A dinner plan is a business workflow. Here is how it moved into an AI channel.
Stewbert takes a household from "what's for dinner" to groceries on the counter and a card to cook from. Underneath, it is a small, complete workflow migration: a job people used to do by hand, with rules, arithmetic, an outside party and money at the end, now run through a conversation with an AI. Every design choice in it is one you would make moving an invoicing, onboarding, claims or procurement workflow. This page draws the lines.
Seven lines that hold in both places
Stewbert talks, reads the photographs, proposes dishes and writes the recipes in its own words. A separate calculator applies the house rules, scales the recipes, subtracts the pantry, sizes the list to packages and checks the delivery. Stewbert never states a count the calculator did not produce. The same inputs give the same answer every time.
Split the work in two. The model handles the conversation, the reading and the drafting; the rules stay in a deterministic engine that is versioned and tested: pricing, policy, eligibility, totals, compliance checks. The model calls it; it never imitates it. That line is what makes the channel auditable, and it is the first decision of any migration.
The cook decides who is eating, what is off the table, which dish goes on which night, whether a substitute is acceptable, and whether to check out. Stewbert proposes; the rules score; the cook picks. Two things are locked on the way, the menu and then the list, and nothing downstream starts until each is settled on something the cook ruled on.
Keep the judgment with the person, and gate the artifacts. Approvals, exceptions and the customer's own choices stay human. Name the two or three artifacts the workflow cannot proceed without (a quote accepted, an order confirmed, a claim documented) and make them gates the AI cannot talk its way past. Explanations can be skipped; artifacts cannot.
Stewbert hands the list to the store, either a shopping page or a filled cart, and stops. It never sees a price, never states a total, and never places the order. The cook checks out on the store's own site, where the prices are.
Hand off, do not take over. The AI prepares the transaction; the system of record executes it; a person or a policy approves it. Money, commitments and anything with legal weight live in the systems built for them. Where an outside API returns something the AI must not repeat (prices, here), strip it at the boundary before it is stored or shown.
A photograph of the shelves is a read the cook corrects, never an inventory. A photograph of the delivery is the record, and nothing becomes pantry until it is in the photo or the cook says so. A check that could not run is reported as not checked, never as a pass.
Evidence over assumption. A document the AI read is a draft to be confirmed, with its confidence shown. A check that did not run is a failure to surface, not a green light. The migration succeeds or fails on whether the channel is honest about what it has not verified.
Everything captured lives in the plan: the household, the rules, the pantry, the menu, the list, the log of what was cooked. Every run is a snapshot; every change is a diff the cook can read. A passport holds the whole plan for a cook who wants to take it elsewhere.
The record is the account. State lives in plain, inspectable files the business owns, not in the model's memory. Every run is reproducible, every change is diffable, and the customer can leave with their data. If the AI vendor changed tomorrow, the workflow would not.
The first version asked twenty questions before it did anything, repeated the answers back, and signed every reply with its internal state. The second asks on one screen with a default for everything, keeps the progress bar out of the chat, shows dishes as boxes to tick, and says less. The model did not change. The shaping did.
Most of the migration is shaping, not model capability. Collect what is known on a form, not in a dialogue. Show state in the interface, not in the prose. Give people boxes where the choice is a choice. Ban the workflow's own vocabulary from what the customer reads. Measure turns, seconds and words per reply, and cut all three.
A turn can take a minute while the rules run, so the page shows what is happening and never loses the thread. Photos add up across several taps. The store's edge refused the service's shared address, so the service got a fixed one. None of that is the model's job.
The plumbing is the same plumbing. Background work with visible progress, retries that do not double-submit, addresses partners will allow, secrets kept out of the code, logs that name the step that failed. An AI channel inherits every operational duty of the channel it replaces, and adds the duty of saying plainly when it could not finish.
Moving a workflow of your own
- Pick one workflow with an end. Dinner ends with groceries and a card. Yours ends with a paid invoice, an approved claim, an onboarded account. If it has no end, it is not a workflow yet.
- Write the rules down and make them a calculator. The house rules became a deterministic engine with its own tests before a single conversation was designed. Anything the AI must never guess goes here.
- Name the gates. The artifacts the work cannot proceed without, and who rules on each. The AI gets people to the gate; it does not open it.
- Decide the hand-offs. Which systems execute, where the AI stops, what it must never repeat from what those systems return.
- Design the capture. One screen with defaults for what is known; photographs and documents as reads to be corrected; the record in files the business owns.
- Shape the voice, then measure it. Short replies, one ask at a time, state in the interface, the workflow's vocabulary kept from the customer. Turns, seconds, words.
Why it is built this way
Because a household will forgive a wrong dinner and a business will not forgive a wrong invoice. The rules that keep Stewbert honest (no number the calculator did not produce, no order it did not hand off, no check it could not run counted as passed) are the rules that let a business put a workflow in an AI channel and still sign its name to the result.
Stewbert is a working example, built by Caterbum LLC. If you are looking at a workflow of your own, write to readysetdata@caterbum.com.