Treat Bank Marketing Content as Data

A new marketing executive came in and found out that a holiday postcard, end to end, was going to take two to three months. Her jaw dropped, and it should have. This was a product we had been promoting for years, on a four by six card, in our standard design system.

Nobody was revolutionizing the postcard. So she sat in every meeting. Brief, copy, design, legal, list pull, proofs. If a review was due Thursday she was in the reviewer’s office Thursday, and it went out in two weeks.

She worked with the people and tools already there, and nothing was bought and nobody was hired. The product had not changed in years and still went through the whole chain from a blank page every time, and no stage in that chain measured how long work waited in front of it.

She could not sit in every meeting for every postcard. We componentized everything on that card that could be componentized, and for the next two to three years the incremental postcard reviews stayed under two weeks without anyone senior in the room. We tracked throughput via Workfront, which counted what shipped and did not time the waits.

Most stages are never measured

A routine asset took one to two months from request to use, and most of that time it sat in queues nobody measured.

I ran the content operation inside marketing at a bank holding company, and before that at a global industrial manufacturer. In both, that was how long a landing page, an email or a branch flyer took.

Requests waited at intake because nobody was allowed to say no, so everything was accepted and nothing was prioritized. Reviews had no clock, and nobody ever got in trouble for sitting on one.

The product owner who has to approve the piece is building the plane after jumping off the cliff. Reading a piece to work out what they do not like about it and what should change is real work, and when they plan their day it is never in the top one or two things on the list.

So it sits. Nobody is being lazy. Review is a second job on top of a full day of other work, and it comes with no deadline.

Translation queued behind the last thing that was sent, and build queued behind everything above it. Six or seven review rounds were normal. Some assets were never final until someone senior said ship it.

In the libraries I have audited, somewhere between a fifth and two fifths of what got made was ever used again. We counted reuse as requests that needed no creative work beyond minor revisions. Regions rebuilt what already existed because finding the thing was harder than making it, and both versions went through review and translation.

You will also see “60 to 70 percent of B2B content goes unused” in every deck on this subject. It comes from a 2013 conference remark, and nobody ever published a method or a sample for it.¹ CreativeX tracked 1,284 core assets at consumer brands in 2023 and found half never showed up in any channel it monitors.²

Content is data

Every asset is a record and every approved component is master data, and your data team already has the vocabulary and the tools to track both.

An approved component is a claim, a disclosure, a rate, a spec, or an image with its rights cleared. Data teams call that master data: the one approved version that every asset built from it should point back to. Most content libraries store the file and nothing else, so you can count your files and cannot say what any of them is made of or where it ran.

Marketing never borrowed the data team’s tools because it has always thought of its output as creative, and creative work does not get IDs.

Lineage is the record of what each asset is made of and where each part is used. Run it downward and you see what a change touches: before you change a rate or expire a disclosure, you see every asset, in English and Spanish, that carries it. Data teams already use open tools for this in their own pipelines; none is built for content yet.³

Legal approves each component once. Every asset that uses the disclosure block inherits that approval, and translation attaches the same way.

Version history is the audit. A regulator picks a day in the past, say March 14, and asks what the customer saw that day. This happens all the time, and people dig through old project records in Workfront by hand to find it. A data warehouse answers that every day by keeping every version of a row with the dates it was current.⁴ Most content systems keep only the version that is live today, so they cannot answer it.

Data engineers also draw a pipeline as a chain of jobs, each depending on the one before, with a timestamp on every handoff. The content chain fits it: brief, create, translate, review, build, publish. Put a timestamp on each handoff and you can see how long work sat between steps, which the postcard executive did by hand, one meeting at a time.

Pharma approves each claim once

With each claim approved once, pharma averages 1.3 review cycles per asset, and most of the field content it approves is rarely or never used.

Pharmaceutical promotion goes through medical, legal, and regulatory review, called MLR, on every piece. The industry’s answer to that load, over the last decade, was to build claims libraries and modular content: each claim approved once with its references, each module a pre-approved block, each asset assembled from modules with the review reading the assembly.

Pharma’s number is lower, and I read the reason as each claim being approved before the asset existed, so the review has less to read.⁵⁶⁷

In 2016, 81 percent of life sciences companies surveyed could not report where their claims and content were in use, and 70 percent had no central global library.⁸ Approving each claim once did not give them lineage.

Veeva’s own CRM data, the two Veeva rows in the unused-content chart above, shows that most approved field content is rarely or never used.⁹

Approval attached to the component fixes the review. Deciding what to make, and what to stop making, is a separate job, done by a person at intake.

Content with no parent costs money and invites fines

An asset with no parent costs as much as any other asset, and no lineage query can find it. The example below prices that with stand-in numbers for you to replace.

The cost

An asset built from a downloaded file has no parent. Pricing them takes three numbers: how many assets you ship, how many have a parent on file, and what an asset costs.

Adobe’s 2025 survey found 70 percent of organizations producing at least 1,000 assets a year and two fifths producing more than 10,000, so 5,000 a year is a round example inside that range.¹⁰

For the share with a parent on file, I have no source. 40 percent is my guess for a shop that has never asked for one, and a generous one.

Fully loaded cost per asset, with agency fees, internal hours, translation, review time, and rework divided by what shipped, is a number finance has to build. No public benchmark for it exists; I looked. $1,800 is a round placeholder, and yours replaces it.

Some of that spend was necessary. Nobody can say which part, and that is before the duplication those assets caused, the translation of things that were already translated, and the compliance exposure of disclosures with no record of where they ran.

Every upload declares a parent, or it does not publish. A new asset is fine, as long as it says it is new.

The regulator’s question

Regulators ask what each customer received on a given day. FINRA Rule 2210 requires the copy, the dates it ran and who approved it, kept three years, and one firm paid $250,000 for keeping only the template.

A bank with a broker-dealer arm answers to FINRA, and Rule 2210 spells out the whole record: every retail communication kept for three years, with a copy, the dates of first and last use, the name of the principal who approved it and the date they did, and the source of any statistical table or chart.¹¹

Credit unions sit outside Regulation DD and follow the NCUA’s version, Part 707. Regulation DD’s commentary says sample copies of advertisements are enough, so FINRA’s list is the stricter of the two, and a content record built to it covers both.¹²

Regulation Z exempts advertising from its retention rule, so a lender’s ability to prove what an ad said rests on its own records and its UDAAP defense, with no mandate behind it.¹³

FDIC Part 328 gives insured banks until April 1, 2027 to put the FDIC’s digital sign on their websites and apps. A sign that repeats across a site and an app is a component: approved once, placed everywhere, with the date it took effect.

H2C kept the template.¹⁴ The version each customer received was not kept, and the record the rule asks for is the version, which is the March 14 case. M1 Finance paid $850,000 the same spring for influencer posts it did not review, approve, or retain.¹⁵

Reuse is the default

A person at intake with the library in front of them closes requests that existing assets already cover, and small changes ship in a day or two.

A request queue will accept the same brochure from four regions. Triage is the person at intake who stops that.

Say a region asks for a flyer for the new small-business checking account, in Spanish, for the chamber breakfast in six weeks. Triage finds two, the business banking one that shipped in March and the Houston branches’ adaptation of it in May, links both, and closes the request as fulfilled.

The region comes back to say the Houston one is close, but the photo is wrong for their branches and the event line needs to go. Triage tells them to reopen it with that, since a photo swap and a deleted line is the variant lane: two days, and review reads only what changed.

Then the region asks what happens if they need a new claim on it. Triage says that makes it a new piece, through the whole chain, compliance included. When triage asks which claim, the region says never mind the claim.

• The variant lane. An image swap, a call to action, a market line, connective copy: nothing that touches an approved component. A production designer spins the variant from the parent, review reads only what changed, and turnaround is a day or two.
• The new-piece lane. Anything that touches a regulated component or the layout. It goes through the full chain.

Give a person who is responsible for the outcome a choice between something new and something that exists, and they will pick new every time. They believe new gives them the best chance to stand out, and they are emotionally invested in the result, which they should be, because the P&L is theirs.

In my experience the belief is wrong. The customers they are targeting have never seen the existing asset, or anything like it, so the new version looks better mainly to the person who asked for it.

Our job as their marketing partner is to tell them how long new really takes. New is six weeks, and it will get pushed, and it will get delayed. The one that is 90 percent right is in their hands by the end of the week.

Put those two next to each other and most people take the asset. Ask them “do either of these work for you?” and they hear that new is on the table, and they take new.

Language and voice

Each translation is a version of the approved English record, and a change to the English rate flags every language it runs in.

About 74 million people in the US speak a language other than English at home, by the Census Bureau’s 2024 count: 44.9 million speak Spanish, 3.7 million speak Chinese, and 28.9 million speak English less than very well.¹⁶

A Spanish savings ad in Houston has to meet the same Regulation DD advertising rules as the English one, and so does a Chinese one in San Francisco. The same compliance team owns the rate, the claim and the disclosure in every language.

The rule for account disclosures already assumes a parent. Regulation DD lets a bank give them in another language “provided the disclosures are available in English upon request,” and the NCUA’s rule for credit unions uses the same words.

Two rules go further. Regulation Z calls a mortgage ad misleading if it gives some required terms, such as the initial rate, only in another language and the rest only in English, and Regulation E requires remittance and prepaid disclosures in another language when the bank principally uses that language to market, package or sell those products.¹⁷ So every translated disclosure needs an English original behind it, and a translated mortgage ad cannot split its required terms between two languages.

We rarely did Spanish, and when we did, it went badly, mostly because people claimed the translations were wrong.

That complaint is why the branch team can rewrite the headline in their own words, since they talk to those customers every day. The rate and the disclosure under it are translated from the approved record, and nobody rewrites them. Translators work faster on a sentence they have translated before:¹⁸

The CFPB’s 2021 statement on customers with limited English proficiency suggests a notice, in the customer’s language, of “the extent and limits of any language services” the bank offers, which is a slot the English ad never needed. The statement is guidance rather than a rule, and as of September 2026 it is not on the bureau’s list of withdrawn guidance.

Its examples include a 2013 enforcement action against a card issuer that enrolled customers in Puerto Rico in an add-on product on Spanish sales calls and sent every written document in English, and supervisory findings about issuers that marketed only some of their cards to Spanish speakers.¹⁹ Lineage answers the question those findings raise in one list: which products have a Spanish version, and which exist only in English.

Every translated disclosure needs an English original behind it.

Shorten the review cycle by letting legal focus on what has changed

Legal approves the template and its allowed combinations once, then gives a full read only to a combination that is new.

Two approved components next to each other can still make a misleading ad, so the combinations need approval too. Pharma’s 1.3 cycles is the number to beat, though the Veeva data does not say how it handles combinations.

Triage flags any combination that is new, and legal decides. Keeping that list of combinations short enough to read in an afternoon is the reason to model only the 20 components behind most of the volume and leave the rest as pages. And the assembly gets a name on it, because when a pre-approved combination turns out misleading in context, legal will ask whose.

Triage is a political job. Diagrams draw it as a box, and in practice it is a person who says no to a regional director once a day and gets escalated over once a week.

Triage holds only with an executive who has said, in front of the regions, that reuse is the default; with a monthly count of requests reopened after triage closed them; and with the agency contract rewritten around variants instead of rounds.

The postcard executive was triage for two weeks.

Buy AI tools last

The variant lane is where AI belongs: image and copy variants, assembled from approved parts inside a locked template. The tools that check generated work against brand and product guidelines check it against whatever guidelines they are given. If those guidelines are the approved components, the tools help. Buy them before the approved components exist and you get one-offs faster, each of which goes through full review and full translation, and the pile in front of legal gets taller.²⁰

Give every asset a parent

Componentizing kept the postcard under two weeks for two to three years with nobody senior in the room. These four changes are how to try the same on the rest of the library.

1. Put a person at intake who can close requests. They search the library before anything is made, link what already exists, and close the request when an existing asset covers it, backed by an executive who has told the regions that reuse is the default.
2. Every upload declares a parent, or it does not publish. New assets are allowed when they are marked new. Everything that ships after that has a parent on record.
3. Run two lanes. A photo swap or a new call to action ships in a day or two, with review reading only what changed. A new claim or a new layout takes the full chain.
4. Buy AI tools last. Model the 20 components behind most of the volume and have legal approve each one once. Then put AI in the variant lane, inside a locked template.

About the numbers

Research, rules, and product documentation checked September 25, 2026. Every sourced figure is traced to a dated primary source and each footnote says what the figure does not show.

The $5.4 million figure is an illustration: the lineage coverage and the cost per asset are placeholders, and 5,000 assets is a round example. The 20 components is my judgment from my own work. The regulatory table is read from the rules as published on eCFR and FINRA’s rulebook; the content model and triage are analysis and recommendation from my own work, and the conclusions are mine.

Each figure, rule and story in this article is recorded with where it came from, whether a footnote, my own work or a stand-in for your number, and the SiriusDecisions figure is recorded as having no source at all. The site will not publish a page that points to a record it cannot find.

If you want to see this with your organization’s own numbers, book twenty minutes.