21 August 2026 · 5 min read
What actually counts as an AI workflow
TL;DR
An AI workflow is a repeatable business process where a machine does the reading and drafting and a named human approves the result. Three conditions: it recurs, it produces a record, someone signs it. Building something with AI help is a different purchase entirely — useful, but not a workflow.
A question I was asked this week: is building a website an AI workflow? It is a better question than it sounds, because the honest answer separates the two things most SMEs are actually shopping for and cannot yet tell apart.
The test
A workflow is work that comes back. Every Tuesday, every invoice run, every inbound enquiry. The machine reads the mess, drafts the output, and stops. A named person approves it. That is the whole shape, and it is why the value is easy to state: this used to take four hours a week and now it takes twenty minutes of review.
If you cannot say how often it runs and who signs it, it is not a workflow yet.
Why a website build fails it
A website is a project with a fixed deliverable and an end date. You do not run it again next month. Building it with AI assistance makes the build faster and the iteration cheaper — genuinely so — but the finished site is not a process that gives hours back, because nobody was spending hours on it in the first place. It is a shop window, not a machine on the shop floor.
This distinction matters commercially. "We build things with AI" is a claim every agency in the country is now making. "We build the process that runs your Monday morning for you, and it has been running ours since May" is a different sentence.
Four that do pass, on this site
The site you are reading is the shop window. The agents behind it are the demonstration — and each one has the same three parts.
Ella, the agent on this site
- Trigger
- Someone lands on the site and starts asking questions.
- What the machine does
- She qualifies the enquiry, captures a contact, checks real availability and books a slot.
- Human gate
- The write-up lands in my inbox before I speak to anyone. I decide who is worth a call.
One Workflow, the weekly letter
- Trigger
- A company signs up with a work email.
- What the machine does
- Public information about that company is read, candidate opportunities are scored on hours saved and difficulty, and a letter is drafted for that specific business.
- Human gate
- Nothing sends until I have read it. A week that produces nothing worth sending sends nothing.
Newsletter triage
- Trigger
- AI newsletters arrive and pile up, as they do for everyone.
- What the machine does
- They are read for one thing only: a workflow an Irish SME could actually run, with the release note as the reason it is possible now.
- Human gate
- The output is a queue of drafts. I publish or bin each one.
Lead routing
- Trigger
- An enquiry finishes, by form or by voice.
- What the machine does
- It is scored on fit against the work Artellis does.
- Human gate
- A strong fit rings a phone. Everything else waits in a list. No lead quietly dies in an inbox.
The distinction worth keeping
There are two purchases here, and they are both real. One is an AI workflow: a recurring process, automated to the point of review, that hands back hours every week. The other is AI-assisted delivery — spec, generate, review, ship — which is how internal tools and sites get built now, at a fraction of what they used to cost. SMEs who want a rota tool, a quoting sheet or a customer portal want the second one. Confusing them is how a project gets sold as a saving that never arrives.
Run it against your own idea
- Does it recur? If it happens once, it is a project. Hours-per-week is the only unit that turns into money, so if you cannot say how often it runs, you cannot say what it is worth.
- What record does it produce? A drafted reply, a posted invoice, a scored lead, a filed job report. If the output is a feeling that things went faster, there is nothing to audit and nothing to improve.
- Who signs it? A named person, not a team. If nobody is accountable for the output, the output does not get used — and under the EU AI Act, that person is also your compliance story.
- What happens when it is wrong? Every workflow worth shipping has a wrong answer somewhere in it. You need to know how it surfaces, who sees it, and what the fallback is. If the answer is nobody notices, do not ship it.
Four honest answers and you have either a workflow worth costing, or a project worth calling a project. Both are fine. Only one of them belongs in a business case about hours.
Not sure which one you have?
A discovery week runs the test across your candidate workflows and comes back with a ranked, costed shortlist — including an honest don't-build on the ones that fail it.