Why Hapex runs on agents, not n8n templates.

Every founder I've talked to who tried automating their work first tried Zapier or n8n. Most of them gave up within six months. Not because the tools were broken. Because the work wasn't shaped like a template.

This post argues that agent-based automation isn't a fashion upgrade over template-based no-code. It's a different category. And for any small business workflow that has even one edge case, the difference is the entire point.

The template assumption

Template tools work like this: you draw a flowchart. "When a new row lands in this Google Sheet, take column A, paste it into ChatGPT with this prompt, take the result, drop it into column E, then send an email." Every run, every row, follows the same nine boxes. No exceptions.

This is fine when the work is genuinely identical every time. Most work isn't.

Take a workflow most of our customers have: customer support inbox triage. The template version goes: "for each unread email, ask GPT if it's a refund request, if yes route to billing@, otherwise route to support@." This breaks the moment a customer asks two things in one email (refund AND a technical question). It breaks when someone sends a refund request that's actually a fraud attempt and needs to go to security@. It breaks when a quote inquiry comes in formatted as a refund question because the customer was confused.

You can keep adding branches. Each branch needs another author, another test, another maintainer. After six months, the template has 40 paths and the original author left the team. Now nobody knows what happens when none of the paths match. Probably nothing. Probably a customer is waiting.

The agent assumption

An agent doesn't have paths. It has a job description. "Read each unread email. Decide what it needs. Route or draft accordingly. When uncertain, flag for human review with a one-line note."

That sounds vague because it is. The model decides per-message. And because the model can read the full context (the customer's history, the products you sell, your refund policy, your tone), it can handle edge cases that no template author would have predicted.

When something genuinely new shows up — a category of email you haven't seen before — the agent doesn't break. It flags. You see the flagged item, decide what should happen, and (on Hapex specifically) the agent learns the pattern for next time without anyone editing a flowchart.

Where the economics actually live

Templates feel cheaper because they look cheaper. $20/month for Zapier vs. $49/month for a Hapex Operator on Plus. But the real cost of a template is what happens after it ships.

If a template gets 20% of cases wrong (a generous estimate for most non-trivial workflows), every wrong case is a customer cleanup. Wrong refund routing means the billing team rejects something they shouldn't have. Wrong support routing means a question waits two days. The cost of those mistakes — in customer churn, in lost revenue, in the human time spent cleaning up — is almost always larger than the price difference between the cheap template and the agent.

The 80% case isn't where the value is. The 20% case is. Templates run the 80% on autopilot and fail silently on the 20%. Agents handle the 20% by reasoning per-case and flag what they can't.

What this looks like in practice

Three concrete examples from our use-cases:

Email triage. A template would route by sender domain or subject-line keyword. Our agent reads the body, classifies intent (urgent / routine / FYI), drafts replies for routine inquiries based on your past replies to similar messages, and surfaces only what genuinely needs you. The output is one digest, not a sorted folder. See the mockup.

Lead research. A template would pull the company's LinkedIn description and paste it. Our agent reads the company website, recent press, LinkedIn, and any relevant news mentions, then synthesizes a one-paragraph brief plus a personalized outreach draft tied to something specific (recent expansion, hiring activity, product launch). See the column F mockup.

Customer follow-up. A template would send the same "checking in" form letter to anyone past N days. Our agent reads your last conversation with that customer and writes a follow-up that references it concretely. "Hope the Cincinnati rollout is landing" is not a template variable. It's the agent remembering what Sarah told you in September. See the mockup.

When templates still win

I'll be honest: there are workflows where templates are the right answer. Pure data movement (Stripe webhook → Sheet row → Slack channel) doesn't need reasoning. Notifications that fire on a clear condition (low inventory, server down) don't either. If your workflow is genuinely "every X, do exactly Y, no exceptions," n8n or Zapier are cheaper and just as reliable.

The mistake most teams make is starting with that assumption for workflows that aren't actually like that. The inbox triage example above looks like "for each email, do Y" until you actually try it. Then the edge cases start eating your weekends.

How to tell which one you need

One question: can a competent new hire learn this job by reading a flowchart, or do they need to read past examples and use judgment?

If a flowchart suffices, use a template. They're great at flowcharts. If the answer is "they need to read examples and use judgment," that's an agent.

If you're not sure, the cheapest way to find out is to write the workflow as a paragraph of plain English. If the paragraph has any sentence containing "depending on," "unless," "based on," or "if it's clearly," it's an agent.

If you want to try it

Open build.hapex.ai, describe whatever you've been trying to automate in your own words. The Operator wires up the connectors and routes each job to the right model automatically, runs a test against your real data, and shows you what activation would cost. No card required to test.

See pricing if you want to know what each plan costs first.

— Shameel Khairi, Hapex AI