Home Use Cases Work Insights About Contact
Back to Our Work

How a Growth-Stage Food Service Company Built Order Forecasting with Human Approval Before Purchases Are Placed

Operations2
Open the walkthrough

Five-Step AI Workflow with Human Oversight

Each step shows the workspace, what the AI completed, and what still needs human review.

Steps 3, 4 and 5 need a person. Step 3 is waiting for approval right now. Click it to open the full walkthrough.

Estimates are directional and based on stated assumptions. All names, organizations, and identifying details have been anonymized in accordance with our confidentiality agreements.

The Transformation

How a Growth-Stage Food Service Company Built Order Forecasting with Human Approval Before Purchases Are Placed

Before
A person used to pull the overnight numbers off each till and supplier email every morning and type them into a spreadsheet before service.
A person used to call tomorrow's prep counts from last week's sales sheet and a look at the weather.
After
every feed landed, checked, with the broken ones held
every item called for tomorrow, with the uncallable ones flagged
the order drafted, tested both ways, waiting on your signature
the day against the forecast, live, with the flags classified
the day scored against the call, and what to call differently
Executive Summary

Tomorrow's order for every perishable line, from the moment the night's numbers land to the moment the day is scored. The agents load and check every feed, forecast every item at every site, draft and test the order, watch the day run against it, and hand a person only the lines where money leaves on their signature. Five stations, run as a loop: what station 5 learns about today's misses is the forecast station 2 publishes tomorrow, and the rule station 1 checks the feeds against.

What Was Broken

A person used to pull the overnight numbers off each till and supplier email every morning and type them into a spreadsheet before service
The real cost
A person used to call tomorrow's prep counts from last week's sales sheet and a look at the weather.

What We Built

Five stations and 12 subagents. Each subagent carries its own tasks and its own refusal.

A4
Feed Sweep
Collect the night's line items off the tills at all three sites
A3
Feed Check
Check each feed against the shape and row count its nights have always had
A4
Signal Builder
Fold weather, bookings and the events calendar into each item's day record
A4
Number Caller
Work out tomorrow's number for every item at every site
A3
Spread Check
Compare the candidate numbers for each item against each other
A1
Order Drafter
Draft each supplier line from the forecast and the stock on hand
A2
What-If Desk
Show what tomorrow's order does under double footfall and under a washout
A3
Commit Record
Record what was committed, by whom, against which forecast and which what-ifs
A4
Floor Watch
Compare each till against its forecast every fifteen minutes
A3
Anomaly Desk
Classify each flag as a rush, a dead item, a broken feed or a missed delivery
A4
Day Scorer
Check every published number against what the till sold at close
A2
Change Proposal
Propose the change that would have cut each miss that keeps repeating

How It Runs

1

Feed Intake

Nobody is asked anything here

Agents collect the night's numbers off every till, supplier confirmation, weather feed and waste log, match each one to the day it belongs to, and hold any feed that arrives broken or short. Nothing waits on a person here. A held feed is never fed to a forecast and never silently filled in.

2

Demand Forecasting

Nobody is asked anything here

Agents fold the checked feeds into each item's record, work out tomorrow's number for every line, publish the ones the history carries, and flag the items where the history is too thin to publish on. Nothing waits on a person here. An item that cannot be called gets a flag with its reason, not a guess with a straight face.

3

Order Setting

A person answers here

Agents write tomorrow's order for each supplier, cut a line where its shelf life carries less than the forecast asks, show what the plan does on a kind day and an unkind one, and leave blank every line that is yours to set. This is where a person works. Money leaves on this signature, and everything upstream exists so that six lines, not sixty, need their eyes.

4

Day Watching

A person answers here

Agents compare the tills against the forecast through the day, watch the deliveries against their windows, classify every flag by what it is, and hold any corrective order for a person instead of placing it. Nothing waits on a person here while the day tracks. A flag that needs money spent stops and asks.

5

Forecast Scoring

A person answers here

Agents score every published number against what the till sold, count the waste at close against the items over-called, trace each big miss to its cause, and propose the changes that would have cut the misses that repeat. A person accepts or rejects each change. How tomorrow gets called is not something an agent changes quietly.

Where a Person Decides

Step 3, Order Setting. Give the three new menu items their numbers, the forecast will not invent them. Commit the order, both what-ifs are on screen. money leaves on your signature.
Step 4, Day Watching. Say yes or no to the held corrective order before Fourth Coffee's cutoff. Leave the day to run while it tracks; a flag reaches your phone when it does not. no corrective order without a person.
Step 5, Forecast Scoring. Accept or reject the two proposed changes. Say what happens to the Tuesday soup, the call has been wrong four weeks running. the call changes when you say.

Operating Model

This changes how work flows through the team.

Role
Responsibility
Order Setting owner
Give the three new menu items their numbers, the forecast will not invent them. Commit the order, both what-ifs are on screen. money leaves on your signature.
Day Watching owner
Say yes or no to the held corrective order before Fourth Coffee's cutoff. Leave the day to run while it tracks; a flag reaches your phone when it does not. no corrective order without a person.
Forecast Scoring owner
Accept or reject the two proposed changes. Say what happens to the Tuesday soup, the call has been wrong four weeks running. the call changes when you say.
Feed Intake, when it goes wrong
Look here only if a supplier says their confirmation vanished.
Demand Forecasting, when it goes wrong
Look here only when a number surprises you, the workings are attached.

What Transfers, What Must Be True

What transfers
A person is in the loop wherever money leaves or the way tomorrow gets called changes. Three stations refuse to proceed without a person: Order Setting and Forecast Scoring outright, and Day Watching on any corrective order, which it holds until a person says yes. Everywhere else the agents load, check, call and watch, and reach you only when the signature is yours to give.
Order Setting stops for a person, and money leaves on your signature.
Day Watching stops for a person, and no corrective order without a person.
Forecast Scoring stops for a person, and the call changes when you say.
Every subagent says what it will not do. 12 of them do.
What must be true in your environment
The agent can read the systems your records already live in. This one reads 20.
Somebody owns Order Setting and has time for it.
Somebody owns Day Watching and has time for it.
Somebody owns Forecast Scoring and has time for it.

Failure Modes

What breaks this pattern:

✗ Filled Gaps Become Real Numbers

A till stops writing at 21:05 and the system fills the gap in. Every forecast downstream now rests on sales that never happened, and nobody knows which ones.

✗ Thin History Gets a Number

An item with two weeks of history gets a published forecast like any item with two years. The buyer cannot tell a guess from a call, so both get ordered with the same confidence.

✗ Orders Go Out Unsigned

The system sends the order and the business pays for it. When the number is wrong, no person ever agreed to it, and the first review happens after the money is spent.

✗ Silence Reads as No Sales

A site's feed goes quiet and the system reads it as a day of zero sales. Tomorrow's order for that site shrinks toward nothing while the site is still selling.

Directional Outcomes

What the agent counts, and the station that counts it.

These counts are the tallies from one monitored run of the agents. They are not monthly or annual totals.

line items collected off eleven feeds
Counted at Feed Intake
2,412
broken feeds silently filled in
Counted at Feed Intake
0
feeds held for a person to open
Counted at Feed Intake
1
items called across three sites
Counted at Demand Forecasting
142
numbers published on thin history
Counted at Demand Forecasting
0
items flagged or marked for a look
Counted at Demand Forecasting
11
Our measurement policy: We do not publish precise ROI without baseline methodology. Every figure above carries its basis.

What Runs Where

Every step names the subagent that does the work, the record it writes, the thing that raises a question for a person, and what it is allowed to touch. This is drawn from the source, not from a diagram somebody kept in sync by hand.

1Feed Intake
subagentfeed-intake
writesdata/day/<date>/<feed>.json
raiseshold-broken-feed
may touchdata/day/**, data/held/**, the source systems read-only
2Demand Forecasting
subagentdemand-forecast
writesforecast/<date>/<item>.json
raisestoo-little-history
may touchforecast/**, data/day/** read-only
3Order Setting
GATE
subagentorder-draft
writesorders/draft/<date>-<site>.md
raisesawait-your-commit
may touchorders/draft/**, everything else read-only
4Day Watching
GATE
subagentfloor-watch
writeswatch/<date>/<flag-id>.json
raiseshold-corrective-order
may touchwatch/**, orders/draft/**, the live tills read-only
5Forecast Scoring
GATE
subagentscore-and-learn
writeschanges/proposed/<change-id>.json, reports/day-<date>.json
raisesaccept-forecast-change
may touchchanges/proposed/**, reports/**, the history read-only

Stack

Every system this agent reads or writes.

System
Read at
Stations
the bookings book
Feed Intake
2 of 5
the change log
Forecast Scoring
1 of 5
the committed order
Forecast Scoring
1 of 5
the day records
Demand Forecasting
1 of 5
the day's till record
Forecast Scoring
1 of 5
the delivery calendar
Order Setting
1 of 5
the delivery windows
Day Watching
1 of 5
the festival and events calendar
Demand Forecasting
1 of 5
the forecast library
Demand Forecasting
2 of 5
the live tills
Day Watching
1 of 5
the par and shelf-life sheet
Order Setting
1 of 5
the site leads' phones
Day Watching
1 of 5
the supplier order forms
Order Setting
1 of 5
the supplier portals
Feed Intake
1 of 5
the three site tills
Feed Intake
1 of 5
the waste count
Forecast Scoring
1 of 5
the waste log
Feed Intake
1 of 5
the waste log as it fills
Day Watching
1 of 5
the weather feed
Feed Intake
1 of 5
the what-if runner
Order Setting
1 of 5
Next Step

Want to see if this pattern fits your predictive analytics?

No build commitment·Real samples, not a demo·Estimate in writing