You Probably Don't Need
an AI Workflow.

I was a believer

Not long ago I was on the AI-workflow side of this argument. Why hand-code rules and pipelines when a model can just read the situation and do the task? It felt obvious — automation was the old way, AI workflows were the future, and I said so.

Then I built one. And building it is what changed my mind.

Where it cracked: consistency and cost

An AI workflow puts a model in the loop, making decisions at run time. That means it's probabilistic — it does not hit 100% consistency. On a one-off task with judgment involved, that's fine. On a repeated job, "usually right" is not good enough; the one time in twenty it drifts is the time that costs you.

The fix everyone reaches for is a more advanced model. But that just trades one problem for another: you pay more per run, forever, to buy stability you should not have needed. I was renting a brain for a job that didn't need one — and paying for it on every single execution. That's when I started to doubt my own decision.

Don't pay for the brain

For a repeated, well-defined task, automation is deterministic: same input, same output, every time, at almost zero cost per run. There's no brain to rent. Once I saw the monthly bill next to the consistency I was actually getting, I started switching things back to plain automation.

Then I read a piece making the same case, and it settled it. I wasn't being nostalgic — the economics simply pointed the other way.

Why everyone jumped to AI workflow

Here's the history that gets skipped. Before AI, automation meant programmers. You paid them to build it, and then — the expensive part — to maintain it. As a system grows, it gets complex: more integrations, more edge cases, more things that break when one changes. Eventually the maintenance load outgrows the people holding it. The codebase becomes a burden nobody can fully carry, and paying developers to keep it alive gets harder and harder to justify.

That pain is real, and I've felt it. It is exactly why people ran toward AI workflows — let a model absorb the mess instead of maintaining brittle code. Seen from inside that pain, "just use AI" sounds like relief.

But the bottleneck just disappeared

Now take one step back. The thing that made automation expensive was never the running — it was the building and maintaining of the code. And that is precisely what AI is now good at. AI writes the automation, and AI helps maintain the codebase. The bottleneck that pushed everyone toward putting a model in every loop is gone.

So the cost of automation becomes measurable and manageable again — and its old advantages come back, magnified: deterministic, consistent, and nearly free to run at scale. The shift is subtle but it decides everything:

To be fair, that's not the whole map. Where a task genuinely needs judgment, natural language, or handling the genuinely unpredictable, a model in the loop earns its cost — that's real AI-workflow territory, and it's worth every ringgit there. The mistake is using it everywhere.

So, do you really need an AI workflow?

This is why I'm writing. A lot of owners and beginners are being talked into going all-in on AI workflows — often by the people selling AI workflows. It sounds modern. It is frequently the wrong tool, at the wrong cost, for the job in front of them.

So think properly, and ask two honest questions. Is the task repeated and well-defined? Then automate it — you don't need to pay for a brain. Does it truly need judgment or language or the unpredictable? Then let a model do that part, and cap what it costs.

AI absolutely belongs in your business. It just doesn't belong in every loop. Use it to build the machine — then let the machine run cheap.

If you're mid-decision on this and want to think it through with someone who has made the wrong call and walked it back, my door is open — no pitch, just a candid sharing session between people building things.

Now or Never.

Published by IMA AI — July 2026.