How it works

Inside the system.

The case study shows what the channels are credited with. This page shows the machinery: how content gets researched, written and checked, how follow-up is designed to run without a person, and how a result gets counted. Bank names, internal figures and prompts are left out.

  • LiveIn use today
  • DesignBuilt spec, shown with generic names
  • PlannedNot built yet

01 · The content pipeline Live

How a guide gets made, and checked four times.

AI does the volume. Four separate gates decide what is allowed out. Scroll to follow one piece through.

  1. Step 1

    A research file comes first

    Before a word is drafted, every fact the piece will use is collected in one file: quoted word for word, with the link it came from. Outside facts come only from official sources.

  2. Step 2

    One row in the content plan

    Each piece starts as a row: the search phrase it targets, the questions it has to answer, its call to action, and a flag for anything compliance will want to see.

  3. Step 3

    An AI writer drafts from a brief

    The writer works from a standing brief of about 1,450 words: voice, structure, banned claims, how to cite. It drafts the piece, its graphics and the page build in one pass.

  4. Gate 1 · AI reviewer

    A second AI, with its own brief, tries to break it

    A separate reviewer checks accuracy and compliance, voice, search and every image. It fixes what it finds and logs each fix. Across ten guides it logged:

    66accuracy & compliance
    66voice
    40search
    19image

    Real catches: a "fixed rate" line that contradicted the product's own footnote, an overclaim that "no cash sits idle", and a fee that didn't match the live page.

  5. Gate 2 · Build check

    The build refuses to ship bad pages

    The build script stops if a banned phrase appears anywhere (body, alt text, title, summary or graphics), if the FAQ and its structured data disagree, or if the piece runs long.

  6. Gate 3 · Me

    I read every piece

    Nothing goes to the bank until I've read it start to finish. The AI passes catch the volume; I catch what a person at the bank would notice.

  7. Gate 4 · Compliance

    The bank signs off

    Each piece carries a "confirm before publish" list that compliance and the product owners work through. Nothing is scheduled until they have. Video goes to them as a PDF with every disclosure, its exact wording and the key frames.

  8. Step 4

    Released to publish

    Only a piece that clears all four gates is released. The current guides are produced and in review; they roll out on the content calendar. The rules behind all of this are in AI governance for marketing teams.

02 · Follow-up that runs itself Design

The abandoned-application workflow, step by step.

Designed in 2026 for a checking-account application, with a holdout so it can be measured. Pick a sample contact and run them through it.

Run a contact:

Pick a contact above to watch the path. The full logic is written out below.

  1. EnrollApplication started, not finishedand marketing consent isn't on file
  2. SuppressExisting customers removeda customer list is checked firstExitAlready a customer: never enrolled
  3. BranchHoldout group?10% held back on purposeYes · endNo email. Measured against the mailed group.
  4. Email 1Pick up where you left offthe destination is spelled out as text
  5. Wait2 days
  6. Email 2A reminder, with an out"Already finished? Ignore this."Goal · exit nowCompleted, banker following up, or withdrawn
  7. Wait3 days
  8. Email 3Last notesame "already finished?" line
  9. EndWorkflow endsno automatic hand-off to nurture
Rules that matter
  • Sends only in the daytime, weekdays and Saturday.
  • Emails never ask for a password or a full Social Security number, and the resume link carries no email address.
  • Re-enrolls only if a new application is started.
  • Consent is tracked apart from the CRM's billing flag, with its own date and source.
Honest limits
  • Application status arrives by manual upload, not a live feed, so it can lag a day or more. That's why emails 2 and 3 carry the "already finished?" line.
  • Designed and built, not yet measured. The holdout is there so the lift can be measured honestly, not guessed.
  • The opt-in nurture series (five emails over two weeks) hands off to this workflow when someone starts an application: highest intent wins.

03 · How a result gets counted Live

One account, one channel, counted once.

The rule behind every dollar and account figure on this site, and where it's weak.

  1. 1A lead is recordedwith the channel it first came through
  2. 2First channel winsmatched on a cleaned-up email address
  3. 3An account opensany new deposit or loan account
  4. 4Reconciledagainst the bank's account-opening reports
  5. 5Counted onceno multi-touch double counting
Where this method is weak
  • First touch under-credits channels that help later, like mail to people already on the email list.
  • Web analytics barely sees social, so social is judged from the CRM and account records, not GA4.
  • The method changed in 2026, so earlier years aren't perfectly comparable, and 2024 detail is missing.
  • Some offline credit, like signage and print, is probably recorded by staff.
Fixes on the way Planned
  • One naming convention for every link and QR code, with a registry and redirect links.
  • A hidden entry-source field on forms, matched to opened accounts and checked again at 60 and 180 days.
  • Cross-domain tracking into the online application, so starts and finishes can be seen.

Email already tags itself: automated sends and one-off sends carry different source codes, so they can be told apart.

04 · What broke

What went wrong, and what I changed.

Systems earn trust by how they fail. These are the ones worth knowing about.

Content

The AI reviewer found errors in every guide it checked.

Rate claims, fee amounts, overclaims: the kind of mistake an AI writer makes with confidence.

Changed: an AI reviewer, an automated build check and my own read, before compliance sees a draft. That's Gate 1 to Gate 3 above.

Website

A web security filter blocked rich pages, even drafts.

Formatted guides and structured data couldn't be saved in the site's editor.

Changed: rebuilt as plain HTML, with the structured data moved into the tag manager.

Video

The first ad cuts tried to say too much.

Seven or eight claims a spot, and too many cuts to follow.

Changed: retired them. One message per spot, longer holds, sound effects cut from 27 to 17, and a sped-up legal read rejected.

Social

One decision, and a direction, not proof.

In August 2025 the social ads moved from gift-led to checking-led. Social-sourced accounts are now 5× where they were (2026 so far vs all of 2022), though they were already climbing before the change.

Honest read: there was no control group, so I report it as a direction, not a result.

The numbers

Want the figures behind it?

Channel dollars, account counts and cost per account are the bank's internal figures, so they live in the private case study. Email me and I'll send access.

Request access Read the public case study

Start a conversation

Let's talk about the role.

Hiring for marketing, or want to compare notes on a system you're building? Email is the fastest way to reach me.

Email Marc