How a Growth-Stage Technology Company Built Forecasting That Answers Every What-If Before the Plan Is Signed

Open the walkthrough

Six-Step AI Workflow with Human Oversight

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

Steps 4, 5 and 6 need a person. Steps 4 and 5 are waiting for approval right now. Click either 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 Technology Company Built Forecasting That Answers Every What-If Before the Plan Is Signed

Before

✗A person used to paste four exports into one spreadsheet every Monday morning.

✗A person used to build the forecast in a spreadsheet, most of a Monday, every week.

✗An analyst used to rebuild the spreadsheet by hand for every what-if, two days each.

→

After

✓four systems pulled onto one clean table, broken feeds held

✓next week's numbers, with how sure, or withheld with why

✓any what-if, answered beside the baseline it moved

✓the plan drafted, evidenced, waiting on your signature

✓the live week against the forecast, a page when it leaves the band

Executive Summary

Every number the business plans next week on, from the moment it lands in four systems to the moment the plan is signed. The agents pull and clean the numbers every night, build the forecast with how sure it is printed beside it, answer any what-if, and watch the live week against the number hour by hour. A person signs the plan, answers the pages, and accepts every change to how the forecast is built. Six stations, run as a loop: what Friday's scoring learns about this week's misses is what Monday's forecast is built with.

What Was Broken

A person used to paste four exports into one spreadsheet every Monday morning

A person used to build the forecast in a spreadsheet, most of a Monday, every week

The real cost

An analyst used to rebuild the spreadsheet by hand for every what-if, two days each.

What We Built

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

Data Pull

Pull the night's records off the CRM, the billing system, the support desk and the warehouse system

Gap Check

Check every field the models read against the range that field has always held

Feature Builder

Compute the rolling sums, seasonal shapes and lags each forecast line reads

Model Runner

Score five models per line on the last eight weeks none of them ever saw

Scenario Runner

Run the standing scenario set against every new forecast: hiring, price moves, stock levels

Sensitivity Check

Test each forecast input by moving it alone and reading what the plan does

Plan Drafter

Draft next week's staffing and reorder lines from the forecast and the scenario sheet

Backing Sheet

Show the forecast, its range and last week's actuals above every plan line

Live Watch

Watch orders, tickets and stock movements against the forecast band, hour by hour

First Look

Draft the most likely causes of the move, from the same data the page carries

Score Keeper

Score every forecast line against what the week actually did, range and midpoint both

Change Proposal

Propose the model or band change that would have cut this week's misses

How It Runs

Source Intake

Nobody is asked anything here

Agents pull the night's records off the CRM, the billing system, the support desk and the warehouse system, match the same customer across all four, and hold any feed that does not look like itself. Nothing waits on a person here. A held feed is never forecast around, and never quietly dropped.

Forecast Building

Nobody is asked anything here

Agents turn the clean table into the signals the models read, score five models on weeks they never saw, and publish each forecast line with its range, or withhold it with its reason. Nothing waits on a person here. A line too wide to plan on is withheld, not rounded into confidence.

Scenario Testing

Nobody is asked anything here

Agents run the standing what-ifs against every new forecast, run any what-if a person types in, show each answer beside the baseline it moved from, and refuse the ones the data has never seen. Nothing waits on a person here. An answer names what it rests on, and a thin answer says it is thin.

Resource Planning

A person answers here

Agents draft next week's staffing and stock plan from the forecast and the scenario sheet, put the evidence above every line, and leave blank any line the forecast cannot support. This is where a person works. Everything upstream exists so that one plan, already evidenced, reaches this screen instead of four spreadsheets.

Anomaly Watch

A person answers here

Agents watch orders, tickets and stock movements against the forecast band hour by hour, absorb the feed hiccups without paging anyone, and page a person the moment a real stream leaves its band. A person answers every page. The next move after a real anomaly is a phone call or a decision, and neither belongs to an agent.

Forecast Scoring

A person answers here

Agents score every forecast line against what the week actually did, replay each proposed change against the past twelve weeks, and hold every change until a person accepts it. A person decides every change. The forecast is what the business plans against, and what builds it never moves quietly.

Where a Person Decides

•Step 4, Resource Planning. Sign next week's plan, or change the lines you read differently. Give the freight number the agents would not invent. the plan moves money and people.

•Step 5, Anomaly Watch. Say whether the silent account is a real problem or a pause you already know about. Decide the next move, the page holds until you do. a page never clears itself.

•Step 6, Forecast Scoring. Accept or reject the two proposed changes. Say whether the new product's line waits for more weeks or comes off the sheet. the forecast changes only by your hand.

Operating Model

This changes how work flows through the team.

Role

Responsibility

Resource Planning owner

Sign next week's plan, or change the lines you read differently. Give the freight number the agents would not invent. the plan moves money and people.

Anomaly Watch owner

Say whether the silent account is a real problem or a pause you already know about. Decide the next move, the page holds until you do. a page never clears itself.

Forecast Scoring owner

Accept or reject the two proposed changes. Say whether the new product's line waits for more weeks or comes off the sheet. the forecast changes only by your hand.

Source Intake, when it goes wrong

Look here only if a number looks wrong further down the line.

Forecast Building, when it goes wrong

Look here only if a published number surprises you.

What Transfers, What Must Be True

What transfers

•A person is in the loop wherever the plan moves money or people, a page needs a decision, or the forecast itself is about to change. Three stations refuse to proceed on their own: station 4 Resource Planning, station 5 Anomaly Watch and station 6 Forecast Scoring. Everywhere else the agents pull, build, answer and watch, and reach you only when the call is yours to make.

•Resource Planning stops for a person, and the plan moves money and people.

•Anomaly Watch stops for a person, and a page never clears itself.

•Forecast Scoring stops for a person, and the forecast changes only by your hand.

•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 Resource Planning and has time for it.

•Somebody owns Anomaly Watch and has time for it.

•Somebody owns Forecast Scoring and has time for it.

Failure Modes

What breaks this pattern:

✗ A broken export orders stock

The feed comes in short and the system forecasts around the gap. Next week's stock order is built on rows that never arrived.

✗ A midpoint reads as a promise

The forecast goes out as one number with no range. The business commits money against that number as if it were certain.

✗ The system invents a cause

Something moves and nothing in the data explains it. The system writes a plausible story anyway, and someone acts on that story.

✗ The plan spends unsigned money

A plan line moves budget and people the moment it is drafted. Nobody signed it, so nobody can say who approved the spend.

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.

Nightly feeds pulled

Counted at Source Intake

970

Forecasts built on a broken feed

Counted at Source Intake

Feeds held for a person to look at

Counted at Source Intake

Forecast lines published

Counted at Forecast Building

1,178

Published lines a person had to rebuild

Counted at Forecast Building

Lines withheld as too wide to plan on

Counted at Forecast Building

Scenarios answered

Counted at Scenario Testing

212

Answers given beyond what the data holds

Counted at Scenario Testing

Scenarios refused with the reason attached

Counted at Scenario Testing

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.

1Source Intake

subagentsource-intake

writesdata/clean/<date>/<system>.parquet

raiseshold-broken-feed

may touchdata/raw/**, data/clean/**, the four source systems read-only

2Forecast Building

subagentforecast-build

writesforecasts/<week>/<line>.json

raiseswithhold-wide-line

may touchforecasts/**, data/clean/** read-only

3Scenario Testing

subagentscenario-run

writesscenarios/<week>/<scenario-id>.json

raisesscenario-outside-data

may touchscenarios/**, forecasts/** read-only

4Resource Planning

GATE

subagentplan-draft

writesplans/<week>/draft.md

raisesawait-plan-signoff

may touchplans/**, everything else read-only

5Anomaly Watch

GATE

subagentanomaly-watch

writesalerts/<stream>/<alert-id>.json

raisespage-stream-owner

may touchalerts/**, the live feeds read-only

6Forecast Scoring

GATE

subagentscore-and-learn

writesmodels/proposed/<change-id>.json, reports/forecast-<week>.json

raisesaccept-model-change

may touchmodels/proposed/**, reports/**, the history read-only

Stack

Every system this agent reads or writes.

System

Read at

Stations

the CRM

Source Intake

1 of 6

the billing system

Source Intake

1 of 6

the cleaned nightly data

Forecast Building

1 of 6

the draft plan

Resource Planning

1 of 6

the forecast bands

Anomaly Watch

1 of 6

the forecast record

Forecast Scoring

1 of 6

the live order

Anomaly Watch

1 of 6

the model store

Forecast Building

2 of 6

the paging channel

Anomaly Watch

1 of 6

the published forecast

Scenario Testing

1 of 6

the published forecast and its ranges

Resource Planning

1 of 6

the purchasing system

Resource Planning

1 of 6

the reporting dashboard

Forecast Scoring

1 of 6

the scenario library

Scenario Testing

1 of 6

the staffing rota

Resource Planning

1 of 6

the support desk

Source Intake

1 of 6

the time-series database

Forecast Building

2 of 6

the warehouse system

Source Intake

1 of 6

the week's actuals

Forecast Scoring

1 of 6

ticket and stock feeds

Anomaly Watch

1 of 6

Next Step

Want to see if this pattern fits your predictive analytics?

Start a Conversation See More Work

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