← All case studies

Case study · Content operations

An AI content engine for a publishing operation

A content creator was publishing a blog and video across social media, with no record of what an audience of tens of thousands kept asking. We built read-only audience tooling, a tri-scoped Researcher, a Drafter, and a privacy-first site, with a person approving every result.

01 / The business problem

Growth created work nobody had planned for

In under three months, the creator went from a standing start to a fast-growing audience across several platforms. Every comment needed attention, every post needed a different format, and the same questions kept returning without a record of what the creator had already answered.

The creator spent the hours that should have gone into making content on re-answering questions and guessing what to publish next.

400+videos
~60Kfollowers
15K+comments
3 mofrom nothing

02 / The manual process before

Repetitive work, one item at a time

  1. Pick the next topic from memory, without a record of what the audience had asked.
  2. Write and format every blog article from scratch.
  3. Answer the same questions again across new videos and posts.
  4. Watch competitors and trends by hand, when time allowed.
  5. Keep the site current on a slower platform with third-party scripts and a cookie banner.

03 / The goal

Remove the repetition and the guesswork

Read what the audience asks across the whole catalogue.
Weigh those questions against competitors and wider trends.
Turn the signal into ranked topics and first drafts in a fixed house style.
Publish into a fast site that sets no cookies and carries no advertising or cross-site trackers.
A person approves every result before it goes live.

04 / Tools used

A small stack, fitted to the job

The site is a static build on a managed edge platform. A small server-side tool reads the platforms’ own comments through read-only APIs. Scheduled AI agents draft against a fixed set of rules. The stack goes no further than the work required.

05 / How the workflow runs

Two loops. Each ends at a person

The answering loop

One review queue for every new comment

One agent keeps a single FAQ current as new questions arrive. A second drafts replies to roughly 90% of new questions from that FAQ. The creator receives one review queue, approves what reads right, writes anything missing, and publishes the batch.

  1. Comments

    Read-only tooling collects new comments across platforms

  2. FAQ

    One agent keeps a single FAQ current

  3. Replies

    A second agent drafts about 90% from the FAQ

  4. Daily review

    The creator approves, edits, or writes the rest

  5. Posted

    Approved replies go out in one batch

The publishing loop

Signals become a complete publishing kit

The Researcher scores candidate topics using audience questions, competitor posts, and wider trends. A person chooses the topic and adds real details. The Drafter then prepares the article, video cue sheet, title, description, tags, thumbnail brief, teaser, clips, and hooks for approval.

  1. Signals

    Own comments, competitor posts, and trends

  2. Researcher

    Scores and ranks candidate topics

  3. Review

    A person picks the topic and adds real details

  4. Drafter

    Drafts the article, video script, and publishing kit

  5. Approve & publish

    A person reviews before it goes live

06 / The result

The repetitive work now fits into one daily review

2h → 15ma day on incoming questions
3h → 15mto draft one article and its video script
~90%of incoming questions drafted from the maintained FAQ

The site scores 100 on standard page-speed tests on mobile and desktop, shows no cookie banner, and hands no visitor data to advertisers.

The creator now clears drafted replies in a 10-to-15-minute daily review. The same recurring questions feed the topic backlog, and a person still approves every published reply and everything the Drafter produces.

How these are counted: the before figures are the operator's own timings of the manual work; the after figures are the time the daily review and a single drafting run take now. The 90% is the share of incoming questions the Drafter produces a usable reply for, measured against the maintained FAQ. These come from one workflow, not from an average across clients.

07 / What we learned

Automation removed the repeated work and left every decision with a person

Decide from the evidence

The maintained FAQ surfaced a contradiction between two published answers. A person settled it; future topics now come from audience, competitor, and trend evidence.

Push fixes into the rules

When a drafting agent drifted off voice, the fix went into its rules so the same drift did not return on the next run.

Leave the right parts human

Filming and on-camera delivery stayed entirely with the creator. The system handles the scripting and publishing kit around that work.

08 / Where this fits another business

The listening half transfers directly

Any business answering the same questions at volume can mine what people ask, draft the repetitive response, and hold it for a person to approve.

That order matters: the FAQ existed before any agent drafted a reply from it.

Discuss Your Workflow

Not sure which of your own processes to start with? Score them with the free checklist.