These two terms get used interchangeably by people selling both, which is unhelpful, because they are genuinely different things with different costs and different failure modes.
Getting the distinction right will save you money, because a surprising amount of what gets sold as AI does not need any AI in it at all.
The difference in one line
Workflow automation follows rules you wrote. AI automation makes judgements you did not write down.
That is the whole thing. Everything else follows from it.
Workflow automation
A workflow is a sequence of steps with no ambiguity in it. When this happens, do that. If this field says X, go down this branch.
- A form is submitted, so add a row to a spreadsheet and send a notification.
- A job is marked complete, so wait three days and send a review request.
- An invoice is unpaid after fourteen days, so send a reminder.
- A call rings out, so send a text.
None of those require judgement. A person following the instructions would get the same answer every time, and so does the software.
Workflow automation is cheap, extremely reliable, and it does not make things up. It has been around for decades. If your problem fits this shape, this is what you want, and you should be suspicious of anyone adding AI to it.
AI automation
AI comes in when a step needs a judgement that cannot be reduced to rules in advance.
- Read this enquiry and decide whether it is a real job or a price shopper.
- Summarise what this customer said in one line.
- Write a reply in our voice to this specific complaint.
- Extract the total, date and supplier from this photo of an invoice, which is laid out differently from every other invoice.
- Decide whether this caller is describing an emergency.
You could not write rules covering every case. The input is messy, open-ended language or images, and the useful output requires interpretation.
AI automation is more capable, more expensive to run, and it can be confidently wrong. All three of those are permanent properties, not temporary limitations.
Why the distinction costs money
Two mistakes, both common.
Putting AI where a rule would do. This is the expensive one. Someone sells you an AI-powered review request system. Read the description and it turns out the actual logic is "wait three days, send this message". That is a rule. Adding a language model to it adds running cost, adds a way for it to go wrong, and adds nothing.
I look for this in every build. If a step can be a rule, it is a rule. The AI is reserved for the steps that genuinely need judgement, which is usually one or two steps in a workflow of ten.
Putting a rule where judgement is needed. Less common and less expensive, but it produces the thing everyone hates: a rigid system that handles the normal case and falls apart on anything unusual. Old-fashioned phone menus are the classic example. Six options, none of which is your problem.
Most useful things are both
In practice almost everything worth building is a workflow with a small amount of AI inside it.
Take answering a new enquiry:
- An enquiry arrives. Rule.
- Read it and work out what they are asking for and how urgent it is. AI.
- If urgent, notify the owner immediately. Otherwise queue it. Rule.
- Draft a reply that answers what they actually asked. AI.
- Send it, log it, schedule a follow-up in two days. Rule.
Four fifths of that is rules. The AI does two steps, and they are the two steps that could not have been written in advance.
That ratio is typical, and it is a useful thing to ask about. If someone describes a build as AI-powered and cannot tell you which specific steps use AI and why, that is worth pushing on.
Which one do you need?
Ask yourself one question about the job: could I write down the rule?
If you can write it down, even if the rule is fiddly with a lot of branches, it is workflow automation. Cheaper, more reliable, no running cost to speak of, and no chance of it inventing something.
If you find yourself writing "well, it depends" or "you can just tell", it needs AI. That is fine. Just know that you are buying something with a running cost, a need for testing, and a failure mode where it sounds certain and is wrong.
What this means for what you pay
Workflow automation should be cheaper, because it is less work to build, cheaper to run, and easier to test. AI automation carries the cost of the model usage plus the evaluation work required to make it trustworthy.
So a quote for a mostly-rules workflow that is priced like a custom AI build is worth a question. And a build where AI has been added because it sells rather than because the job needs it is worth a different question.
For what it is worth, my own published prices do not separate the two, because almost every real build is a mix. What the price reflects is how many systems it touches, how much of your process needs formalising, and how much testing the judgement steps need.
The short version
Do not buy AI for a job a rule would do. Do not force a rule onto a job that needs judgement. Most useful automation is mostly rules with a little judgement in the middle, and anyone building it for you should be able to point at exactly which parts are which.
If you want help working out which of your problems are which, the automation audit is free, and pointing out that something needs no AI at all is a normal outcome of it.