Process automation: a field guide for operators

The first process automation I shipped in 2022 wrote the wrong customer name to the wrong invoice for six straight days before anyone noticed. The workflow ran clean, the API returned 200, the logs were green. The trigger payload had two fields with the same case insensitive key and I picked the wrong one. That was the lesson. Process automation does not eliminate human judgment, it concentrates it at the design layer. Everything below is what we actually do at Moore IQ when an operator asks us to take a manual process off their plate.

What process automation actually means

Process automation is the practice of replacing a repeated human workflow with software that performs the same steps reliably and on schedule. The term covers a wide range. At the simple end it is a single trigger that copies one piece of data to another system. At the complex end it is a multi step pipeline that ingests unstructured input, makes a judgment call, writes to several systems of record, and notifies a human only when something falls outside the rules.

Most operators get confused because the marketing definitions span everything from "macro record a spreadsheet" to "deploy autonomous agents". The version that pays back is narrower. It is a documented manual process, with a clear input and a clear output, that someone is paid to do more than three times a week.

The three kinds that actually pay off

We have run this matrix across roughly 500 production workflows. Three categories generate almost all of the return.

Boundary handoffs. A record exists in System A and needs to land in System B with some field mapping in between. CE21 to GoHighLevel. Forms to QuickBooks. Calendly to a CRM contact. These are unglamorous and they pay back in a week. The pattern is always the same. Trigger, transform, upsert, error notify. If your team currently does this with copy paste or a CSV export, this is the first thing to automate.

Inbox triage. A bucket of unstructured input (email, chat, voicemail transcripts, form free text) needs to be classified, routed, and have a draft response written. The judgment used to require a human. With Claude or another modern model behind a prompt with a few examples, the model is right enough to draft and a human approves. Approval shrinks from twenty minutes to fifteen seconds. The Anthropic API docs cover the basics if your team has not used it before.

Scheduled rollups. A team needs the same report every Monday morning, pulling from three or four systems. The data exists, the format is consistent, the audience is known. Automating this removes a recurring tax on the operator who used to assemble it manually.

Anything outside these three categories is harder to scope and slower to pay back. We do build them. We will tell you if it is worth it.

Where the money actually goes wrong

The two failure patterns we see most often:

Automating before documenting. If the manual process is undocumented and the person doing it cannot describe the rules they apply, the automation will codify the wrong rules. Spend the first hour watching the operator do the task and writing down every branch, including the ones they do unconsciously. The cost of skipping this step is rebuilding the workflow three times. This is the tool problem versus architecture problem split we wrote about separately.

Buying the wrong tool first. Most teams pick a vendor before they pick a process. They sign up for Zapier or Make or Power Automate because they heard a podcast, then they try to bend their process to the tool. The correct order is the other way around. Pick the process, write the data flow on paper, then pick the tool that matches the flow. We compared the three platforms in n8n vs Zapier vs Make for 2026 if you want the side by side.

The stack we ship with

Our default for new client work is n8n self hosted on a VPS the client owns. The opinion, unsoftened: we prefer n8n over Zapier and Make because you can self host it, you can version the workflows in git, you can see every node execution, and you do not pay per task. We have built workflows that would cost an extra few hundred dollars a month on Zapier and run for the cost of the VPS we already maintain. The full cost picture lives in the n8n self hosted versus cloud breakdown.

For AI judgment inside the workflow we default to Claude (Sonnet for most tasks, Haiku for cheap routing). When a workflow needs to talk to a system that has no good API, we either build a custom MCP server or fall back to a scraping marketplace. When state needs to persist across runs, we use Supabase because pgvector and the dashboard story beat Airtable for anything past 10,000 rows.

This stack saves clients 300 plus hours a week in manual tasks when it lands well. It is not magic. It is the result of designing the workflow to match the operator, not the other way around. The full n8n documentation is worth a careful read before any team commits to it.

A real example

Last quarter we shipped a process automation for a law firm operator who needed an answer system on top of years of case documents. Verbatim from the Loom we recorded after the build:

I built a RAG system for law firm documentation to ask questions about it with referencing to the actual PDF that you are working with.

The flow had four moving parts. A document ingest step that watched a shared drive. An embedding step that wrote to Supabase using pgvector. A query interface that took a natural language question and returned a grounded answer. A citation layer that included the source PDF and the page number. The first version was wrong about citations one in five times. The third version was wrong less than one in fifty times. The difference was a re ranking step we added after watching three real lawyers use it.

The point of the example is not the stack. It is that we shipped something useful, then shipped something better after we watched the work get done. That is the loop. The longer write up on the law firm specific patterns is in n8n workflow automation for law firms.

How to scope your first process automation

Five questions, in order:

1. What is the input, and how does it arrive today?
2. What is the output, and what system owns it?
3. Who currently does this work, and how long does it take per occurrence?
4. How often does the same step go wrong, and what does the operator do to recover?
5. If the automation fails silently, what is the worst case for the business?

Question five is the one most teams skip. The answer determines how much monitoring and human approval you need to design in. An invoice classifier that mis tags a vendor is recoverable. A wire payment trigger that fires on a wrong condition is not. Microsoft has a clean version of this scoping rubric inside their Power Automate guidance even if you do not end up using the platform itself.

When this is not for you

If your process changes meaningfully every two weeks, do not automate it yet. Stabilize it for a quarter, then automate the stable version. Automating a moving target produces a workflow you have to rebuild constantly.

If your team has fewer than three people and the process happens twice a week, the math probably does not work. The build time and the maintenance load are not free.

If the only reason you are buying automation is because a competitor announced they did it, talk to the operator first. They will tell you whether the work is actually worth automating. We have walked clients away from builds in the first call when the answer turned out to be "your team needs documentation, not software".

What this looks like end to end

A first engagement with us usually goes:

1. Sixty minute discovery on the highest payback process you have today.
2. A two page scope doc with the workflow drawn out and the build estimate.
3. Build in a one to three week window depending on integration count.
4. A handoff Loom that shows your team how to run it and how to modify it.
5. Optional quarterly retainer for changes and new builds.

If you want us to look at your own process queue, the automation priority scan gives you a ranked top five back in about three minutes. If you want to talk it through, book a build call.