What an AI automation actually is, explained without the vocabulary

If you have been sent a proposal full of words like workflow, agent, n8n and RAG and would like to know what you are buying before you pay for it, this is the page. It explains the pieces in the order a business owner meets them, using the systems we run rather than the ones in the brochures.

A workflow is a list of steps a computer runs

Start here, because most of what is sold as AI is this. A workflow is a fixed sequence: when a form is submitted, look up the customer, write them into the CRM, send the confirmation, notify sales. It runs the same way every time, and when it breaks you can see which step broke. It has been possible for years and needs no AI at all. The tools people use to build them, Zapier, Make and n8n, are visual editors for exactly this: boxes for steps, arrows between them. Which one fits a given business is a question of cost at volume and who owns the server, and it is rarely the reason anything works or fails.

The AI is one step in that list

A language model is the part that can read, sort, draft, extract and answer. It reads an invoice and pulls out the supplier, the amount and the date. It sorts an inbox into five folders. It drafts a quote from your price book. It answers a question from your documents. It listens to a caller and talks back.

What it does not do, in any system we would put our name to, is decide. The workflow decides what happens next. Every step is logged. Anything that touches money or goes out to a customer waits for a person to approve it, until the error rate has been measured on real work. So when the model reads an invoice wrong, the wrong number lands in a review queue with the reason it was flagged, not in your accounting software. When the phone agent is unsure about a booking, it takes a message rather than guessing a time.

The systems we run that have never woken anybody up at night are the boring ones. That is on purpose.

What "agent" means, and when to be suspicious

An agent is a model that is allowed to choose its own next step: which tool to call, whether to look something up, when it is done. That is impressive in a demo and expensive at three in the morning when it chose something nobody expected. Most things sold as agents are workflows with one model step in the middle, and that is usually the right design. If somebody pitches you an agent and cannot tell you what it does when it is wrong, they have not run one in production. The question is worth asking about anything with a model in it, and it belongs with the other questions to ask before hiring an agency.

Why it goes wrong, in plain terms

The model is rarely the cause. In the systems we run and the broken ones we are called into, the failures are dull: a login expired and nothing alerted anyone, a step assumed a field would never be empty and one day it was, a spreadsheet got a new column and everything shifted by one, a workflow finished green every day while processing zero rows. A demo never meets any of that. Production meets all of it in the first month. The six usual causes are written up in why AI automations die in production, and the list of things a system needs before it counts as finished is the production-ready checklist. You do not need to understand the checklist to use it. Hand it to any vendor and watch how they react.

What you should own

Every account the system runs on, in your company's name from day one: the automation platform, the model provider, the phone number, the hosting, the database. The builder is invited in as a user. If the builder disappears, the system keeps running and anyone you hire next can log in. The most common call we get is not "build us something". It is "the person who built our automations is gone and every login was theirs". What to do when that has already happened is its own page, and the way to make sure it never does is to put ownership in the contract before the first build.

What it costs to keep running

There are two prices. The build is the one in the proposal. The second is the monthly bill for as long as the system runs, and it is usually small: the platform, the model usage, phone minutes if there is a phone, and a little hosting. A single workflow tends to land at $40 to $150 a month, all of it on your own accounts where you can read the meters. What makes that bill spike, and how to cap it, is in the running costs nobody mentions. A vendor who marks up those meters earns more when your system wastes more, which is the clearest warning sign in a proposal.

How to buy it

If you already know the one job you want gone, describe it in two paragraphs and get a build quoted against a written scope. If you suspect there is money on the table and cannot say where, start with a free consultation: 30 minutes on how the work runs today, then a written proposal with a fixed price. What a paid audit buys you covers the deeper version other agencies sell, and when to skip it. Either way, start with one job, get a price in writing before anything is built, and make sure the accounts are yours. How to spot the first job is the companion to this page.