Five-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3 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.
How a Growth-Stage Technology Company Built Request Handling with Human Approval Before Customer Changes Go Live
Every operations request, from the moment it arrives to the moment its change is live and read back in every system it touches. The agents take each request in, pull what billing, the CRM and the product database already say about it, draft the action, apply the ones policy already answers, and hand a person only the changes to what a customer pays or gets. Five stations, run as a loop: the rules a person accepts at station 5 are the rules station 1 holds against and station 3 answers tomorrow.
What Was Broken
What We Built
Five stations and 11 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Request Intake
Nobody is asked anything hereAgents collect every request off the form, the shared inbox and the system queues, match each one to its account, merge the duplicates, and hold anything missing the fields its kind needs. Nothing waits on a person here. A held request is never worked from a guess; the missing piece is named and asked back.
Record Assembly
Nobody is asked anything hereAgents pull every field the request touches from billing, the CRM and the product database, trace any disagreement to the write that caused it, and stop a request whose systems disagree on money. Nothing waits on a person here. A money conflict is never settled by picking a side; it goes forward with both lines named.
Action Drafting
A person answers hereAgents draft the action for every request from its assembled record, pass the ones policy already answers straight to write-back, and hold every change to a price, a plan or an access right for a person. This is where a person works. Everything else exists so that three actions, not a hundred and twenty, reach this screen.
System Write-Back
Nobody is asked anything hereAgents queue each cleared action as one job per system, apply the writes in the order the systems need, read every record back, and stop any job whose target moved since the draft. Nothing waits on a person here. A record that changed under a job is never written over; the job stops and the draft is redone from the new record.
Day Close
A person answers hereAgents check every request against where it ended, count what settled without a person, propose the intake and policy rules that would have stopped today's repeated mistakes, and show what each rule would hold wrongly. A person accepts or rejects each rule. What policy answers on its own, and what gets held at the door, 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:
An incomplete request comes in and the agent fills the blank itself. The wrong account or the wrong plan then moves through every step that follows.
Billing shows one price, the CRM shows another, and the agent picks one. The customer pays the wrong amount, and nobody knows the two systems ever disagreed.
The agent changes what a customer pays without a person seeing the evidence first. The first person to catch the mistake is the customer, on their invoice.
The record changed after the draft was made, and the job writes over it anyway. Someone else's fix is erased, and nothing in the system shows it happened.
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.
request-intakerequests/intake/<req-id>.jsonhold-incomplete-requestrequests/intake/**, requests/held/**, the form and the queues read-onlyrecord-assemblyrequests/assembled/<req-id>.jsonsystems-disagree-on-moneyrequests/assembled/**, billing, the CRM and the product database read-onlydraft-actionrequests/drafts/<req-id>.mdawait-your-clearrequests/drafts/**, everything else read-onlywrite-backrequests/applied/<req-id>.jsonrecord-changed-under-the-jobrequests/applied/**, the three systems through their connectors onlyclose-and-learnrules/proposed/<rule-id>.json, reports/requests-<date>.jsonaccept-rule-changerules/proposed/**, reports/**, the history read-onlyStack
Every system this agent reads or writes.