# What I look for before automating anything.

> How I decide which workflow to automate first, and which ones to leave alone. A first-person view from production automation work and 13 years in enterprise IT.

Canonical: https://alexcinovoj.com/services/ai-automation-consultant

Most automation projects fail at selection, not implementation. Here is the filter I run before writing a single node.

Written from production automation work and 13 years of watching enterprise processes break in the same three places.

## The four questions that decide the first build.

### Is the process actually defined?

If two people describe it differently, there is nothing to automate yet. Write it down first, then decide.

### What does a wrong answer cost?

Cheap, reversible errors are ideal first candidates. Irreversible ones need review gates before they need automation.

### Who owns it after launch?

Name a person. Automation without an owner rots quietly until the day it matters.

### Can I measure the before?

No baseline, no way to prove the automation earned its place. I capture the manual numbers first.

## The work that consistently pays off.

Lead qualification, support triage, document extraction, system-to-system sync, reconciliation, and internal question answering over documents. All high volume, all low individual stakes, all reversible. That combination is why they survive contact with production.

## What I talk people out of.

Automating an approval nobody trusts. Putting a model in front of data nobody has cleaned. Replacing a person whose real job was judgment. And building anything on top of an operations layer that is already at its limits, because that ceiling becomes your ceiling.

## Boring tools, sharp glue.

n8n for orchestration, Claude and OpenAI for reasoning, MCP for tool access, Postgres with row-level security for state, and tracing that lets me replay any failure. Nothing exotic. Exotic is how systems become unownable.

## Common questions.

## If the workflow is picked and the system will not hold, that is the production path.

Hardening, evals, permissions, and handoff run through Systems Decision Audit at TechTide AI, where the scope and pricing live.