sorting them.

Zapier catches each new email and sends it to the OpenAI API, which drops the message into one of five buckets. New inquiry, existing deal, spam, complaint, billing. Zapier then routes the mail to the right Slack channel, tags the record in the CRM, or answers on its own when the message is spam or the question is obvious.

A full agent would be overkill here. The flow never varies, and AI does exactly one job, which is classification. Everything downstream is deterministic, so there are fewer errors and a smaller API bill.

A helpdesk agent with a knowledge base and escalation

Support answers the same 80% of questions over and over. The answers still need context though, meaning the customer's plan, the history of the conversation, and whatever is currently on fire.

So we connect an AI agent to the knowledge base, the CRM and the monitoring system. A customer asks something, the agent finds the relevant docs, checks that customer's history, and judges whether it can answer alone. If it can, it does. If it cannot, it opens a ticket with the context attached and routes it to the right person.

Automation cannot do this job, because the questions are unpredictable and the context keeps shifting. The agent has to understand meaning rather than match keywords. The part that matters most is the boundary. The agent knows exactly when to escalate, and it never rules on refunds, contract terms or technical interventions by itself.

Reports with Make and AI summarization

Management wants a weekly performance overview. The data sits in three different systems and an analyst loses every Friday to assembling it.

A Make workflow pulls from the CRM, accounting and the project management tool every Friday at 6:00 AM. It aggregates everything into structured JSON and hands that to the OpenAI API, which writes an executive summary covering key metrics, trends and recommendations. The result goes out by email and to Slack.

There is nothing for an agent to do here. The data is structured and the process repeats identically every week. AI summarizes and interprets, it does not decide.

Anti-patterns to avoid

An agent for everything

The most common mistake by far. A company reads about agents and wants one on every process. What it gets is a large API bill, unpredictable behavior, and bugs nobody can trace. Use a simple rule of thumb. If the process draws as a flowchart with no decision diamonds in it, you do not need an agent.

No human in the loop

The agent decides on its own and nobody reviews the output. Month one is wonderful. Month two it starts producing answers that are not quite right. Month three your customers are writing complaints. Every agent needs a defined escalation path and a threshold where a human steps in.

No logging

When an agent makes a bad call you need to know why. Without a log of the inputs, the reasoning and the outputs, you cannot debug it at all. Log the inputs, the prompt, the model's response, the action taken and the result.

Security is the foundation, not a bonus

Four things hold on every AI deployment we put into production.

No data travels in plain text. Traffic between tools, APIs and databases runs over TLS.

The agent reaches only what it needs. An agent that processes email has no route to financial data, so when it does go wrong, it goes wrong inside a fence.

Every decision the agent makes lands in an audit log. Who, when, what and why. A regulator will ask eventually, and you will need an answer.

That leaves rate limits and cost caps. An agent with no ceiling on API calls is a security risk and a financial one, because a workflow that loops on itself can drain a monthly budget before anyone notices.

An illustrative workflow

Picture a Make workflow with an AI node that sorts incoming orders by complexity. The order lands in the system, the workflow passes it to the chosen model, and the model proposes a category from the item count and any special requirements.

Simple orders (standard products, ordinary address, nothing unusual requested) go straight through to automatic processing. Complex ones (large volumes, non-standard requirements, new high-value customers) get flagged for a human to look at.

This is an architecture example, not a claim about a client outcome. A real pilot has to be measured on real data. Watch the share of orders classified correctly, the number of manual corrections, and processing time before and after deployment.

No agent anywhere in it. A deterministic workflow with one AI decision node and an obvious route to manual review.

Start simple

You do not need an agent for everything. Most business processes automate perfectly well with ordinary tools. Add an AI node where a decision has to be made. Deploy a full agent only where the complexity is real and where the monitoring, logging and escalation infrastructure is already in place.

If you want to work out where AI makes sense in your company and where automation is enough, a two-week AI Sprint answers that. We map the processes, design a solution and hand over a working prototype. Get in touch and we will set up an intro call.