Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 4 and 6 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 Construction Company Built Office Paperwork That Arrives Ready to Sign
Every request that reaches the office of a growing construction company, from the moment somebody asks to the moment it is entered, filed and answered. A change order, a new vendor, a rental extension, an invoice, a payroll change: the agents log it, pull the facts it needs out of the four systems they are scattered across, write the paperwork, and hand a person only what carries their signature or their judgment. Six stations, run as a loop: the rules Friday's close proposes are the rules Monday's intake sorts by and Monday's packets are checked against.
What Was Broken
What We Built
Six stations and 13 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 email, the request form and the site chat, match each one to the workflow its kind has always used, merge the asks that arrived twice, and hold anything too vague to route. Nothing waits on a person here. A held request is answered with what is missing, never dropped.
Record Gathering
Nobody is asked anything hereAgents search the accounting system, the project folders, the HR files and the estimating spreadsheets for what each request names, pull the latest version of every record, and quote where each fact came from. Nothing waits on a person here. A fact two systems state differently is carried forward as two facts, with neither chosen.
Packet Assembly
Nobody is asked anything hereAgents draft each form or packet from the gathered records, fill every field with the record behind it quoted beside it, and leave blank any field no record supports. Nothing waits on a person here. A blank a signer can see beats a guess nobody can.
Sign-Off
A person answers hereAgents pass what the sign-off policy already approves, queue what remains by the job's own deadlines, and put the evidence above each packet. This is where a person works. Everything else exists so that four packets, not thirty, reach this screen.
Entry and Filing
Nobody is asked anything hereAgents enter each approved item into the accounting system and the job records, file the signed packet on its job, read every entry back against the signature page, and tell the requester what was done. Nothing waits on a person here. Nothing unsigned is entered, and an entry that does not match its packet is redone before it is filed.
Weekly Close
A person answers hereAgents count the week's requests to where each ended, write the Friday report a person used to spend a morning building, flag the figures that do not fit the job's own run, and propose the rules that would have stopped this week's mistakes. A person decides the rules. A checklist and a sign-off policy are the company's rules, and an agent never changes one 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:
Two systems state the same fact differently and the agent picks one. The signer approves a guess and never learns a second version existed.
Someone says yes in a hallway and the agent records it as approval. The company's money moves and no record says who committed it.
A figure gets typed into the books that differs from the signed packet. The books now say something nobody signed, and every report downstream repeats it.
A strange number gets folded into the weekly average so the close looks clean. The one figure that could have warned you is gone before anyone reads the tally.
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.
log-requestsoffice/requests/<request-id>.jsonhold-unroutable-requestoffice/requests/**, office/held/**, the inbox and the form read-onlygather-recordsoffice/gathered/<request-id>.jsonrecords-disagreeoffice/gathered/**, the source systems read-onlyassemble-packetoffice/packets/<request-id>.mdcannot-fill-fieldoffice/packets/**, office/gathered/** read-onlyqueue-for-signatureoffice/queue/<request-id>.jsonawait-your-signatureoffice/queue/**, office/packets/** read-onlyenter-and-fileoffice/entered/<request-id>.jsonentry-differs-from-signatureoffice/entered/**, the accounting system and the job records write, office/queue/** read-onlyclose-the-weekoffice/reports/week-<date>.json, office/rules/proposed/<rule-id>.jsonaccept-rule-changeoffice/reports/**, office/rules/proposed/**, the history read-onlyStack
Every system this agent reads or writes.