Most businesses that tried automation have the same story. They built some workflows. The workflows worked. Then reality happened, the workflows started producing wrong results in edge cases, someone quietly went back to doing it manually — and now they maintain both the manual process and the automation that was supposed to replace it.
This is not a failure of Zapier, Make or n8n. Those are excellent tools and we use them daily. It is a failure of the model underneath them.
What Fixed Automation Actually Is
A traditional workflow is a chain of deterministic steps: when X happens, do Y, then Z. It executes exactly what you configured, every time, without deviation. That is its greatest strength and its hard limit.
It works perfectly while the input matches what you anticipated. The moment reality deviates — a missing field, an unexpected reply, a request nobody wrote a rule for — the chain either fails loudly or, far worse, succeeds incorrectly and writes bad data into your systems.
The 80/20 That Defeats Automation Projects
Here is the pattern we see constantly. A business maps a process, finds that 80% of cases follow a clean path, and automates that path. Genuine progress.
Then the remaining 20% turns out to contain 80% of the actual labour — because exceptions are exactly the cases that require someone to look, think and decide. So a human still monitors the queue, still handles the awkward cases, still checks that the automation did the right thing. The headcount does not drop. And now there is a workflow to maintain as well.
That is why so many automation projects produce a fraction of their promised return. Not because they were badly built, but because the hard part was never automatable with fixed rules.
What an Agent Does Instead
An agent is not given a script. It is given a goal, a set of tools, and boundaries. It works out how to reach the goal, adapts when the situation is not what it expected, and escalates when it hits something it genuinely should not decide alone.
Concretely, in the cases that break chains:
- A form arrives with a missing field. Fixed automation fails or writes a broken record. An agent asks the customer for the missing detail and continues.
- A supplier replies with a question instead of a confirmation. Fixed automation has no rule and nothing happens. An agent answers it, or escalates with full context.
- An unusual request arrives. Fixed automation drops it into a queue nobody watches. An agent reasons about it and either handles it or routes it to the right person.
- The process changes. Fixed automation requires rebuilding the chain. An agent needs its goal, policy or tools updated — and adapts.
The Right Mental Model: Hands and Judgement
This is not an argument for abandoning your automation platform. n8n, Make and Zapier remain the execution plumbing — reliable, observable, cheap connections between systems.
What changes is what sits above the plumbing. Think of the automation platform as the hands and the agent as the judgement. The agent decides what should happen given the situation; the platform reliably carries it out. Most well-built agentic systems we deliver run on exactly this arrangement, and the automation investment a client already made keeps earning.
When Fixed Automation Is Still the Right Answer
Plenty of the time, and it would be dishonest to pretend otherwise. If a process is genuinely deterministic — same input shape, same steps, no judgement, no exceptions — a fixed workflow is cheaper to build, cheaper to run, easier to debug and completely predictable. Syncing a form submission into a spreadsheet does not need reasoning.
Use agents where variability lives. Use rules where it does not. Anyone recommending agents for everything is selling, not advising.
How to Tell Which One a Process Needs
Ask two questions about the process.
First: how often does something arrive that the rules do not cover? If the answer is essentially never, use fixed automation. If your team regularly says "this one is different, let me handle it manually", you have found a case for an agent.
Second: does correctness require understanding the content, or only its shape? Moving a file based on its name is shape. Deciding whether a customer's message is a complaint, an order change or a cancellation is content — and content is where fixed rules become an ever-growing pile of if-statements that nobody can maintain.
The Practical Upgrade Path
You do not throw away what you built. In most engagements we start by looking at which existing workflows have the highest failure and manual-intervention rate — the team usually knows exactly which ones without checking — and place an agent at the decision point that keeps breaking. The rest of the chain continues to run as it always did.
That is a small change with a disproportionate effect, because the exceptions were consuming most of the human attention in the first place.
The Bottom Line
If your automation only works when everything goes perfectly, you have not automated the process — you have automated the easy part of it and kept the hard part on a human. Agents are how the hard part finally moves.
Harzotech builds agentic workflow systems and works with the automation stack you already have. Book a workflow audit and we will tell you which of your processes need agents and which are better left as rules.