Five-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 4 and 5 need a person. Step 4 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.
How a Growth-Stage Data-Driven Company Built Forecasting Every Operating Number with Human Approval Before Money Moves
Every operating number the company plans on, from the overnight data pull to the decision it pays for. The agents collect and clean the numbers, forecast every line with a range, watch the day against the forecast hour by hour, and hand a person only the calls that move money. Five stations, run as a loop: what Friday's scoring learns about a model or a feed changes what Monday's forecast says and what the intake holds at the door.
What Was Broken
What We Built
Five stations and 11 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Data Intake
Nobody is asked anything hereAgents collect every feed from every system, match the same order across systems that name it differently, merge duplicates, and hold any feed whose totals disagree with the system it came from. Nothing waits on a person here. A held feed never enters a forecast and is never quietly patched over.
Forecast Building
Nobody is asked anything hereAgents shape the cleaned data into signals, run the model set against every line, pick the best model per line on weeks it never saw, and publish each forecast with the range it honestly carries. Nothing waits on a person here. A line no model can beat guessing on gets no number, and the withholding is flagged, not hidden.
Anomaly Watch
Nobody is asked anything hereAgents compare each hour's actuals against the band the forecast drew, wait out a single odd hour, and turn a swing that holds into one alert carrying its own numbers. Nothing waits on a person here. A real anomaly reaches the person who owns that line as an alert, days before it would have surfaced in a weekly report.
Allocation Decisions
A person answers hereAgents run each pending stock and staffing call through three futures, draft a recommendation where the futures agree, and put the forecast line under every number. A person picks. This is where a person works. Everything before it exists so that three calls, not forty lines of data, reach this screen.
Forecast Scoring
A person answers hereAgents score every published forecast against what actually happened, trace each big miss to its cause, and propose model changes with their downsides written beside their gains. A person accepts or rejects each one. A person decides here. How the company forecasts is not something an agent changes quietly.
Where a Person Decides
Operating Model
This changes how work flows through the team.
What Transfers, What Must Be True
Failure Modes
What breaks this pattern:
A day's feed arrives with a total that disagrees with the system it came from, and it goes in anyway. Every forecast built that day inherits the error, and nobody can say later which numbers it touched.
A model that cannot beat the naive guess publishes its numbers anyway. The board plans stock and hours against a figure no better than guessing, and treats it as a prediction.
Every odd hour sends an alert, with no pattern behind it. People learn that alerts mean nothing, so the real event gets the same silence.
An allocation executes before any person picks it. Stock and hours move on a model's output, and the first person to see the call is the one who has to unwind it.
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.
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.
feed-intakedata/clean/<feed>/<date>.jsonhold-broken-feeddata/**, the source systems read-onlyforecast-runforecasts/<week>/<series>.jsonno-model-beats-guessingforecasts/**, data/clean/** read-onlylive-watchanomalies/<date>/<event-id>.jsonraise-anomaly-alertanomalies/**, forecasts/** read-only, the live feeds read-onlyallocation-deskdecisions/<week>/<call-id>.jsonawait-your-allocationdecisions/**, scenarios/**, everything else read-onlyscore-and-tunescores/<week>.json, changes/proposed/<change-id>.jsonaccept-model-changescores/**, changes/proposed/**, the history read-onlyStack
Every system this agent reads or writes.