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 an Enterprise HR Services Company Built a Paperwork Agent That Follows Up on Missing Documents
Every document a client sends, from the tray it lands in to the record it becomes in the client's own system. The agents sort it, read it, check it, chase what is missing, post what is complete, and hand a person only the records where a value needs their judgment or a rule change needs their name. Six stations, run as a loop: the corrections a person makes at station 4 and the failures station 5 holds become the rules station 6 proposes, and an accepted rule changes how station 1 sorts and station 2 reads tomorrow.
What Was Broken
What We Built
Six stations and 14 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Document Intake
Nobody is asked anything hereAgents collect every document off every route, name its type and its client from the pages themselves, merge duplicates, and hold anything that cannot be read or placed. Nothing waits on a person here. A held document is never discarded and never keyed by hand.
Data Capture
Nobody is asked anything hereAgents read every page, typed print, tables, checkboxes and handwriting alike, put each value into the field its document type carries, and mark every field they are not sure of. Nothing waits on a person here. A field the page cannot support is left empty and marked, never filled with a guess.
Record Validation
Nobody is asked anything hereAgents check every captured record against what it must carry: required fields, dates that agree with each other, an employee the client's roster knows. Incomplete records are chased with the missing item named. Nothing waits on a person here. A standard chase goes out on its own; a record that needs judgment routes to station 4.
Exception Review
A person answers hereAgents put each doubtful record on one screen, the field beside the page that produced it, and recommend a correction where the page supports one. A person confirms, corrects, or sends it back. This is where a person works. Everything else exists so that 27 records, not 340, reach this screen.
System Posting
Nobody is asked anything hereAgents post every cleared record into the client's own system under that client's field mapping, read each post back, and hold any post that failed or came back changed. Nothing waits on a person here. A failed post is held with the system's own error, never retried into a second copy.
Batch Close
A person answers hereAgents check every document against where it ended, keep the trail per document, and propose the sorting, reading and mapping rules that would have stopped today's repeats. A person accepts or rejects each rule. A person decides here. How a document type is read, and how a client's fields are mapped, 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:
The page shows no value, so the system borrows one from a similar record. A guessed field looks the same as a read one, and it enters a payroll run with nobody checking it.
Two employees share a name, and the document attaches to the wrong roster row. One person's deduction is applied to another person's pay.
The send returns success, so the record is marked done. The client's system dropped a field on the way in, and nobody reads the record back to notice.
The system adjusts a client's mapping on its own after a few odd pages. Every later document for that client is read by a rule nobody agreed to.
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.
doc-intakedocs/intake/<doc-id>.jsonhold-unplaceabledocs/intake/**, docs/held/**, the drop folders read-onlypage-capturedocs/captured/<doc-id>.jsonfield-unreadabledocs/captured/**, docs/intake/** read-onlyrecord-validatedocs/validated/<doc-id>.json, chase/<doc-id>.jsonjudgment-gapdocs/validated/**, chase/**, the client rosters read-onlyexception-reviewdocs/confirmed/<doc-id>.jsonawait-your-confirmdocs/confirmed/**, everything else read-onlysystem-postposts/<doc-id>.jsonpost-failedposts/**, docs/confirmed/** read-only, the client systems through the mapped fields onlybatch-closerules/pending/<rule-id>.json, reports/docs-<date>.jsonaccept-rule-changerules/pending/**, reports/**, the history read-onlyStack
Every system this agent reads or writes.