Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 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 Articles Scored for Voice with Editor Signoff Before Publishing
Turns a client's calendar slot into a published, tested, reported piece, and never lets a sentence it could not source go out under a client's name. Written for an agency where the record is the point: every brief carries the search evidence behind it, every draft carries its voice score and its sources, and the only things left for a person are the signoff before publication and the numbers the client hears. Six stations, run as a loop: what the June report learns is what the July briefs are built from.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Brief Building
Nobody is asked anything hereAgents research every calendar slot before anything is written: what people search, who already ranks, and what angle is still open. Nothing waits on a person here. A slot only reaches somebody when the research cannot defend an angle for it.
Article Drafting
Nobody is asked anything hereAgents write the piece from the brief, score every sentence against the client's own voice, and check the page against the term it is written for. A person sees the draft at the next station, with the flags already attached. Nobody proofreads blind here.
Editorial Signoff
A person answers hereAgents lay the draft, its voice score, its flags and its sources on one screen, so the editor decides from one read instead of a hunt. This is where a person works. Every piece is read by an editor before the client's name goes on it, and the editor's notes come back as a revision, never a rewrite.
Channel Publishing
Nobody is asked anything hereAgents slot each approved piece into the calendar, cut the version each channel needs, put it live and check the live page before the next channel fires. A person signs nothing here. The signature already happened at station 3, and a failed live check holds the run on its own.
Headline Testing
Nobody is asked anything hereAgents write three headline and description variants for each live piece, split the traffic between them, and promote a winner only when the numbers can carry the call. A person used to pick the headline by taste. Nobody picks here now: the test settles it, or the test says it cannot.
Client Reporting
A person answers hereAgents trace every lead back through the pages it read, build the month's report per client, and turn what worked into next month's briefs. A person signs the report. That is the one signature this side of the loop that stays human, because the client hears the numbers from a person who is answerable for them.
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:
When the brief gives a section no facts, the drafter fills the gap with a number that sounds right. That number reads like research, and it stays in the article until a reader checks it.
A claim with no source reaches the editor looking like every other sentence. The editor signs off on something the brief cannot back, and the client owns the correction.
The site version fails its check, and the email, the social posts, and the newsletter go out anyway. Every channel now points at a page that is wrong, and you fix five surfaces instead of one.
When a lead's path is too broken to trace, the report credits whatever page it touched last. The client reads that as proof and puts next quarter's budget behind content that did not earn the 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.
brief-researchbriefs/<client>/<slug>.jsonno-defensible-anglebriefs/**, rankings/** read-onlydraft-articledrafts/<slug>.mdvoice-below-floordrafts/**, briefs/** read-only, voice/** read-onlyreview-packetreview/<slug>.jsonawait-editor-signoffreview/**, drafts/** read-onlypublish-runpublished/<slug>/<channel>.jsonlive-check-failedpublished/**, calendar/**, review/** read-onlyvariant-testtests/<slug>.jsontest-inconclusivetests/**, published/** read-onlymonth-reportreports/<client>-<month>.json, briefs/proposed/<client>-<month>.jsonsign-client-reportreports/**, briefs/**, tests/** read-onlyStack
Every system this agent reads or writes.