Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 1, 3 and 6 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 Marketing Agency Built Keyword Gap Articles with Human Approval Before Budget Is Committed
Turns a client's keyword gaps into published pieces that sound like the client wrote them, and never lets a claim without a source go out under a client's name. Written for an agency where the record is the point: every brief carries the data that argued for it, every published value traces back to its draft, and the only things left for a person are the calendar, the clearance and the signature on the month. Six stations, run as a loop: what station 6 learns about August is what station 1 briefs for September.
What Was Broken
What We Built
Six stations and 12 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Topic Planning
A person answers hereAgents track every keyword the client competes on, find the gaps competitors have not filled, and draft a brief for each topic before anybody writes a word. A person approves the calendar, because the plan is where the client's budget gets committed.
Draft Writing
Nobody is asked anything hereAgents write each piece to its brief in the client's voice, weave the keywords in as sentences, and draft the title and meta variants the test will later need. Nobody is asked anything here. A draft that stops on a missing source, or drifts off voice, reaches the editor at station 3 with the lines already marked.
Editorial Review
A person answers hereAgents check every claim against its source and assemble each draft into a card the editor can clear from a few lines, with the fix for every flag already drafted. This is the station where a person works. Everything before it exists to make this read short, and nothing publishes without it.
Channel Publishing
Nobody is asked anything hereAgents convert every cleared piece into each channel's own format, publish on schedule, and confirm each one is live and indexed where it should be. Nobody presses publish any more. A piece that breaks in conversion holds itself, and reaches the editor with the broken element named.
Headline Testing
Nobody is asked anything hereAgents run the editor-approved title and meta variants against each other, split the traffic evenly, and lock the winner only when the gap between them is real. Nobody picks a headline by feel any more. The variants were approved at station 3; here the numbers decide between them.
Performance Reporting
A person answers hereAgents trace every lead back to the piece and the keyword that produced it, draft the month's report, and map what ranked and converted into next month's candidate topics. A person signs the report before the client sees it. That signature stays human, because somebody must be answerable for numbers a client pays against.
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 writer fills a source gap with a sentence that sounds right. The client publishes a claim nobody can defend when a reader asks where it came from.
The calendar schedules a topic the site already ranks for. The new page and the old page split the search traffic, and both rank lower than the old page did alone.
The test picks a winner while the gap between variants is still random. Every headline after that gets shaped by a result that was never real.
A lead the system cannot trace gets credited to the nearest page. The client pays for more of the content that never produced a lead.
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.
topic-researchplan/briefs/<slug>.jsonapprove-calendarplan/**, reports/** read-onlydraft-writerdrafts/<slug>.mdclaim-without-sourcedrafts/**, plan/briefs/** read-only, voice/** read-onlyreview-briefreview/<slug>.jsonawait-editor-clearancereview/**, drafts/** read-onlypublish-runpublished/<slug>.jsonbroken-renderpublished/**, review/** read-onlyvariant-testtests/<slug>.jsontest-refusedtests/**, published/** read-onlymonth-reportreports/<period>.jsonsign-reportreports/**, tests/** read-only, plan/** read-onlyStack
Every system this agent reads or writes.