The ASI Wine Guide is a working demonstration of a floor system for a steakhouse: the
list, the menu, and the allergen sheet on every phone, an assistant that answers from
tonight's list, menu changes by camera, training with graded answers, and shift notes
that become the morning briefing.
The founder built it to show what a build can do, on a full list and menu rather than a
toy one. A few working servers asked to use it, and do, which is a better test than any
demo script. It is ASI's own system, not a client engagement, and it is not for sale as
is. It is what we would build for your house, starting from your floor.
Every build starts the same way: a shift beside a closing server, half an hour with
whoever trains new starters, and the list and the allergen sheet as the kitchen really
uses them. These are the six places most houses turn out to be losing time, and the six
the demonstration build was designed around.
The wine list
Maintained by hand. The printed version and the working version drift apart within a week of any change.
The allergen sheet
The column rarely matches what the kitchen actually does. On most floors this is the single riskiest gap.
Training material
Thorough, and almost nobody opens it after their first week. The problem is reach, not content.
Shift notes
Written and discarded daily. Everything learned on a Friday is gone by Saturday.
Supplier sheets
Re-typed by hand into the list, which is where most of the errors enter.
Sales history
The cleanest data in the building, used for nothing but accounting.
What got built
An agent at every moment of the shift.
Everything below works in the demonstration build today. Nothing on this page is planned.
Before service
A pre-shift briefing: what is out tonight and what changed since you last worked, derived from the live list and the change log rather than a memo somebody has to write.
A wine of the day chosen each night, with a sommelier’s take and a pairing checked against the real menu. The pairing rule exists because an early version once suggested a dish the menu did not have.
During service
An assistant a server can talk to or type to, answering from the same list the floor is pouring, with bottle cards and prices.
An allergen protocol: a guest mentions shellfish and every answer is checked against the sheet, conflicts are flagged, and where the sheet is silent it says so.
"86 the Caymus," said in plain language, resolved against real bottles before anything is proposed, and applied only when a manager approves.
After service
An end-of-shift report drafted from the closer’s rough notes and cross-checked against the night’s tables, for the closer to correct and sign.
Ticket capture built with redaction first: where a house wants the tickets kept, payment and identity regions are blacked out before anything is stored and the originals go to a private archive.
A morning digest for the manager: the live 86 list, menu changes, feedback filed, shifts committed, and the day’s AI spend. Every number is a query result, so it can be trusted the way a report is trusted.
Between shifts
Menu and list changes by camera or PDF, proposed as a batch a manager reviews line by line. Every phone updates within a minute of approval.
Training tracks for wine, whisky, and agave, with written answers graded against a rubric, a certification at the end, and a manager view of who has finished what.
A guest simulator: a difficult table, played by a model, graded on what the server actually said, with a grader that checks facts against the live list so it never marks a real wine as invented.
The screens
On a server's phone.
Captures from the demonstration build. No venue or staff are named anywhere, and no
results are quoted, because a demonstration does not have any worth quoting. In a
walkthrough, it runs on your own list.
The learning path. Modules unlock in order; progress carries over.A guest mentions shellfish. The protocol switches on, and every answer after it is checked against the sheet.
What it demonstrates
The habits every build gets.
A person approves every change
The model proposes menu edits, drafts sell lines, and resolves "86 the Caymus" to a bottle. Nothing is written until a manager says yes, because a confidently wrong price is worse than no price.
The agent reads the floor’s list, not a copy
The by-the-glass list the assistant answers from is the one the floor is pouring. One store, one version, every phone within a minute.
Trust is built from what can be checked
The morning digest is queries, not prose. The simulator’s grader verifies against live data. The allergen mode refuses to guess. Where a model could be confidently wrong, code decides instead.
Guest data is handled like guest data
Redaction runs before storage. The archive is private. Staff sign in with a whitelist, and every AI call is metered to the account that made it.
Six things a build can do
What a restaurant build can include.
Each one has a "today" that costs a house something every night, and each one is scoped
for your list, your menu, and your rules. Five work in the demonstration build; the sixth
runs on our own paperwork. A build usually starts with the one that costs you the most.
01 · During service
Answers at the table
Wine, spirits, dishes, and allergens in one place, on the phone already in their pocket.
Today
Four sources, none of them searchable, none of them current, and none of them usable while standing at a table. A server who hesitates defaults to the bottle they remember.
With the agent
They ask out loud or by text and get an answer from tonight’s list in seconds, with the bottle, the price, and why it works. The by-the-glass list the agent reads is the one the floor is pouring, because they are the same list.
What a build of it involves
Your wine, spirits, and food lists in a structured form, with a version that reaches every phone within a minute of a change
A voice and text assistant that answers only from that list and says when a wine has been eighty-sixed
Bottle images, tasting notes, and a sell line drafted by a model and approved by a person
Working in the demo build
02 · During service
Allergen answers that cite their source
The line from your own allergen sheet, or "not recorded, ask the kitchen."
Today
Servers walk to the kitchen, wait, and sometimes guess when the line is slammed. Guessing is the part that carries the real risk.
With the agent
A guest mentions shellfish and the assistant switches into a protocol: every answer is checked against your sheet, the dishes that conflict are flagged, and where the sheet is silent it says so instead of reassuring anyone.
What a build of it involves
Reconciling the allergen sheet with what the kitchen actually does, which is where most of the work is
A protocol mode in the assistant that flags conflicts and refuses to guess
Allergen filters on every menu screen and a dish page that shows the components
Working in the demo build
03 · Between shifts
Menu changes that reach the floor tonight
Point a camera at the supplier sheet. Approve the changes. Every phone updates.
Today
New arrivals and vintage changes are re-typed by hand into the list, which is where most of the errors enter, and the floor hears about them a shift late.
With the agent
The app reads a printed list or a supplier sheet and proposes the changes, item by item. Nothing is written until a manager approves each one, because a confidently wrong price is worse than no price. "86 the Caymus," said out loud, works the same way.
What a build of it involves
Camera and PDF import that proposes additions, price changes, and removals as a reviewable batch
Plain-language commands for the nightly changes, resolved against real items before anything is proposed
A change log the pre-shift briefing reads from, so the floor sees what moved
Working in the demo build
04 · Between shifts
Training that fits before a shift
Short modules on the phone, written answers graded, and practice on a difficult guest.
Today
A thorough set of onboarding documents that almost nobody reopens after week one. Training quality depends on who is running it that day.
With the agent
The material becomes short modules with examples drawn from your own list rather than textbook wines nobody pours. Quizzes grade written answers, not just multiple choice. A simulated guest pushes back, asks for something off-list, or wants a cheaper bottle, and the server is graded on what they actually said.
What a build of it involves
Your existing material rebuilt as tracks with progress that carries over, and a manager view of who has finished what
Graded written answers with a rubric you can read
A guest simulator whose grader checks facts against the live list, so it never marks a real wine as invented
Working in the demo build
05 · After service
Shift notes that outlive the shift
What was 86’d, who complained, what changed, captured at close and briefed tomorrow.
Today
Written on a pad and binned. Institutional memory lasts one shift, and everything learned on a Friday is gone by Saturday.
With the agent
The closer jots a few lines at close and a service report is drafted from them, cross-checked against the night. Where a house wants tickets captured too, payment and identity details are blacked out before anything is stored. The manager gets a morning digest built from the numbers.
What a build of it involves
Optional ticket capture with redaction that runs before storage, and an archive nobody outside the house can read
An end-of-shift report drafted for the closer to correct and sign
A morning digest for managers and a pre-shift briefing for the floor, both read from the same record
Working in the demo build
06 · Between shifts
Agents for the paperwork
Invoices, supplier sheets, and the inbox, read and filed, with an interruption only when something changed.
Today
Supplier invoices, receipts, and bulletins arrive by email and photo and are handled by hand, or not at all, until the end of the month.
With the agent
An agent reads what arrives, files it where it belongs, records what it found, and only raises a hand when something needs a person: a price that moved, a document that does not match the order, a notice with a deadline.
What a build of it involves
An inbox and document pipeline scoped to the paperwork your house actually receives
Rules you can read for what gets filed silently and what gets escalated
A daily or weekly summary of what came in and what changed
Running on our own paperwork
What it is not
It is not a product, and you cannot buy it as it stands. A steakhouse's list, menu, house
rules, and training are its own, and a hotel's are different again. What you can have is a
build that starts the way this one did: on your floor, with your list, scoped to the moments
that cost you the most, and owned by you when it is done.
Artificial Systems Integration
Start where this one started, on your floor.
Tell us what goes wrong most nights. We will tell you what it would take to fix it, what it would cost, and whether it is worth doing at all.
Hi, I'm Amanda — ASI's AI concierge, still a prototype, and a
live example of the kind of system we build. Happy to answer questions or show you around —
in any of about 70 languages, so use whichever you're most comfortable in.
Amanda
ASI's AI concierge
Meet Amanda — our AI concierge, built the way we build for clients.
Ask what we do, how an engagement works, or whether your data problem is one we can solve.
She can speak, listen, show you around the site, and pass your details to the team —
in any of about 70 languages, so use whichever you'd rather speak.
Voice and text are processed by Google's Gemini API to run this conversation
(privacy policy).
Please don't share sensitive personal information. Prefer a form?
Contact page.
