When a restaurant should not use AI.
Most friction does not resolve to automate.
Three cases settle it before any scoring starts. When failure cannot be taken back. When the task never repeats the same way twice. When the friction is the hospitality.
Everything else routes through six verdicts, and only one of them is automate. We publish the whole thing, including the anchors, because a criterion you cannot inspect is a sales position rather than a method. Use it without us if it is useful.
A model can tell you a sauceis the wrong colour tonight.
It cannot decide whether this plate is right for this room, tonight, with these guests. Building toward a kitchen where it does is building toward a venue where nobody tasted the food. That is a human preserve, and it is the category most rubrics do not have.
The six verdicts
Every friction point resolves to exactly one. The rubric exists to make the choice explicit and defensible, not to maximise the amount of automation.
Eliminate
The work should not exist. A report nobody reads, a double entry, an approval that has never once been withheld. The cheapest possible outcome and the most frequently skipped, because eliminating work produces no invoice for anybody.
Redesign
The work is real and the process around it is wrong. Fix the sequence, the ownership or the physical layout first. Automating a broken flow makes it fail faster and at greater volume.
Standardize
The work varies and it should not. Some variation is genuine and some is drift, separating the two is a judgment call, and getting it backwards is how a venue accidentally sands the character off its own service.
Automate
Deterministic, repeating, tolerant of failure, with clean data. Conventional software usually beats a model here and costs less to run.
Augment
A person stays accountable and a model widens what they can see or how fast they see it. They can still overrule it, and the process still functions if the model is unavailable.
Human preserve
Removing the friction would remove the hospitality. Not a concession, and not a stage on the way to automating it later. Some of what looks like inefficiency is the product being delivered.
Three gates that beat any score
A high aggregate should never outvote a fatal single dimension, which is the standard failure of every weighted model. So these run first and they override.
Irreversible failure
If failure cannot be taken back, automate is unavailable and the best verdict is augment, with a named person accountable. Allergen handling, safety, and anything a guest experiences once and cannot be given back.
No recurrence
If the task never repeats in the same form, conventional automation has nothing to standardize. The useful question becomes a different one: can this be modeled and pressure-tested before it runs the only time it will ever run?
The friction is the hospitality
If the interaction is the product, the default verdict is human preserve and the burden of proof sits with anyone arguing otherwise.
A worked example, and it is the one people get wrong
Reconciling dietary and allergen counts across four revisions of an event order looks like the most automatable task in the building. Repetitive, numerical, rule-bound.
It is not, because the failure is not reversible. Gate one applies, automate comes off the table, and the verdict is augment with a named person accountable.
The more useful finding is upstream. The real problem was never the counting. It was four revisions in circulation with no single current version, which is a redesign, costs nothing, and removes most of the risk before any technology question is asked.