The Saltwater Cowboys experiment (two AI employees, one very human outcome)
Saltwater Cowboys is a wildly popular restaurant on Shem Creek in Charleston, South Carolina. Recently I built them two AI employees:
- Amy — the front-line receptionist who answers the phone in production
- John — the AI manager who handles anything “not normal,” and escalates the truly human-required edge cases
Amy now answers roughly 50 calls a day. Before this, the human managers would end up having to personally deal with around 30 of those 50 because phone calls aren’t just “information.” They’re emotion, confusion, exceptions, special requests, complaints, hiring inquiries, vendor coordination, and the constant little surprises that drag a manager away from the floor.
Today, that human attention load is down to about three calls a day.
That’s the part people like to hear because it feels like automation.
But the more important part is why it worked.
The real trick wasn’t training. It was constraining.
Most people assume building an AI employee is like onboarding a new hire:
“Here’s the menu.”
“Here’s how reservations work.”
“Here’s the hours.”
“Here’s how to handle to-go orders.”
“Here’s the daily special.”
That’s not what happened.
I didn’t train Amy and John on how to be restaurant staff.
They already know how.
They already know what a restaurant is.
They already know the shape of phone calls.
They already know what customers tend to ask.
They already know how complaints typically go.
They already know how hiring inquiries sound.
They already know what an “owner request” feels like.
They even know a thousand plausible menu items and a thousand plausible “daily special” formats.
That’s the point.
These models arrive pre-loaded with general competence. In many ways they show up like a hyper-experienced employee who has worked everywhere.
Which creates a new problem:
They also arrive pre-loaded with plausible nonsense.
The bull riding problem (the clearest example I’ve ever seen)
“Saltwater Cowboys” sounds like it could be a place with bull riding.
So the AI will happily infer a whole vibe:
- rodeo nights
- mechanical bull
- themed events
- cowboy trivia
- “yeehaw” energy
Not because it’s stupid.
Because it’s helpful — and its job is to complete patterns.
If you don’t constrain it, it will invent a reality that sounds reasonable.
So you don’t “train” it by feeding it a lesson on restaurants.
You constrain it with truth:
- We do not have bull riding.
- We do not have a mechanical bull.
- Do not imply rodeo events.
- If asked, answer clearly and move on.
That’s not training. That’s governance.
Nuance is mostly negation
When people say, “AI can’t handle nuance,” what they often mean is:
AI can’t reliably guess your specific version of the world.
Because nuance isn’t more knowledge.
Nuance is the shape of the local truth.
And local truth is largely made of:
- exceptions
- constraints
- house rules
- edge cases
- what you don’t do
- what you never promise
- what you must escalate
- what you must not invent
That’s why my prompts for Amy and John are not giant encyclopedias.
They’re mostly negation.
Examples from this project looked like:
- “Don’t make up items that aren’t on the menu.”
- “Don’t claim you can do things we don’t do.”
- “If you don’t know, say you don’t know.”
- “Never promise a manager will call back unless explicitly instructed.”
- “Certain situations must be escalated.”
- “Certain phrasing must be used for clarity.”
Even when you do provide menu details, the real value isn’t that the AI “learned the menu.”
The real value is that the AI is now constrained from confidently hallucinating a menu that sounds right.
Why Amy + John works: an escalation ladder, not a brain transplant
Here’s the structural move that made the whole system feel sane:
- Amy handles the predictable 80–90%.
- John handles the messy 10–20% — complaints, employment, vendor weirdness, “I need the owner,” “I need a receipt,” “I’m a contractor rewiring the mezzanine,” etc.
- John then texts the HMOC (human manager on call) with a clean summary when a real human needs to step in.
That means humans aren’t “replaced.”
They’re protected.
And the restaurant doesn’t lose the human touch where it matters most: the truly weird situations where judgment and authority live.
Training is expensive. Constraints are scalable.
If you approach AI employees like training, you end up chasing an impossible goal:
“I need them to know everything about my business.”
That’s infinite work, and it’s the wrong model.
If you approach it like constraining, you’re doing something much more realistic:
“I need them to know what not to do, what not to promise, how we specifically operate, and when to hand off.”
That’s finite.
That’s documentable.
That’s testable.
That’s maintainable.
And it maps perfectly to how great managers actually run organizations:
- define boundaries
- define escalation
- define tone
- define what “good” looks like
- define what is never allowed
The new craft: writing constraints that sound human
The art isn’t “prompt engineering” in the cringe sense.
The art is writing operational constraints that:
- keep the AI truthful
- keep the AI calm
- keep the guest experience smooth
- keep staff protected
- keep the business from accidental commitments
- keep escalation clean and rare
That’s why I call this constraining, not training.
Training is trying to make the AI smarter.
Constraining is trying to make the AI faithful.
The takeaway
AI workers are arriving with something we’ve never had in labor before:
A new employee who already knows how the world works.
Your job isn’t to teach them what a restaurant is.
Your job is to carve a narrow tunnel through their broad knowledge so they reliably operate inside your reality.
That tunnel is made of constraints.
And once you see it, you can’t unsee it:
The future of “managing AI employees” is less about giving them more information…
…and more about telling them, with precision, what is not true here.