Hand Next Week's Nine BEOs to a Work Agent and Get Back the Prep List, the Purchase Orders and the Staffing Plan

How many hours does Thursday cost you?

Not Saturday. Saturday is the part that looks like work — the trucks, the flip, the plated 220. Thursday is the part nobody counts. Thursday afternoon your executive chef sits down with next week's nine confirmed banquet event orders and turns them into a consolidated prep list by station, a purchase order for the broadline vendor, another for produce, a seafood order due before Tuesday's cutoff, a Restaurant Depot run list, a staffing plan by event, a rental order due Wednesday, and a truck load schedule. Four to six hours, every week, from the two most expensive people in the building.

Then on Monday two of the guarantees change, and half of it gets redone by hand.

A work agent is software you give an outcome to rather than a series of clicks — it reads your files and systems, works on its own for hours, and hands back a finished document that you review. That is a different animal from a chat box, and 2026 is the year it stopped being a demo.

The packet nobody has a name for

Ask ten caterers what they call the Thursday packet and you get ten answers: the production sheet, the prod list, the weekly build, the pull sheet. No standard name because there is no standard tool. Caterease and Total Party Planner print a beautiful BEO. Restaurant365 or MarketMan track food cost after the fact. Nowsta or 7shifts hold the shifts. Nothing turns nine BEOs into the week's execution plan, so a human does it in a spreadsheet, from memory and last year's file.

Which is why the same three failures repeat. You over-order proteins, because the chef pads every number and nobody wants to be short on a wedding. You under-staff the Friday corporate lunch, because it looked small on paper and nobody counted the fourth-floor load-in with one freight elevator. And you miss a dietary line buried in a BEO notes field — six vegetarians, one celiac — until the captain radios the kitchen mid-salad course.

None of these are competence problems. They are what happens when one person does a multi-source assembly job under time pressure, after a week that already had two tastings and a site visit in it.

What changed: you hand over the outcome, not the clicks

Anthropic shipped Claude Cowork on 12 January 2026 and OpenAI shipped ChatGPT Work on GPT-5.6 on 9 July 2026. Both do the same basic thing and both were aimed at people who are not technical: you state a goal, it connects to your apps and files, breaks the job into steps, works on its own for hours, and hands back finished work — a completed spreadsheet, a priced quote packet, a deck, a filled set of forms. Cowork expanded to mobile and web in July, so your chef can start a job from the parking lot.

The difference from the 2024 assistant is worth being precise about. Then, you asked a question and got an answer you still had to act on. Now you state the finished thing you want on your desk and it assembles it, across the systems where your information lives, and tells you what it did. That is a work order, not a search.

flowchart TD
  A["9 confirmed BEOs for next week"] --> E["Work agent builds the production packet"]
  B["Current broadline and produce price files"] --> E
  C["Guarantees, dietary notes, venue load-in details"] --> E
  E --> F["Consolidated prep list by station"]
  E --> G["One purchase order per vendor"]
  E --> H["Staffing plan and shifts to post in Nowsta"]
  E --> I["Rental order and truck load schedule"]
  F --> J{"Chef and Director sign off by Thursday 3 p.m."}
  G --> J
  H --> J
  I --> J

One Thursday, run as a goal

The chef opens it at 9 a.m. and writes one instruction: build next week's production packet from the confirmed BEOs, price it against current vendor files, staff it to our matrix, flag anything that conflicts.

By eleven there is a folder. The prep list is sorted by station and by day, not by event, which is how a kitchen actually works — every roasted vegetable across four events on one line, with pans and oven time. The purchase orders are split by vendor with seafood pulled out separately for the Tuesday cutoff. The staffing plan lists servers, bartenders and a captain per event against your own ratio, and flags the Friday corporate lunch for an extra runner because the BEO says fourth floor, one freight elevator, no dock.

And there is a conflicts page. Two Saturday events both want the eight-burner and the combi oven between 1 and 3 p.m. The Thompson wedding BEO lists 220 guests but the guarantee on file is 206. One BEO has "GF" in the notes with no count attached, which the agent will not guess at — it flags it for the chef to call the client.

The chef spends 45 minutes on that folder instead of five hours building it. He argues with the staffing on two events, moves a prep item to Wednesday, and sends it. That is the change: the assembly is done, the judgment is his.

Where the money is: over-ordering, not overtime

Most owners assume the win is hours. The hours are real but they are the smaller half. The bigger half is what happens when ordering is built from actual guarantees and real yields instead of a padded guess at 4 p.m. Thursday.

These are illustrative figures — put in your own purchase number and your own waste percentage, and if you do not know your waste percentage, that is the first thing this exercise teaches you. Against that, a handful of business seats on Claude Cowork or ChatGPT Work runs in the low hundreds of dollars a month, plus real setup time getting your price files and BEO data into a state the agent can read.

The honest caveat sits in one line of that table: the food saving only appears if guarantees are recorded in your event software the day they come in. If final counts live in text messages on the sales manager's phone, the agent builds the packet off stale numbers faster than a human would, which is not progress.

What you have to change about how work gets assigned

This is the part that trips owners up, and it has nothing to do with software. You have spent years assigning tasks — "pull the BEOs, price the seafood, post the shifts." Now you assign an outcome and a definition of done, then review the result. Three habits change.

Write the standard down once. The agent needs your staffing ratios, par levels, yield assumptions, load-in rules and vendor cutoffs stated plainly, because today those live in your chef's head. Getting them onto paper is a one-week job you have meant to do for six years, and it is the prerequisite.

Name an owner for every document. The prep list belongs to the chef. The purchase orders belong to whoever signs invoices. The staffing plan belongs to the events manager. Unreviewed work is not finished work, and "the agent made it" is not a person.

And set the deadline earlier than you need it. If the packet is due Thursday at 3, start it at 9. It works on its own for hours; that is the design. Owners who treat it like a search box and expect an instant answer get the worst of both worlds.

What it must not be allowed to do

It should not send a purchase order. Draft it, price it, total it — a human presses send, because a fat-fingered case count on a seafood order is a four-figure mistake with a truck attached.

It should not decide dietary substitutions. It can count "6 vegetarian, 1 gluten-free, 1 shellfish allergy" off the BEO notes and put that at the top of the prep list in bold. What goes on the plate is the chef's call, and if a note is ambiguous the correct behavior is to flag it, not resolve it.

It should not post shifts on its own. A staffing plan that hits your team's phones without a human look is how a captain ends up double-booked across two venues on the same Saturday night.

And it will not sell anything. The tasting, the walk-through, the moment the mother of the bride decides she trusts you — none of that is on this list.

Frequently asked questions

Can it read our BEOs if they only exist as PDFs?

Yes, and that is how most operators start — point it at a folder of next week's BEO PDFs and the vendor price sheets. Slower and slightly more error-prone than a direct connection to your event software, but it works from day one and requires nothing from your software vendor.

Which one should we use, Claude Cowork or ChatGPT Work?

Run the same Thursday packet through both for two weeks and keep the one your chef stops complaining about. They overlap heavily; the deciding factor is usually which connects more cleanly to where your files already live, not which is smarter.

What happens in October when we have four Saturdays in a row?

That is when it earns its keep and when it is most dangerous, because nobody has time to review carefully. Build the habit in February when the calendar is thin, so by peak season your chef knows which two pages he actually needs to read.

Our sales manager already does this in a spreadsheet she built. Is she out of a job?

No, and if you frame it that way she will make sure it fails. She becomes the person who owns the standard and reviews the output — a better use of someone who knows why the Henderson account always wants linens set the night before. Her spreadsheet is a head start; it already encodes rules the agent needs.

One related note: the packet is only as good as the information coming in, and much of that arrives by phone — the guarantee that changes Wednesday, the planner adding a bar hour, the client asking about vendor meals. CallSphere builds AI voice and chat agents that answer those calls and your website chat, capture the details in usable form, and book tours and tastings around the clock — so the numbers your Thursday packet is built from are written down before Thursday.