Where AI actually belongs in an operation
Every tool on your stack shipped an AI feature in the last eighteen months, and most of them are the deterministic automation that was already there, with a new price on it. The useful question is not whether to turn AI on. It is which single step in a process should be a judgment, what belongs on either side of it, and what it costs when nobody is watching the meter. These three demos are how I answer that.
Rules on the outside. Judgment in the middle. A person where it counts.
Here is a real intake process, seven steps from a request landing to work starting. Six of those steps are pure logic and should never cost a credit. One is a judgment, and that is the one worth paying an AI to make. One should never leave a human at all. Click any step to see which, and why.
The distinction most vendors blur
| What it is | Who decides the next step | What it costs |
|---|---|---|
| A rule status changes, owner gets notified |
You did, when you drew it | free on every plan |
| A rule with a judgment inside it is this urgent, who owns it |
You drew the path, the model fills one gap | metered per run |
| An agent given a goal, picks its own tools |
The model does, including what to try next | the most expensive line on any price list |
Notice what the third row tells you. Every vendor has priced autonomy at exactly the level of their own confidence in it. When the same company charges nothing for a rule and three hundred credits for an agent run, that is not a pricing decision, it is an admission.
The classifier below is real logic running in your browser, not a recording. Same input gives the same output, and you can feed it something it has never seen. What it is not is a deployed agent: in production this step calls a model, reads the live board, and writes the result back. Run it twice, break it on purpose, paste in something absurd. A demo that cannot survive that is not worth showing you.
An inbound request
The decision trace
Nothing yet. Every signal it checks will be listed, including the ones that did not fire, because the misses are how you audit a decision.
The pricing model matters more than the feature list.
Three incompatible ways to charge for the same thing are in play across the tools you already own, and the difference decides whether an AI feature is a budget line or a fuse. This is the conversation I have before anything gets switched on, and it is the one nobody has with you at the point of sale.
The failure that should be on every implementation checklist
A team wired inbound email into their work system, added an AI step to prioritise each new item, and left the platform's own notification emails switched on. Those notifications arrived as email. Email created items. Items sent notifications. The AI ran on every cycle. They burned through four million credits in a month and every AI rule they owned was disabled until renewal.
Nothing about that is an AI problem. It is a loop, and loops are the oldest bug in operations. It only became expensive because somebody put a metered step inside one. The rule I work to: never place an AI step in a workflow whose output can re-trigger its own input, and check that before anything is enabled, not after the invoice.